Mobile Crowdsourcing: From Theory to Practice 3031323963, 9783031323966

Table of contents :
Preface
Contents
Part I Introduction
Crowdsourcing as a Future Collaborative Computing Paradigm
1 Definition and History
1.1 HPU and CPU
1.2 Basic Components
1.3 History
2 Crowdsourcing Events in Recent History
2.1 Help Find Jim Gray (2007)
2.2 Malaysia Airlines Flight MH 370 (2014)
2.3 DARPA Network Challenges (2009)
2.4 Tag Challenges (2012)
2.5 Kasparov vs. IBM Deep Blue (1997)
2.6 A Big Picture: Human vs. Machine
3 Crowdsourcing Overview
3.1 Workflow of Crowdsourcing
3.2 Types of Crowdsourcing
4 Platform
4.1 Amazon Mechanical Turk
4.2 Crowd4U
4.3 gMission
4.4 UpWork
4.5 CrowdFlower
5 Sample Applications
5.1 Image Processing
5.2 Commonsense Knowledge
5.3 Smart City
5.4 Other Science Projects
5.5 Mixed HPU and CPU Applications
6 Algorithmically and Theoretically Challenging Issues
6.1 Paradigm
6.1.1 Sequential Implementation
6.1.2 Parallel Implementation
6.1.3 Divide and Conquer
6.2 Multi-Armed Bandit (MAB)
6.3 Incentive Mechanisms
7 Opportunities and Future Directions
7.1 Beyond Simple Workflows
7.2 Beyond Simple Worker Selection
7.3 Beyond Independent Workers
7.4 Beyond Simple Training
7.5 Beyond Simple Interactive Mode
7.6 AI Applications
7.7 Crowdsourcing 2.0
8 Conclusion
References
Urban Mobility-Driven Crowdsensing: Recent Advances in Machine Learning Designs and Ubiquitous Applications
1 Introduction
2 An Overview of Urban Mobility-Driven Crowdsensing
3 Advancing Machine Learning Designs for UMCS
3.1 Machine Learning Advances in Crowdsensed Signal Reconstruction
3.2 Machine Learning Advances in Understanding Crowd Mobility Distributions
4 Expanding Ubiquitous Use Cases for UMCS
4.1 Indoor Crowd Detection and Group Identification
4.2 Urban Mobility Reconfiguration with Crowdsourced Information Fusion
5 Conclusion
References
Part II Key Technical Components: User Recruitment and Incentive Mechanisms
Unknown Worker Recruitment in Mobile Crowdsourcing
1 Introduction
2 Related Work
3 System Model and Workflow
3.1 System Model
3.2 System Workflow
4 Unknown Worker Recruitment Scheme
4.1 Modeling and Formulation
4.2 Algorithm Design
4.3 Theoretical Analysis
4.4 Extansion: Budget-Limited UWR Scheme
5 Privacy-Preserving Unknown Worker Recruitment Scheme
5.1 DP-MAB Model
5.2 Problem Formulation
5.3 The DPF Algorithm
5.4 Performance Analysis of DPF Algorithm
5.5 DPU Algorithm
5.6 Performance Analysis of DPU algorithm
6 Conclusion
References
Quality-Aware Incentive Mechanism for Mobile Crowdsourcing
1 Introduction
2 Related Work
3 Quality-Aware Incentive Mechanisms for MCS Systems
3.1 System Overview
3.2 Auction Model
3.3 Design Objective
3.4 SRC Auction
3.4.1 Mathematical Formulation
3.4.2 Mechanism Design
4 Quality-Aware Incentive Mechanism Considering the Bid Privacy for MCS Systems
4.1 System Overview
4.2 Aggregation Method
4.3 Auction Model
4.4 Design Objective
4.5 Mathematical Formulation
4.6 Mechanism Design
5 Conclusion
References
Incentive Mechanism Design for Mobile Crowdsourcing Without Verification
1 Introduction
1.1 Motivations
1.2 Key Challenges
1.3 Chapter Outline
2 Model
2.1 Workers' Decisions and Payoffs
2.1.1 Task and Workers
2.1.2 Worker Effort Exertion Strategy
2.1.3 Worker Solution Reporting Strategy
2.1.4 Incentive Mechanism
2.1.5 Worker Payoff
2.2 Platform's Decisions and Payoff
2.2.1 Platform Decisions
2.2.2 Platform Payoff
2.3 Platform–Worker Interaction
3 Approaches to Worker Heterogeneity
3.1 Motivating Examples and Key Questions
3.2 Solution: Majority Voting Mechanism
3.3 Results and Insights
4 Approaches to Worker Collusion
4.1 Motivating Examples and Key Questions
4.2 Solution: Truth Detection Mechanism
4.3 Results and Insights
5 Approaches to Information Incompleteness
5.1 Motivating Examples and Key Questions
5.2 Solution: Randomized Learning Mechanism
5.3 Results and Insights
6 Approaches to Information Asymmetry
6.1 Motivating Examples and Key Questions
6.2 Solution: Bayesian Persuasion Mechanism
6.3 Results and Insights
7 Conclusion and Open Problem
7.1 Future Challenges and Open Issues
7.1.1 Joint Optimization of Information Elicitation and Aggregation
7.1.2 Competitive Market
7.1.3 Worker Bounded Rationality
7.1.4 Worker Privacy and Moral Issues
7.2 Conclusion
References
Part III Key Technical Components: Task Allocation
Stable Worker–Task Assignment in Mobile Crowdsensing Applications
1 Introduction
2 Background
2.1 Worker–Task Assignment in Mobile Crowdsensing
2.2 Matching Under Preferences
3 Why Should We Care About Stability in MCS?
4 Stable Task Assignments in Different MCS Applications
4.1 Participatory MCS
4.2 Opportunistic MCS
4.3 Hybrid MCS
5 Conclusion and Open Problems
References
Spatiotemporal Task Allocation in Mobile Crowdsensing
1 Introduction
2 Optimized Allocation of Time-Dependent Tasks for Mobile Crowdsensing
2.1 Problem Statement
2.2 System Overview
2.3 Problem Formulation
2.4 Task Allocation Algorithm
2.5 Performance Evaluation
3 Heterogeneous User Recruitment of Multiple Spatiotemporal Tasks
3.1 Problem Statement
3.2 System Overview
3.3 Problem Formulation
3.4 Model Analysis
3.4.1 Heterogeneous Task Priority Model
3.4.2 Platform Payment Incentive Model
3.4.3 User-Contributed Task Coverage Ratio Model
3.4.4 Binary-Based Representation of Level
3.5 HURoT Problem-Solving Approaches
3.5.1 Utility Function with Dual Objectives
3.5.2 Utility-Based User Recruitment (UURe)
3.5.3 Level-First and Utility-Based User Recruitment (L-UURe)
3.5.4 Global Level-First and Utility-Based User Recruitment (GL-UURe)
3.6 Performance Evaluation
3.6.1 Experiment Settings
3.6.2 Experimental Results and Analysis
4 Conclusion
References
Part IV Key Technical Components: Data Inference
Joint Data Collection and Truth Inference in Spatial Crowdsourcing
1 Introduction
1.1 Challenges and Motivations
2 Model of Truth Inference and Task Allocation
2.1 System Overview
2.2 Truth Inference
2.2.1 Numerical Task
2.2.2 Categorical Task
2.3 Task Allocation
2.4 Process of Crowdsourcing System
3 Online Expertise-Aware Truth Inference
3.1 Maximum Likelihood Numerical Inference
3.2 Expectation Maximization Categorical Inference
3.3 Algorithm Design for Truth Inference
4 Online Location-Aware Task Allocation
4.1 Probability Improvement-Based Allocation
4.1.1 Numerical Tasks
4.1.2 Categorical Tasks
4.2 Entropy-Reduction-Based Allocation
4.2.1 Numerical Tasks
4.2.2 Categorical Tasks
4.3 Algorithm Design for Task Allocation
5 Performance Evaluation
5.1 Dataset and Settings
5.1.1 Dataset
5.1.2 Parameter Settings
5.1.3 Comparison Algorithms
5.1.4 Evaluation Metric
5.2 Results of Truth Inference
5.3 Results of Task Allocation
5.4 Results of Running Time
5.5 Results on Larger Dataset
6 Chapter Summary
References
Cost-Quality Aware Compressive Mobile Crowdsensing
1 Background
2 System Model and Problem Statement
2.1 System Model
2.2 Data Inference
2.3 Importance Assessment
2.4 Cost Assessment
2.5 Quality Assessment
2.6 Problem Formulation
3 Advanced Cell Selection Strategies in CCS
3.1 Randomized Sampling Strategy
3.1.1 Recovery Accuracy Prediction Based on Regularized Column Sum
3.1.2 CACS via Convex Optimization
3.2 Active Sampling Strategy with Multiple Steps
3.2.1 Use Case Study
3.2.2 Cost Estimation
3.2.3 Cost–Quality Beneficial Cell Selection
3.3 Active Sampling Strategy Based on Bipartite Graph
3.3.1 Representing Matrix Factorization Based on Bipartite Graph
3.3.2 Sampling to Form a Complete and Robust Linear System
4 Evaluation
4.1 Experimental Setup
4.1.1 Datasets
4.1.2 Baselines
4.2 Experimental Results
4.2.1 Errors of Inferred Value
4.2.2 The Number and Total Costs of Selected Cells
5 Summary
References
Part V Key Technical Components: Security and Privacy
Information Integrity in Participatory Crowd-Sensing via Robust Trust Models
1 Introduction
2 Architecture for Participatory MCS
3 Security Threats and Challenges
3.1 Types of Dishonest Behaviors
3.2 Cold Start Problem and Other Challenges
3.3 Categories of Vulnerabilities and Attack Types:
4 Quality and Quantity Unified Architecture for Secure and Trustworthy Crowd-sensing
4.1 Robust Quality of Information Model
4.1.1 Posterior Estimation of Probability Masses
4.1.2 Non-linear Weighing of Probability Masses:
4.1.3 Link Function
4.2 Robust User Reputation Scoring Module
4.2.1 Modified Link Functions
4.2.2 One-Hot Encoded Sum
4.2.3 Output Activation and Classification Criterion
5 Analytical Case Study
6 Conclusion
References
AI-Driven Attack Modeling and Defense Strategies in Mobile Crowdsensing: A Special Case Study on Fake Tasks
1 Introduction
2 Background on Mobile Crowdsensing
2.1 Use Cases of MCS
2.2 System Architecture of MCS
2.3 Quality of Service in MCS
3 Security and Threat Models in MCS
3.1 Threat Models in MCS
4 AI-Driven Attack Anticipation in MCS
4.1 Fake Task Injection Modeling
4.2 Types of Task Movement
4.2.1 Zone-Free Task Movement (ZFM)
4.2.2 Zone-Limited Task Movement (ZLM)
4.3 Self-organizing Feature Map Implementation for Attack Modeling
4.4 Region-Based SOFM Structure
4.5 Locally Reconfigurable SOFM for More Impactful Attack Region Selection
5 AI-Driven Defense Strategies in MCS
5.1 AI-Backed Legitimacy Detection
5.2 Machine Learning Model Development to Increase the Performance of Legitimacy Detection
6 Conclusion
References
Traceable and Secure Data Sharing in Mobile Crowdsensing
1 Introduction
2 Related Work
2.1 Mobile Crowdsensing
2.2 Privacy-Enhancing Techniques for Mobile Crowdsensing
3 Traceable and Privacy-Preserving Non-interactive Data Sharing (TIDS) Scheme
3.1 Problem Statement
3.1.1 System Model
3.1.2 Threat Model
3.1.3 Design Goals
3.2 Preliminaries
3.2.1 Bilinear Pairings
3.2.2 Access Structure
3.3 The TIDS Scheme
3.3.1 The TIDS Framework
3.3.2 The Detailed Description of TIDS
3.4 Security Analysis
3.5 Performance Analysis
3.5.1 Theoretical Analysis
3.5.2 Experimental Evaluations
4 Conclusion and Future Work
References
User Privacy Protection in MCS: Threats, Solutions, and Open Issues
1 Introduction
2 User Privacy Threats and Requirements
2.1 Threat Model
2.2 Privacy Attacks
2.3 Privacy Threats
2.3.1 Privacy Threats from Task
2.3.2 Privacy Threats from Data
2.4 Privacy Leakage in the Whole Data Flow Process
2.4.1 Privacy Leakage in Task Allocation
2.4.2 Privacy Leakage in User Incentive
2.4.3 Privacy Leakage in Data Collection
2.4.4 Privacy Leakage in Data Processing and Publishing
2.5 Requirements for User Privacy
3 Privacy Protection Technologies
3.1 Anonymization-Based Technologies
3.1.1 Generalization
3.1.2 Suppression
3.2 Perturbation-Based Technologies
3.2.1 Randomized Response
3.2.2 Differential Privacy
3.3 Encryption-Based Technologies
3.3.1 Fully Homomorphic Encryption
3.3.2 Partially Homomorphic Encryption
4 Privacy Protection for Mobile Users
4.1 User Privacy Protection in Task Allocation
4.2 User Privacy Protection in Incentive
4.3 User Privacy Protection in Data Collection and Publishing
5 Open Issues for Mobile User Privacy Protection
5.1 Full Lifecycle Privacy Protection Framework
5.2 Breaking the Privacy-Overhead-Utility Trilemma
5.3 Incorporation of Novel Privacy-Preserving Computing Technologies
5.4 Privacy Protection for Users' Fresh Time-Series Data
6 Conclusion
References
Part VI Applications
Crowdsourcing Through TinyML as a Way to Engage End-Users in IoT Solutions
1 Introduction
2 Strategies for End-User Engagement
3 Basic Concepts of TinyML
3.1 Building a TinyML Application
3.2 TensorFlow Lite Micro
3.3 The Edge Impulse Platform
4 Example Applications with TinyML
5 TinyML on Device Development
6 Beehive-Application
7 Conclusions
References
Health Crowd Sensing and Computing: From Crowdsourced Digital Health Footprints to Population Health Intelligence
1 Introduction
2 Preliminaries and Fundamentals
2.1 Fundamentals for Population Health and Epidemiology
2.2 Crowd Sensing and Its Applications in Health Care
2.3 AI Technologies as Enablers of HCSC
3 Conceptual Framework for HCSC
4 Case Study: Compressive Population Health
4.1 Approach
4.2 Experimental Results
5 Research Opportunity and Proposal for Future HCSC
6 Non-scientific Considerations for HCSC: Privacy, Ethics, and Security
7 Conclusion
References
Crowdsourcing Applications and Techniques in Computer Vision
1 Introduction
2 Computer Vision
2.1 Machine Learning Applied to Computer Vision
3 Crowdsourcing Applications in Computer Vision
3.1 Computer Vision Datasets
3.2 Labelling Software
3.3 Crowdsourcing Use Cases
4 Crowdsourcing Data Aggregation and Evaluation
5 Concluding Remarks
References
Mobile Crowdsourcing Task Offloading on Social Collaboration Networks: An Empirical Study
1 Introduction
2 Related Work
3 Preliminary Understanding by Questionnaire
4 Further Experiment: Field Experiments and Analysis
4.1 Experimental Settings
4.2 Study I: The Basic Situation
4.2.1 Basic Observation in Task Offloading
4.2.2 The Impact of Punitive Measures
4.3 Study II: Task Offloading Pattern Investigation
4.3.1 The Patterns of MCS Task Offloading
4.3.2 Multi-hop MCS Task Offloading
4.4 Study III: Incentive Mechanism in Task Offloading
4.4.1 Effect of Bid Incentive Rewards
4.4.2 Incentive Reward-Sharing Mechanism
5 Conclusion
References

Citation preview

Wireless Networks

Jie Wu En Wang   Editors

Mobile Crowdsourcing From Theory to Practice

Wireless Networks Series Editor Xuemin Sherman Shen, University of Waterloo, Waterloo, ON, Canada

The purpose of Springer’s Wireless Networks book series is to establish the state of the art and set the course for future research and development in wireless communication networks. The scope of this series includes not only all aspects of wireless networks (including cellular networks, WiFi, sensor networks, and vehicular networks), but related areas such as cloud computing and big data. The series serves as a central source of references for wireless networks research and development. It aims to publish thorough and cohesive overviews on specific topics in wireless networks, as well as works that are larger in scope than survey articles and that contain more detailed background information. The series also provides coverage of advanced and timely topics worthy of monographs, contributed volumes, textbooks and handbooks.

Jie Wu • En Wang Editors

Mobile Crowdsourcing From Theory to Practice

Editors Jie Wu Department of Computer and Information Sciences Temple University Philadelphia, PA, U.S.A.

En Wang Jilin University Changchun, China

ISSN 2366-1186 ISSN 2366-1445 (electronic) Wireless Networks ISBN 978-3-031-32396-6 ISBN 978-3-031-32397-3 (eBook) https://doi.org/10.1007/978-3-031-32397-3 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors, and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Preface

Crowdsourcing involves a large crowd of participants working together to contribute or produce goods and services for the society. The essence of crowdsourcing can be captured by the words: “crowd” and “outsourcing.” The history of crowdsourcing can be traced back to the early eighteenth century when the British government try to outsource the Longitude Problem to a crowd in pursuit of a solution toward safe sailing in the sea. The Oxford dictionary was the product of a crowdsourcing effort in the nineteenth century. In the twentieth century, Toyota selected its company logo through a crowdsourcing competition. The advancement of the Internet has enabled crowdsourcing to flourish in the early twenty-first century. The Internet has lowered the bar for a crowd to enter a crowdsourcing task. The term crowdsourcing was first proposed in 2006 by Howe and Robinson. The first scholarly paper that used the term crowdsourcing was by Brabham in 2008. Since the use of word crowdsourcing, there have been nearly 40 different interpretations of crowdsourcing. As time goes on, the notion of crowdsourcing and its key components have converged to the following: (1) there is an organization that has a task that needs to be performed or solved; (2) there is a crowd that is willing to perform the task either by incentive or voluntarily; (3) there is an online environment, also called a platform (e.g., Amazon’s Mechanical Turks), that facilitates the interaction between the organization and the crowd. The early twenty-first-century applications of crowdsourcing can be called crowdsourcing 1.0, which includes businesses using crowdsourcing to accomplish various tasks, including the ability to offload peak demand, accessing cheap labor, generating better results in a timely matter, and reaching out to a wider array of talent outside the organization. In the past decade, the widely used smart phones further broaden the horizon of the crowdsourcing applications. A smart phone can be used to perform basic sensing activities, perform basic computations, and communicate with other devices and with edge/cloud. With the explosive popularity of small phones and other small IoT devices and the rapid development of wireless networks and communication technology, we have reached an era of crowdsourcing 2.0, also called mobile crowdsourcing.

v

vi

Preface

Mobile crowdsensing, also known as spatial crowdsourcing, can be described as an extension of crowdsourcing to the mobile network to combine the idea of crowdsourcing with the sensing capacity of mobile devices. As a promising paradigm for completing complex sensing and computation tasks, mobile crowdsensing serves the vital purpose of exploiting the ubiquitous smart devices carried by mobile users to make conscious or unconscious collaboration through mobile networks, to complete large-scale and fine-grained sensing and computing tasks. Considering that we are in the era of mobile Internet, mobile crowdsensing is developing rapidly and has great advantages in deployment and maintenance, sensing range and granularity, reusability, and other aspects. Due to the benefits of using mobile crowdsensing, many emergent applications are now available for individuals, business enterprises, and governments. In addition, many new techniques have been developed and are being adopted. The main goal of this book is to collect the recent development on the principle, techniques, and applications in mobile crowdsourcing. This book will be of value to academics, researchers, practitioners, government officials, business organizations (e.g., executives, marketing professionals, resource managers, etc.), and even customers – working, participating, or those interested in fields related to crowdsourcing, and in particular, mobile crowdsourcing. The content of the book will be especially useful for students in areas like Internet of Things, Distributed Systems, Computer Networks, Data Mining, and Mobile and Pervasive Computing, but also applies to students of education, economy, or law, who would benefit from the information, cases, and examples herein. This book is suitable to serve as a supplemental or reference book for a graduate course in crowdsourcing, as well as for developers in the crowdsourcing industry. The focus of this book is to expose readers to the technical challenges in building a crowdsourcing system, together with various applications of mobile crowdsourcing. This book is organized into six parts with a total of 17 chapters. Each part corresponds to an important snapshot, starting from an introduction and overview of general crowdsourcing and mobile crowdsourcing. This book complements several books that have recently emerged on the topic, including Guo, Liu, and Yu’s book on Crowd Intelligence with the Deep Fusion of Human, Machine, and Things. For some introductory books on crowdsourcing and its history, check Brabham’s Crowdsourcing and Howe’s Crowdsourcing: Why the Power of the Crowd Is Driving the Future of Business. • Part I: Introduction (Chaps. 1 and 2) • Part II: Key Technical Components: User Recruitment and Incentive Mechanisms (Chaps. 3, 4, and 5) • Part III: Key Technical Components: Task Allocation (Chaps. 6 and 7) • Part IV: Key Technical Components: Data References (Chaps. 8 and 9) • Part V: Key Technical Components: Security and Privacy (Chaps. 10, 11, 12, and 13) • Part VI: Applications (Chaps. 14, 15, 16, and 17)

Preface

vii

Part I gives an overview of crowdsourcing/mobile crowdsourcing, their definitions, and relevant concepts. Chapter 1 presents an overview of crowdsourcing, including its history and basic components. This chapter takes a unique of view of CPU versus HPU (Human Processing Unit, a.k.a. crowdsourcing) to contrast crowdsourcing solutions with the traditional solutions carried out by CPUs. Some philosophical discussions on the intelligence and problem-solving capabilities of CPU versus HPU have also been included. Chapter 2 presents an overview of mobile crowdsensing from the perspective of urban mobility, called Urban Mobility-driven CrowdSensing (UMCS). This chapter introduces and analyzes the recent advances and applications of UMCS, especially with regard to the learning algorithm designs and emerging use-cases, which serve as a guideline for new researchers and practitioners in this emerging and interesting research field. Parts II to V focus on the main body of technical components in designing a mobile crowdsourcing system. Part II deals with the way and incentive to attract a crowd to get involved in a crowdsourcing event. Chapter 3 focuses on the basic worker recruitment problem in crowdsourcing, where the crowdsourcing platforms must recruit a set of workers under a certain budget without pre-knowledge of workers’ information. This chapter models this problem as a multi-armed bandit game. Then a recruitment scheme is presented to balance the exploration and exploitation processes of unknown worker recruitment. Chapter 4 studies another critical problem of incentivizing user participation in crowdsourcing in order to perform high-quality sensing tasks. By incorporating the Quality of Information (QoI) and the bid privacy, this chapter proposes two quality-aware incentive mechanisms. Based on both the single-minded and multi-minded combinatorial auction models, the proposed mechanisms minimize the total payment of platforms and satisfy the truthfulness and individual rationality. Chapter 5 investigates the incentive mechanisms, where a crowdsourcing platform intends to incentivize a group of workers to put effort and truthfully report their solutions to a given task. This chapter also surveys approaches to dealing with such problems using gametheory and online learning. Part III takes on resource allocations, including allocating the crowd to cover the task across the time-spatial domain. Chapter 6 studies a central problem in mobile crowdsourcing, i.e., the assignment of sensing tasks to workers. This chapter reviews the recent studies in various scenarios of mobile crowdsourcing and highlights the differences in each scenario together with a summary of corresponding solutions. Chapter 7 presents an overview of task allocation under time and space requirements. This chapter focuses on spatiotemporal tasks by considering two scenarios. The corresponding task allocation algorithms are proposed with both theoretical analysis and extensive simulations. Part IV focuses on data inference, including inference from sparse data. Chapter 8 outlines the use of prior knowledge for data inference. In this chapter, considering the problem that various expertise of workers and dynamic information of crowdsourcing systems are often ignored in data inference, the authors propose a framework of joint data collection and truth reference. Extensive evaluations demonstrate the superiority of these algorithms over the other state-of-the-art

viii

Preface

approaches. Chapter 9 presents the overview of Compressive Crowdsensing (CSS). Due to the high cost of mobile crowdsourcing, compressive sensing is introduced as a data inference method, forming a new method (i.e., CSS). This chapter focuses on the latest development of CSS and gives a solution to collect sensing data effectively and recover the unsensed data accurately. Part V studies security and privacy issues related to crowdsourcing activities. Chapter 10 presents a recipe for a trust and reputation scoring framework, under a cold start scene in participatory mobile crowdsourcing applications without ground truth and prior knowledge. The recipe is robust against attacks targeting operational and AI vulnerability in the form of fake event reporting attacks and feedback weaponizing attacks. As shown in this chapter, the effects of these attacks on the scoring framework are mitigated significantly by modifying existing methods. Chapter 11 focuses on implementing intelligent attacks and defense mechanisms in mobile crowdsourcing to protect mobile users and open platforms from cyberattacks. Specifically, the chapter builds several AI-empowered approaches from both the anticipatory and defensive perspectives regarding fake task injections. Open issues and challenges in mobile crowdsourcing systems are then carefully discussed. Chapter 12 discusses the security issues that appear during the datasharing procedure in mobile crowdsourcing, especially when the cloud server is not honest. As such, providing a traceable and secure data-sharing method is quite necessary. This chapter reviews several pioneering, secure data-sharing schemes, describes a special scheme (a Traceable and privacy-preserving non-Interactive Data Sharing scheme (TIDS) as an example), and then deeply discusses future research topics about these schemes. Chapter 13 investigates significant privacy concerns in mobile crowdsourcing, an emerging paradigm that engages a large scale of mobile users for data collection and processing. This chapter lists specific privacy concerns and requirements for users involved in mobile crowdsourcing systems, and then presents several effective approaches. Several open issues left for further research are discussed at the end of the chapter. Part VI surveys several applications of crowdsourcing. Chapter 14 introduces the application of crowdsourcing in Tiny Machine Learning (TinyML), which allows more end users to participate in crowdsourcing data for the IoT world. In this chapter, the current status of TinyML is reviewed by illustrating its underlying technologies and methodologies. Relevant examples show that crowdsourcing has a good application prospect in TinyML as a way to attract users. Chapter 15 introduces the application of crowdsourcing in public health operations, including population health monitoring and modeling. This chapter proposes a new vision called Health Crowd Sensing and Computing (HCSC), which leverages opportunistic and crowdsourced digital health footprints to realize the goal of more intelligent population health monitoring. The combination and application of AI techniques and HCSC are also discussed. Chapter 16 utilizes mobile crowdsourcing to solve computer vision problems. This chapter compares crowdsourcing solutions with traditional solutions created by experts from a unique perspective of combining AI and crowdsourcing. In addition, several applications of crowdsourcing in computer vision tasks are also explored. Chapter 17 conducts an on-campus empirical study

Preface

ix

on mobile crowdsourcing task offloading on social collaboration networks. This chapter investigates over 1000 workers and comprehensively examines the intrinsic characteristics and behavioral patterns in task offloading through field experiments. By analyzing the collected operation logs of the workers, several important findings about task offloading in mobile crowdsourcing applications are summarized. We would like to express our gratitude to all the contributing authors. This book would not have been possible without their generous contributions and dedications time-wise. Our special thanks are given to the Springer managing editor Susan Lagerstrom-Fife and production editor Arun Siva Shanmugam, who gave us both initial encouragement, support, and continuous guidance during the book editing process. Finally, we would like to thank our families in the USA and in China for their great understanding and patience during this project. This book is dedicated to them. We hope readers will find this book useful in their study or in their workplace! Philadelphia, PA, U.S.A. Changchun, China

Jie Wu En Wang

Contents

Part I Introduction Crowdsourcing as a Future Collaborative Computing Paradigm . . . . . . . . . . Jie Wu, Chao Song, and Wei Chang Urban Mobility-Driven Crowdsensing: Recent Advances in Machine Learning Designs and Ubiquitous Applications . . . . . . . . . . . . . . . . . . . Suining He and Kang G. Shin

3

33

Part II Key Technical Components: User Recruitment and Incentive Mechanisms Unknown Worker Recruitment in Mobile Crowdsourcing . . . . . . . . . . . . . . . . . . Mingjun Xiao, Yin Xu, and He Sun

61

Quality-Aware Incentive Mechanism for Mobile Crowdsourcing . . . . . . . . . . Haiming Jin and Lu Su

91

Incentive Mechanism Design for Mobile Crowdsourcing Without Verification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 Chao Huang, Haoran Yu, Jianwei Huang, and Randall Berry Part III Key Technical Components: Task Allocation Stable Worker–Task Assignment in Mobile Crowdsensing Applications . . 143 Fatih Yucel, Murat Yuksel, and Eyuphan Bulut Spatiotemporal Task Allocation in Mobile Crowdsensing . . . . . . . . . . . . . . . . . . . 163 Honglong Chen, Guoqi Ma, and Yang Huang Part IV Key Technical Components: Data Inference Joint Data Collection and Truth Inference in Spatial Crowdsourcing . . . . . 193 Xiong Wang

xi

xii

Contents

Cost-Quality Aware Compressive Mobile Crowdsensing . . . . . . . . . . . . . . . . . . . . 225 Yong Zhao, Zhengqiu Zhu, and Bin Chen Part V Key Technical Components: Security and Privacy Information Integrity in Participatory Crowd-Sensing via Robust Trust Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 Shameek Bhattacharjee and Sajal K. Das AI-Driven Attack Modeling and Defense Strategies in Mobile Crowdsensing: A Special Case Study on Fake Tasks. . . . . . . . . . . . . . . . . . . . . . . . . 275 Didem Cicek, Murat Simsek, and Burak Kantarci Traceable and Secure Data Sharing in Mobile Crowdsensing . . . . . . . . . . . . . . 299 Jinwen Liang and Song Guo User Privacy Protection in MCS: Threats, Solutions, and Open Issues. . . . 321 Zhibo Wang, Xiaoyi Pang, Peng Sun, and Jiahui Hu Part VI

Applications

Crowdsourcing Through TinyML as a Way to Engage End-Users in IoT Solutions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359 Pietro Manzoni, Marco Zennaro, Fredrik Ahlgren, Tobias Olsson, and Catia Prandi Health Crowd Sensing and Computing: From Crowdsourced Digital Health Footprints to Population Health Intelligence . . . . . . . . . . . . . . . . 387 Jiangtao Wang, Long Chen, and Xu Wang Crowdsourcing Applications and Techniques in Computer Vision . . . . . . . . . 409 Miloš Stojmenovi´c Mobile Crowdsourcing Task Offloading on Social Collaboration Networks: An Empirical Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433 Liang Wang, Yong Cheng, Dingqi Yang, Haixing Xu, Xueqing Wang, Bin Guo, and Zhiwen Yu

Part I

Introduction

Crowdsourcing as a Future Collaborative Computing Paradigm Jie Wu, Chao Song, and Wei Chang

1 Definition and History Crowdsourcing is a word coming from the combination of “crowd” and “outsourcing,” particularly through the Internet. Crowdsourcing coordinates a group of people called crowds to perform small jobs that solve problems that computer systems or a single user could not easily solve. Presently, crowdsourcing techniques are the key to turning smartphone sensing into a powerful tool and leveraging massive user engagement to gather data and perform required tasks. Advantages of crowdsourcing include, but are not limited to, reduced costs, higher efficiency, more flexibility, higher quality, better scalability, and more diversity. Crowdsourcing applications include virtual labor markets, tournament crowdsourcing, open collaboration, and others, including data donation [118]. According to the literature, several definitions of crowdsourcing have been proposed [7, 12, 14, 57]. Howe in [44] introduces the first definition of crowdsourcing, which represents the activity of an organization in outsourcing its tasks previously performed by employees to an external crowd of persons. Various incentive mechanisms have been proposed in crowdsourcing. Several crowdsourcing studies used Amazon’s Mechanical Turk [4] and its associated incentive mechanisms to

J. Wu () Department of Computer and Information Sciences, Temple University, Philadelphia, PA, U.S.A. e-mail: [email protected] C. Song School of Computer Science and Engineering, University of Electronic Science and Technology of China, Chengdu, China e-mail: [email protected] W. Chang Department of Computer Science, Saint Joseph’s University, Philadelphia, PA, USA e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 J. Wu, E. Wang (eds.), Mobile Crowdsourcing, Wireless Networks, https://doi.org/10.1007/978-3-031-32397-3_1

3

4

J. Wu et al.

solve assigned tasks [51, 85]. Based on eYeKa, there has been a high adoption of crowdsourcing in business (e.g., 85% of the top global brands). The remainder of this section reviews the basic components of crowdsourcing and gives an overview of several historical events associated with crowdsourcing. Section 2 provides several motivating examples of crowdsourcing applications, discusses tasks completed by humans via HPU vs. machines via CPU, and introduces the events of humans vs. machines in chess and Go games. Section 3 offers an overview of crowdsourcing workflow and different types of crowdsourcing. Section 4 lists several popular crowdsourcing platforms. Section 5 looks at several crowdsourcing applications. Section 6 discusses challenging issues with a focus on algorithmic and theoretical aspects. Section 7 examines future opportunities associated with crowdsourcing. Section 8 concludes the chapter by presenting our view on the human vs. machine debate through the lens of HPU vs. CPU.

1.1 HPU and CPU The benefits of crowdsourcing are its inexpensiveness and fast speed. In addition, it supports the notion of the whole (i.e., crowd) being greater than the sum of its parts (i.e., individuals). Crowdsourcing can be viewed as consisting of Human Processing Units (HPUs) (i.e., human brains) that are analogous to CPUs in the traditional computer system [17, 77]. In this case, a small job executed by a worker equates to one instruction to the HPU, and the time of execution equates to the physical time in real life. For certain applications, HPU is more effective than CPU in the following areas: (1) verification and validation, such as image labeling, (2) interpretation and analysis, such as language translation, and (3) surveys, such as social network surveys. The networks formed by the interactions between humans (HPUs) and machines (CPUs) are termed as Human–Machine Networks (HMNs) [91]. HPU as human computation has emerged as a new paradigm of computation [17] relatively recently. In many applications, human computation is a useful complement to computer computation, and it enables tasks, like identifying emotions from human speech, recognizing objects in images, and so on, to become possible or to be done more efficiently. With the help of HPU, many data-intensive applications have been proposed, including (1) crowd-powered databases [31, 68, 80] and useful operators like filtering [79] and max [40], group-by [24], (2) various data processing technologies like image labeling [131], entity resolution [97], and schema matching [134], and (3) combinatorial problems like mining [6] and planning [65].

Crowdsourcing as a Future Collaborative Computing Paradigm

ʕ Feedback ʒ Request Requester

5

ʔ Results ʓ Tasks

Crowdsourcing Platform

Workers

Fig. 1 Components of crowdsourcing

1.2 Basic Components Crowdsourcing involves three components, which are requester, worker, and platform. Figure 1 shows the basic components of crowdsourcing as follows: • Requester: A requester is a person who submits a request with a task to the platform. In Amazon Mechanical Turk (AMT), a requester can publish a Human Intelligence Task (called HIT) with some requirements, such as the price, the time constraint for solving, the expiration deadline, and the qualification requirement. • Worker: Workers in crowdsourcing are people who will perform HIT issued by the requester. Workers either can choose tasks or are assigned to tasks by the platform. In some cases, the qualifications of workers need to meet the requirement in order to be eligible. • Platform: The platform connects the requester and the workers. The main functions of the platform include assigning tasks to appropriate workers, filtering and merging multiple outputs from the assigned workers, setting up the rewards, and privacy projection for both the requester and the workers. Note that the last step of the workflow is the final feedback from the platform back to the requester. Feedback can either be raw results or the processed result of raw results from the workers.

1.3 History Even though crowdsourcing was proposed in 2006, the basic idea of using a crowd to solve a problem was applied in the early eighteenth century, and since then, crowdsourcing appeared in many applications. In 1714, the British government provided .£20,000 to reward those who could solve the Longitude Problem [110], which killed thousands of sailors per year [89]. When a ship was damaged, its accurate location coordinates to help seamen are unavailable to be obtained due to the problem. The Longitude Prize was set up at that time for the public to find a method to measure the longitudinal position of a ship. This event is considered to be the first example of crowdsourcing, and John Harrison, an English carpenter, won the award.

6

J. Wu et al.

The following lists several major events that can be considered crowdsourcing applications [22]. Note that before the invention of the Internet, information gathering in crowdsourcing was done through other means, rather than through the Internet. However, the Internet has lowered barriers to entry for crowdsourcing activities. • The Oxford English Dictionary was written in 1884 based on a large crowd across the country to catalog English words [66]. Overall, 800 volunteers contributed to creating the original Oxford English Dictionary. • Japanese car manufacturer Toyota held a logo designing competition held in 1936 [109] and chose the current logo from close to 30,000 candidate logos. • Wikipedia, started in 2001, is a free content Internet encyclopedia open to the public, and it acquires collective crowd wisdom through crowdsourcing. • TopCoder, a crowdsourcing software development company, was founded in 2001. It offers a platform for collaboration and competition. • Threadless.com, started in 2005, has the members creating their own designs. It is considered to be the first crowdsourcing example in the modern era. • YouTube is an example of crowdsourced entertainment, which was founded in 2005. Similarly, there are many large companies in Fortune 500 relied on crowdsourcing for various business-related tasks [86]. • In 2006, Howe and Robinson were the first to introduce the term “crowdsourcing” in the June issue in Wired. • Brabham published the first scholarly work using the term crowdsourcing in 2008. All the above examples show the power of crowdsourcing and the evolution of its applications.

2 Crowdsourcing Events in Recent History In this section, we introduce four samples of crowdsourcing in recent history as motivation. These failed and successful samples include Help Find Jim Gray, Malaysia Airlines Flight MH 370, DARPA Network Challenges, and Tag Challenges. At the end of this section, we will discuss an important debate on human vs. machine through the lens of HPU vs. CPU.

2.1 Help Find Jim Gray (2007) Jim Gray, Turing Award laureate, went missing when he sailed outside San Francisco Bay in January 2007. Dr. Gray is known most notably for his contribution to several major database and transaction processing systems. After 4 days of extensive searching, the search effort called off the effort. However, private efforts

Crowdsourcing as a Future Collaborative Computing Paradigm

7

for searching for Jim Gray were ongoing for a while with the help of a special Amazon link to the search effort. This private effort involves collecting and reviewing individual satellite images to determine if there are any photos that warrant further investigation. However, all these efforts generated no results.

2.2 Malaysia Airlines Flight MH 370 (2014) Flight MH 370 Malaysia Airlines appeared with a trace during the flight between Manila and Beijing on 8 March 2014. DigitalGlobe, a Colorado-based tech company, immediately posted over 3100 square kilometers of imagery that could be examined by the crowd. Within the first 24 h, thousands of volunteers were viewing 2 million pages of satellite imagery every 10 min. They tagged more than 60,000 objects that have the potential for further examination, and these results were made available to the public to review. However, the crowdsourcing search effect was not fruitful. In fact, the whereabouts of MH 370 remains a mystery.

2.3 DARPA Network Challenges (2009) In 2009, DARPA, a US military agency, issued an interesting challenging problem for the public to solve. It called on participating groups to find the locations of 10 red balloons scattered in the air around the country (i.e., the USA). A total prize of $40,000 would go to the first participating team to find all balloons. This purpose is to test the effectiveness of social networking (including crowdsourcing) and webbased technologies that can complete a time-critical large-scale task. The MIT team won the competition by finding all 10 balloons in under 9 h with the help of social networking. The winning team adopted a special incentive mechanism called multi-level marketing to recruit participants, with the prize money ($4000 per balloon) to be distributed in the chain of participants leading to an identified balloon, with $2000 allocated to the person who spots the balloon, $1000 (half of the remaining $2000) to the person who recruited the winner in the chain, and so on. For a short chain, the leftover funds will be donated.

2.4 Tag Challenges (2012) The Tag Challenge in 2012 extends the idea of the DARPA Network Challenge, but in a much harder way. The objective of each participating team was to be the first to locate and photograph five volunteer “suspects” in five different cities in the world: New York City, Washington DC, Stockholm, London, and Bratislava (capital

8

J. Wu et al.

in Slovakia). Again, social networking and crowdsourcing play an important role; the challenge is how information propagates through social networks and what it takes for a message to spread out widely. The winning team from UCSD, which found 3 out of 5 individuals, included one team member from the winning team of the DARPA Network Challenge. The winning team adopted a unique incentive mechanism for the award distribution: recruiters of the first 2000 recruits will receive $1 for each recruit. The success of the UCSD team gave some interesting insights into designing incentive mechanisms for crowdsourcing. The question is which factor plays a dominating role: money, personal satisfaction, or justice under a shared belief?

2.5 Kasparov vs. IBM Deep Blue (1997) If HPU refers to human intelligence and capability, while CPU refers to computer intelligence and capability, which one will dominate in the future? This is a rather philosophical question that goes beyond crowdsourcing through people vs. problemsolving using computers. We focus first on a narrow field of human vs. machine in chess and Go games. In May 1997, Deep Blue, a computer built by IBM, beat the then world-chess champion Garry Kasparov. If AI could beat the world’s sharpest chess mind, it seemed that AI would overtake humans in everything. Still, some people felt that the chess game is still not sophisticated enough and AI may find it hard to beat humans in more complex games, such as the Go game. However, in October 2015, DeepMind technologies’ AlphaGo became the first AI program to beat a professional Go player on a full-sized board. In 2017, AlphaGo beat the then No. 1 player in the world. In fact, the self-taught AlphaZero, without using any human knowledge, is currently the world’s top player in the Go game.

2.6 A Big Picture: Human vs. Machine Since the historical event of Kasparov vs. IBM Deep Blue in 1997, people introduced the freestyle chess game [50] that allows humans unrestricted use of computers during the games. This freestyle format introduced a powerful idea for the future mix of HPU/CPU and applications. In Freestyle chess tournaments dated 2005, there are four types of players: Grand-master (.>2500 points), Machine (Hydra, then best AI chess machine), Grand-master + machine, and Amateur (.>1500 points) + machine. The winner came from the last group, where an Amateur teamed up with the machine. How can such a team become the champion? Perhaps the amateur knows how to collaborate with the machine. He knew when and how to explore ideas, i.e., when to take a machine’s suggestion and when to ignore it.

Crowdsourcing as a Future Collaborative Computing Paradigm

9

Note that a chess or Go game has a clear set of rules for achieving a welldefined goal. In addition, a player has all the information needed to make the right decision toward the ultimate goal. In many crowdsourcing applications discussed in this chapter, such information is vaguely presented or incomplete.

3 Crowdsourcing Overview We give a quick overview of crowdsourcing basics in terms of workflow in crowdsourcing and three major types of crowdsourcing.

3.1 Workflow of Crowdsourcing As shown in Fig. 2, the workflow of crowdsourcing can be divided into three steps: preparation, execution, and termination [22]. The preparation includes all preparation before a crowdsourcing task is given to the platform to be solved by a group of workers. First, the requester should design the task adequately and calculate the workforce needed. If a task is too complex, a divide-and-conquer approach can be used to partition the task into several small ones. Each small task should be easy to solve and independent of any other small tasks, which means the execution of a small task should not affect any other tasks. Once the requester creates the task suitably, incentives should also be designed properly so that a sufficient number of workers will sign up for the tasks. The execution part starts after the preparation step. A requester needs to find the workers for his/her task. The requester recruits the workers through different crowdsourcing platforms. The selection of the workers is essential for the success of crowdsourcing tasks which depend on the quality of the results performed by these workers. Based on the requirements of the task, the requester via the platform will recruit workers with some skill levels and assign the sub-tasks to them. Termination is the last step of crowdsourcing where various steps are executed to complete the task. After receiving all the results from workers, in addition to

Execution • Design task • Divide task • Incentive scheme

Preparation

Fig. 2 Workflow of crowdsourcing

• Find crowd • Assign task

• Quality control • Aggregation • Output result

Termination

10

J. Wu et al.

passing the raw results directly to the requester, in many cases, the platform refines the raw results by separating the outputs provided by ordinary workers from the outputs provided by experts, with the methods of quality control (such as majority voting). Finally, these outputs are merged and calculated to find the right result for the requester.

3.2 Types of Crowdsourcing There are three major types of crowdsourcing based on the nature of applications. • Virtual labor markets: this crowdsourcing provides a platform where workers can complete tasks for monetary compensation, e.g., Amazon Mechanical Turks. • Tournament crowdsourcing: this crowdsourcing offers a platform for ideas competitions, where usually only the winner is compensated, e.g., CrowdFlower and TopCoder. • Open collaboration: this crowdsourcing serves as a platform for collaboration to complete an assigned task. Typically, it does not provide monetary compensation; people who are willing to help are recruited through social media, e.g., Wikipedia. According to [13], there is no consensus on the scope of crowdsourcing. There are nearly forty different interpretations of crowdsourcing. Here, we focus on the above three types of applications throughout this chapter when we cover the history, platforms, and applications of crowdsourcing.

4 Platform Nowadays, there are many crowdsourcing websites and applications. Essentially, they provide a virtual marketplace where a requester can post a variety of tasks online and workers are able to seek, join, and complete some tasks at their discretion. In general, these websites and applications are called crowdsourcing platforms. This section introduces some important crowdsourcing platforms.

4.1 Amazon Mechanical Turk Among all the existing general-purpose crowdsourcing platforms, Amazon Mechanical Turk (AMT) is the most famous one. It provides business owners and developers with an on-demand workforce at a small monetary cost. The idea of AMT was first mentioned by Venky Harinarayan in an Amazon patent filed on October 12, 2001

Crowdsourcing as a Future Collaborative Computing Paradigm

11

[1, 45]. On November 2, 2005, AMT was launched officially as part of Amazon Web Services (AWS). Since then, it has received worldwide attention, and by the year 2011, workers of AMT are from over 190 countries [56, 99]. AMT creates a flexible and convenient labor force market. According to AMT, each worker can freely select a HIT (i.e., Human Intelligence Task) based on the task’s description, keywords, expiration date, the amount of reward, and time allotment. Some tasks even require a qualification test before selecting them. Workers are able to find their submission status and rewards once they complete the assigned workload. The AMT platform supports various types of tasks. The following list summarizes six commonly used types and corresponding examples: • • • • • •

Content Creation: Composing a description of a local spot Data Collection: Finding and collecting data that satisfies certain features Labeling and Categorization: Interpreting content, such as labeling an image Surveys: Taking certain actions based on the survey Transcription: Writing word descriptions from a given video Verification: Checking the authenticity of the information

The AMT platform is well maintained by Amazon and it is updated continuously. The current version allows requesters to build HITs in three different ways [58]: the web-based user interface, command-line tools, and APIs. AMT is rich in templates, and a requester can easily build his task page by using those templates. By the time of writing this chapter, there is a total of 11 project templates available on AMT.

4.2 Crowd4U Unlike the commercial crowdsourcing platforms, such as AMT, whose internal technique details are hidden from the public, Morishima et al. [71] constructed an open, generic, and non-profit platform called Crowd4U [103]. Crowd4U was first opened to the public in November 2011, and by June 2022, there are over 2 million tasks that have been performed on Crowd4U. One unique feature of Crowd4U [70] is its compatibility with other platforms: via an HTTP-based API, Crowd4U is able to retrieve data from others. Supporting complex crowdsourcing tasks is another feature of Crowd4U. It is achieved by using CyLog, a language that can implement complex logic and dataflows. Since Crowd4U is all-academic open, researchers can test various incentive mechanisms, task assignments, or recruitment strategies on it. Besides conducting research experiments, Crowd4U can also be used as a general-purpose crowdsourcing tool [123], and several interesting crowdsourcing tasks have been done on it, such as identifying the course of tornadoes. Due to all these special characteristics, Crowd4U gets a lot of attention, especially from academia.

12

J. Wu et al.

4.3 gMission gMission [21] is a smartphone-based platform, which focuses on spatial crowdsourcing tasks, e.g., detecting the crowdedness level in a cafeteria. It was designed by a group of researchers from HKUST. In gMission, smartphones are used as sensors, and tasks are bound to specific locations. To accomplish a task, each worker has to physically appear at a specific place so that his smartphone’s internal sensors can collect data. gMission consists of four components: an interface module, a data manager, a function manager, and a quality control module. Each module handles different aspects of the crowdsourcing procedure. The interface module is responsible for the user interface, the data manager stores data and action information, the function manager handles location-sensing tasks’ management, task recommendation, and work allocation, and last but not least, the quality control module verifies locations and controls the life cycle of tasks. Unlike the existing commercial web-based crowdsourcing platforms, such as AMT, gMission is one of the first few attempts at smartphone-based mobile crowdsourcing systems. Its design inspired the implementation of other mobile crowdsourcing platforms.

4.4 UpWork UpWork [115] (formerly called Elance-oDesk) is another well-known crowdsourcing platform. There are a few unique features that distinguish it from AMT. First, UpWork was designed to support job hunting and remote collaboration. A requester can interview, hire, and work with freelancers. Second, unlike AMT, where a task typically consists of a collection of questions, each task in UpWork is indecomposable. Third, UpWork uses a unique payment strategy by which the platform gets commission from the freelancers: in UpWork, a worker is allowed to set a price and submit proposals for jobs [75]. The platform connects the worker with employers. Once the task is done, the platform cuts a 5% to 20% service fee, depending on the total amount the worker billed the employer. Due to all those features, UpWork has become a very popular online freelancing platform. By 2017, UpWork had 14 million users from 180 countries [116].

4.5 CrowdFlower In 2007, Lukas Biewald and Chris Van Pelt together founded CrowdFlower [105], an aggregator crowdsourcing platform that aims at involving human beings in machine learning. Unlike general-purpose crowdsourcing systems, CrowdFlower is mainly used by students and data scientists. Typical types of CrowdFlower tasks are data

Crowdsourcing as a Future Collaborative Computing Paradigm

13

collection, transcription, sentiment analysis, labeling, etc. CrowdFlower asks the requester to appraise a task. Like UpWork, when the task is done, CrowdFlower takes approximately 20% of the payment as a commission [76]. What makes Crowdflower unique is its recruitment strategy: Crowdflower is partnered with multiple other platforms, and a Crowdflower task can be delegated to the workers on those platforms [25]. However, perhaps because of the delegation and commissions for the related platforms, some workers are discontented with their amount of payment [104].

5 Sample Applications The crowdsourcing platform organizes the crowd workers with assigned tasks and provides services to different applications. In this section, we introduce the typical applications on image processing, commonsense knowledge, smart city, and other science projects, with a focus on software engineering and coding. Finally, we discuss several projects that use a mix of HPU and CPU, i.e., both crowd workers and traditional computer-based solutions.

5.1 Image Processing For some applications, humans are better at labeling than machines [22]. This subsection discusses three applications of image labeling and classification. The work in [130] focuses on real-time applications for searching for target images on smartphones. A requester sends a target photograph with buildings and some candidate photos to the platform and queries which candidate photos contain the same building in the target photograph. The workers receive the task with the target and candidate photographs and answer “yes” or “no” to indicate whether a candidate photograph has the same building as the target photograph. For the challenge of classifying galaxies, astronomers hope to classify millions of galaxies (based on color and shape) in pictures taken by the Hubble Space Telescope [138]. While computers are good at recognizing colors in images, they find difficulty in recognizing shapes. However, recognizing shapes is very simple for humans. Imagine a case where scientists could use crowdsourcing in an entertaining game format and get help from ordinary people to classify these galaxies. This project (called Galaxy Zoo) was the starting point for the Zooniverse [119], a website that connects scientists with research collaborators and has since expanded to many other science projects [137]. Tohme [113] provides a customized interface to find the curb ramps from Google Street View scenes, by the technologies of computer vision (CV) and machine learning. Tohme dynamically schedules the two workflows of human (worker) labeling and computer vision with human (worker) validation according to the

14

J. Wu et al.

prediction of performance. The performance of Tohme has been verified with 1086 Google Street View scenes in four North American cities and 403 recruited workers.

5.2 Commonsense Knowledge In many applications, it is hard for a machine to accomplish a given task, while it is common sense knowledge for human beings. Therefore, it can easily be solved through crowdsourcing [22]. Gwap [95] is a website with games for users to play, and the users collaboratively do the tasks provided in the games. For example, two users as the players are presented with a photograph in the game, and each user inputs a list of tags for that photograph. Thus, the system will feed back the points when one tag of the two users is matched. Therefore, the system can help to label the photograph with tags by the users as workers in crowdsourcing. The technology of CAPTCHA (Completely Automated Public Turing test to tell Computers and Humans Apart) displays the out-of-order sequence of letters for the users to transcribe, in order to distinguish whether the user is human or machine [94]. Moreover, by the method of CAPTCHA, the transcribing of old books can be performed by users around the world. To improve the performance of the CAPTCHA system, Google created and developed a new system reCAPTCHA in 2007. This system utilizes artificial intelligence and helps protect the website from malicious attacks by robots [108].

5.3 Smart City A smart city [120] utilizes different types of information and communication technologies to increase operational efficiency through data collection and information sharing. The information gained from the smart city could help to manage resources and services efficiently to improve the quality of both government services and citizen welfare. Here, we discuss three particular areas that are related to crowdsourcing: traffic monitoring, trajectories of mobile users, and monitoring of air pollution. Waze [117] is a crowdsourcing application that provides services for transportation. In Waze, the drivers share the events (such as accidents or traffic jams) where they meet along their trajectories. In 2013, Digital China [107] also proposed an “Integrated Citizen Service Platform” for the users to publish the events they encounter in the city. In 2021, Didi Chuxing [106] in China proposed a new application called “Long-Distance Eyes,” which also allowed drivers to share traffic events with photographs or short videos to improve the travel plans of users and migrate traffic congestion.

Crowdsourcing as a Future Collaborative Computing Paradigm Fig. 3 Crowdsourcing platform harvests the mobile trajectories

15

Applications Personalized trip recommendation

Pedestrian flow prediction

Recommendation

rediction

Crowdsourcing platform Harvesting trajectories

The crowdsourcing platform harvests the trajectories of mobile users and provides services to the applications of individual users or business institutions, as shown in Fig. 3. In order to improve user experience in the personalized trip recommendation, the work in [39] fully takes advantage of the Foursquare dataset, a crowdsourced large-scale check-in LBSNs dataset, and discovers attractive routes to improve personalized trip recommendation. In detail, apart from POIs, the routes’ connected POIs also attract visitors. These routes have many crowd flows and have high popularity. This kind of route is termed an Attractive Route and brings extra experience to users. Pedestrian flow prediction is used to help the operators make decisions. Moreover, the operators held events (such as sales promotions) to attract nearby crowds. This kind of event is termed a business event. This work in [38] investigates the impact of business events on pedestrian flows from the crowdsourcing trajectories and proposes a model for pedestrian flow prediction. For air pollution problems (for example, PM2.5), many crowdsourcing platforms are proposed to monitor air conditions. Third-Eye [61] is a crowdsensing platform to monitor fine-grained PM2.5, which is developed by BUPT and Microsoft Research Asia. The platform utilizes the photographs taken by the users’ smartphones and identifies PM2.5 levels of air conditioning by deep learning algorithms. The system helps the government to collect information on air conditioning and adopt related strategies for protection.

5.4 Other Science Projects Crowdsourcing finds many applications in science [118] in the areas of astronomy, energy system research, genealogy research, ornithology, and seismology. Here, we discuss its applications in software engineering and coding in the field of computer science. The crowdsourcing platform of software engineering recruits global software engineers, to execute various kinds of software engineering tasks, including require-

16

J. Wu et al.

ments development, detailed design, programming, and testing. Many platforms implement such crowdsourced software engineering, such as TopCoder, uTest, and TestFlight [54, 67]. TopCoder [114] is a famous crowdsourced software engineering company, which has a process model called TopCoder Competition Methodology [67]. The platform decomposes complex software development into multiple sub-tasks. With a waterfall model, the development is divided into phases. After the online competitions with crowd developers, the qualified winning solutions of each development phase are accepted by the platform.

5.5 Mixed HPU and CPU Applications This subsection discusses several applications that make use of a mixed HPU and CPU approach, i.e., crowdsourcing that draws resources from both humans and machines. One of the early adoptions of both HPU and CPU is collaborative crowdsourcing language transition over the web. In general, the requirement of professional translators or the lack of bilingual speakers makes translation difficult and costly. In [5], a three-step cost-effective approach is used to minimize the cost of hiring bilingual speakers: (1) context-sensitive lexical translation by CPU only for initial word-by-word translation, (2) assistive translation by bilingual HPU who know both original and target language for accurate sentence-level transactions, and (3) target synthesis by monolingual HPU for final polishing in the target language. A mix of HPU and CPU has found its use in databases. Databases give incorrect answers when there are missing data or when an understanding of the data semantics is required. In this case, humans can provide inputs for the missing information. In [30], an extended database called CrowdDB is provided that combines both CPU (for normal database functions) and HPU (for abnormal cases). CrowdDB uses SQL and can be operated on two platforms: AMT and custom-made mobile phones. Nowadays, in medical imaging research, AI is the most discussed topic, both in diagnostic and in therapeutic contexts [81, 90]. The mix of HPU and CPU can start with CPU-based diagnoses for initial screening and elimination, followed by HPUbased decisions by medical doctors. AI can also help in precision medicine and risk assessment. However, fully machine-based approaches still have a long way to go for two reasons: (1) unlike a quantitative task, the knowledge of decision-making with medical imaging requires long-term experience and medical training. (2) There are legal implications when fully machine-based diagnostic and therapeutic treatments are adopted for medical imaging.

Crowdsourcing as a Future Collaborative Computing Paradigm

17

6 Algorithmically and Theoretically Challenging Issues This section discusses challenging issues, with a focus on algorithmic and theoretical aspects that are related to computer science. The coverage is by no means complete. Some aspects are not included, such as privacy in crowdsourcing [27, 129].

6.1 Paradigm The tasks to be accomplished can sometimes be complex and difficult to solve, and the requester takes a different approach to solving such the tasks, such as breaking the task down to reduce complexity [22]. There are three typical methods for task decomposition: (1) by sequential implementation, each task is partitioned into several small sub-tasks which are executed sequentially, and the output of a subtask is set as the input of the next sub-task. (2) By parallel implementation, each task is also partitioned into several small sub-tasks which are executed in parallel, and the outputs of these sub-tasks will be aggregated to the final output for the task. (3) By the method of divide and conquer, the original problem of the task is divided into several smaller problems recursively which can be solved easily. According to the results of these solved smaller problems, the original problem can be solved and the final result is generated. Moreover, we will discuss multi-armed bandit methods and incentive mechanisms in crowdsourcing.

6.1.1

Sequential Implementation

In the method of sequential implementation, the original task is partitioned into several small sub-tasks, which are executed sequentially. The platform will select and execute these sub-tasks one by one until the last one is finished and generates the final result. The output of a sub-task is set as the input of the next one, as shown in Fig. 4. For sequential execution of sub-tasks, [10] adopts identify, filter, and extract stages. Hirth et al. [42] first divides the original task into many sub-tasks by a method of the control group. Then, the platform assigns the sub-tasks to the workers

Crowdsourcing Task sub-taskk 1 Fig. 4 Sequential implementation

sub-taskk 2

sub-taskk 3

Final output

18

J. Wu et al.

Fig. 5 Parallel implementation

Sub-task

Sub-task

Sub-task

Final output

for evaluation. At last, the platform obtains the outputs from the workers for the final output. The model in [135] has the sequential stages of training, refinement, and evaluation of the results. The platform in [74] first tags all the photographs of food. Then, each photograph is identified. At last, the qualities of the identified foods are evaluated.

6.1.2

Parallel Implementation

The method of parallel implementation divides a task into several small sub-tasks which are executed in parallel, and the outputs of these sub-tasks will be aggregated to the final output of the task, as shown in Fig. 5. SCRIBE [55] is a system to provide instantaneous captioning for deaf people. The system recruits the captionists as workers from AMT [4] or quickTurk [11]. Each recording is divided into several parts for capturing by different workers. Then, the results are aggregated and sent back to the user by the server. The system in [92] recruits the users to take different photographs at different locations and then aggregates these photographs into a three-dimensional model. Moreover, the systems in [42, 74] also divide the original task into multiple sub-tasks, and the sub-tasks are executed in parallel.

6.1.3

Divide and Conquer

The method of divide and conquer recursively decomposes the problem of the original task into smaller problems until the problem cannot be divided anymore. Each problem can be solved easily, and the outputs of the problems will be aggregated for the final output, as shown in Fig. 6. The method of divide and conquer in Turkomatic [3, 53, 72] has been used in many applications, such as labeling. First, the requester submits an original task to the platform. Due to the complexity of such a task, the platform recursively decomposes it into several smaller tasks which are much easier to be solved than the original one. Then, the recruited workers execute these sub-tasks and feed the outputs back to the platform. At last, the platform aggregates these outputs for the final result of the original task.

Crowdsourcing as a Future Collaborative Computing Paradigm

19

Fig. 6 Divide-and-Conquer implementation Problem 1 Divide and Conquer

Problem 2

Original problem Problem 3

6.2 Multi-Armed Bandit (MAB) A dynamic procurement using multi-armed bandit (MAB) defines the following problem: when a gambler in a casino faces a row of slot machines, he/she needs to decide which one to play, by how many times, and in which order when multiple machines are selected. Suppose a machine has a random reward with a fixed distribution, the objective is to maximize the sum of rewards earned through a series of sequential lever pulls. The total number of pulls is restricted to a given number. Worker recruitment in crowdsourcing resembles the above MAB process with uncertainty regarding the reward from each worker (e.g., the probability of completion or the quality of a worker). Badanidiyuru et al. [8, 9] proposed a worker selection algorithm to maximize the total rewards through a sequence of worker selections, similar to MAB. In some cases, tasks may not even be successfully completed. For example, the task assigned to a worker can be completed with probability. Hassan et al. [41] used the MAB process for task assignment among all the workers to maximize the success rate under a given budget. In another extension, workers are associated with a unique completion cost, but unknown completion quality. The objective of crowdsourcing is to maximize the total completion quality under the given budget constraint [33, 35]. More specifically, the platform would divide the task allocation process into multiple rounds. In each round, the platform will select several workers (i.e., each worker is assigned one task [34]). Here, the platform has to face the dilemma between exploitation and exploration. The term “exploitation” means that the platform would select the users who had good performance in the past, while the term “exploration” indicates that the platform would try some users with not-so-good performance in the past to discover the potentially optimal users who will generate high completion quality in the future. For example, the traffic monitoring application can leverage various network users’ terminals to collect real-time traffic data. However, since the external environment is changing, the data quality for different users is usually unknown in priority.

20

J. Wu et al.

6.3 Incentive Mechanisms Incentive mechanisms are the main motivation to do a task in crowdsourcing. There are two kinds of incentive mechanisms, which are monetary and non-monetary [22]. In the monetary incentive mechanism, a worker selects a task sent from the requester from the platform, and the worker will get a payment when the requester accepts the result of the work. It is usually assumed that workers are rational and will go after tasks that offer more monetary reward [22, 28, 43, 69, 128]. Not all workers are looking for monetary compensation. Among non-monetary incentives, the motive may come from personal satisfaction, such as free contributions to knowledge-sharing sites like Wikipedia [122] and wikiHow [121] and question-and-answer sites like Quora [111], and Stack Overflow [112]. Some incentives are more natural, such as fun and entertainment. Making tasks more entertaining will help to increase user participation and ease the recruitment effort. Like Gwap [95], the platform provides games to attract users for solving various complex tasks. Foldit is a special type of multi-player game used to predict the structure of proteins [23]. Other non-monetary incentives come from social recognition, self-esteem, or honor, such as a $1 award for the first 2000 participants in the winning team of Tag Challenges. Still, other incentives come from actions that benefit a cause that is considered to be just, based on some common briefs or religions, such as in the case of searching for debris of Malaysia Airport MH 370.

7 Opportunities and Future Directions This section discusses some opportunities and future directions of crowdsourcing. It focuses on the traditional model that selects workers for tasks issued by the requester. Issues related to task partition and uncertainty associated with workers in terms of their qualities and honesty will be covered.

7.1 Beyond Simple Workflows The workflow structure controls how a task is completed by a crowd of workers. It influences not only the completion time of a task but also the quality of the results. One of the possible research opportunities in crowdsourcing is to include more complex workflow structures. Most of the existing crowdsourcing systems adopt sequential, iterative, or parallel workflow structures. By blending bottom-up and open processes with top-down organization goals, many difficult tasks can be accomplished by crowdsourcing. For instance, instead of presenting a fixed set of task questions to crowdsourcing workers, HumanGS [78] selects optimal sets

Crowdsourcing as a Future Collaborative Computing Paradigm

21

of questions to minimize the overall cost. It uses a graph search approach to optimally allocate crowdsourcing questions to different workers. Jigsaw Percolation [16] designs a new model to let workers collaboratively solve a puzzle on social networks: each worker is assigned a piece of the puzzle, and acquaintances’ puzzle pieces can be merged if they are compatible. If all pieces eventually merge together, then the puzzle is solved. This model essentially represents the natural dynamic of how compatible ideas and innovations are merged in the real world, and the underlying workflow structure forms a graph. We strongly believe that the research on workflow design will be a fruitful research direction. The readers should consider not only the workflow for task allocations but also the workflow for collecting results. This problem becomes more important especially since many distributed crowdsourcing applications have emerged. In addition, in Crowdsourcing 2.0, human intelligence has been integrated with other advanced technologies, such as artificial intelligence, big data, cybersecurity, and the Internet of Things. The workflow design for those integrated new crowdsourcing applications may also need to be addressed.

7.2 Beyond Simple Worker Selection In crowdsourcing, the quality of the workers affects the results’ accuracy and the tasks’ completion time and costs. Usually, a requester has no choice but to hire more workers and extend the tasks’ expiration time if the existing ones are doing poorly. Therefore, the selection of workers is a crucial step in crowdsourcing. The initial version of crowdsourcing simply assigns tasks to any worker who is willing to participate. However, considering the fact that workers may have distinct accuracy and spend different amounts of time working on the tasks, this simple worker selection mechanism usually cannot optimize the system’s overall rewards (e.g., the completion time of the entire project, costs of hiring workers, or the accuracy of the results.) To design a better worker selection strategy, several attempts have been made, including the MAB-based approaches discussed earlier. However, in the real world, a crowdsourcing platform does not know workers’ reliability a priori. Also, workers may be dishonest and report a higher cost in order to receive a better payment. Gao et al. [36, 126] explored this situation and designed an auction-based combinatorial multi-armed bandit mechanism. This approach can help a crowdsourcing platform select workers and optimize the overall quality achieved under a limited budget. Challenges still remain in crowdsourcing when multiple workers are recruited and their areas of coverage may overlap. Some form of extended MAB beyond superarm [20] may be needed. There are still open issues in MAB-based approaches, including the way to introduce fairness in addition to merit. We believe that more theoretical works about the recruitment strategy of workers are needed in the field of crowdsourcing.

22

J. Wu et al.

7.3 Beyond Independent Workers The majority of the existing crowdsourcing platforms hire independent workers, which results in a common problem that they can only process simple and independent tasks [52]. To resolve the limitation of independent workers, new crowdsourcing systems have been designed. For instance, Social Crowdsourcing (SC) [19] explores the social networks of individual workers, and it outsources tasks to a crowd of socially related workers, instead of a single independent one. In SC, a task is accomplished by repeatedly recruiting new workers through their social connections. In other words, a task’s owner simply outsources the workloads to a few workers and leaves them to further propagate the sub-tasks to their friends, friends of friends, and so on. However, since SC is a fully distributed system, there is no central control. How to estimate the amount of workload that propagated from a worker to his friends is a crucial problem, which significantly influences the completion time of the entire project. In addition, the distributed workload allocation problem is not trivial since there are common friends of some workers. Chang et al. [18] found that the workload allocation strategy should be changed at different stages of the workload propagation, and they proposed an adaptive approach to assign workload to SC workers. Xiao et al. designed an optimal offline method and a greedy online scheme to solve the problem [125]. When shifting from independent workers to teams, many new problems emerge, such as workload allocation, incentive strategy, and task results aggregation from workers. Recently, several novel mechanisms have been designed for social networks-based crowdsourcing. For example, Jiang et al. [46] study how to measure the overall capacity of a group of workers on social networks and propose a new algorithm for team-based task allocation. Wang et al. [98] consider the impacts of selfish workers and design new incentives to promote workers to behave honestly. In short, the research about social networks-based crowdsourcing is a new trend. Since more complicated worker recruitment schemes have been developed, such as hiring a crowd of friends, the workload assignment, result collection, and incentive mechanism should be re-designed completely.

7.4 Beyond Simple Training Since the tasks of crowdsourcing are completed by independent workers, their responses contain large variances in terms of quality [48]. In the real world, some workers are even malicious and may aim to corrupt a crowdsourcing structure [64]. To resolve this issue, many papers have proposed new crowdsourcing models or schemes, such as hiring redundant workers, involving experts [132], designing an assessment for workers’ quality [22], or constructing statistical models to reduce the influence of the variance in responses [37]. Besides those traditional approaches,

Crowdsourcing as a Future Collaborative Computing Paradigm

23

some platforms [32, 60, 96] even provide initial training or tests for the non-expert workers [26, 32]. However, there are two open questions about the training in crowdsourcing. The first question is about the construction of proper training data for a variety of users. The majority of the existing systems only provide a single set of data for all workers. However, people with different backgrounds may need diverse training before participating in some crowdsourcing tasks. In other words, the state-ofthe-art crowdsourcing systems focus only on the tasks themselves but ignore the opportunity to analyze workers, learn their learning needs, and adaptively create different training examples. There are some existing works [29, 93] on classifying workers and aggregating responses or assigning new tasks based on the community feature of workers. However, as far as we are aware, the research on adaptively generating differential tutoring is limited. Instead of simply discarding the responses from workers with lower accuracy, one may explore them to generate more efficient training data for future workers with similar backgrounds. We believe this could be a promising research direction. Secondly, in the field of crowdsourcing, the impacts of subjectivity are underestimated widely. Not all variance in workers’ responses is related to the error. In fact, many of the crowdsourcing questions are subjective to a certain degree, but most people simply regard an uncommon response as mislabeling. The lack of the study of subjectivity further brings in three sub-questions: (1) how to quantitatively measure the subjectivity in crowdsourcing questions and responses, (2) how to design training data to minimize the influence of subjectivity, and (3) how to aggregate subjective responses and produce a more comprehensive result. To date, only a few papers [47, 48, 73] mathematically model the subjectivity in crowdsourcing. It is apparent that a theoretical study of subjectivity is necessary for the field. To conclude, as more and more crowdsourcing platforms provide training and qualification tests to workers, the creation of subjectivity-sensitive and differential training data for various groups of workers is an open question.

7.5 Beyond Simple Interactive Mode Traditional crowdsourcing platforms are web page-based, where individual workers proactively find a task from a web page, accept it, and complete a series of subtasks alone. This kind of interaction makes workers bored and tired soon [82]. In addition, their accuracy rate may drop due to boredom or tiredness. In recent years, the increasing popularity of conversational agents has enabled a more interactive communication mode between workers and crowdsourcing platforms. The Conversational Crowdsourcing platforms [15, 49, 88] adopt some conversational agents to assist the workers. Based on their preferences, many conversational styles are available, such as “High-Involvement,” “High-Considerateness,” casual chatting, or formal description styles [83]. Some agents can even invite workers to take a

24

J. Wu et al.

break after a certain number of tasks. Based on the studies [83, 84], Conversational Crowdsourcing platforms not only enhance the engagement of workers but also improve their retention rate. The appearance of conversational agents in crowdsourcing brings many interesting research problems, ranging from the design of conversational agents to the auto-recognition of workers’ preferred interactive modes. For instance, can we include some social features when designing the agents? Can we let crowdsourcing workers be the agents or machines only? How can initial training exercises be added to the interactions between workers and agents? Is there some way to automatically find out the preferred conversational styles of a new worker based on his/her social profile? In short, conversational crowdsourcing is a newly emerged topic, and there are plenty of open questions.

7.6 AI Applications In recent years, the blooming of AI significantly affects the field of crowdsourcing. Many novel AI-based crowdsourcing applications have emerged. Those applications range from the environment and traffic monitoring to the preservation of data privacy. Crowdsourcing has wide applications in Federated Learning (FL) [59], which is a machine learning (ML) technique that trains an ML algorithm across multiple decentralized devices, i.e., IoT devices of edges in a crowdsourcing application, holding local data samples, without exchanging them directly. One important aspect of FL is the ever increasing demand for privacy-preserving AI [2] and truthfulness for each participant in collective contribution. In the application of sparse crowdsensing, the spatiotemporal matrix of a sensing task has the following spatiotemporal correlations: the adjacent rows (or columns) are linearly dependent. The matrix with spatiotemporal correlations is called a low-rank matrix. Therefore, several data inference methods combining data-driven methods, and unique attributes of a low-rank matrix have been proposed for urban traffic flow [87, 136], air quality [62], humidity and temperature, and road speed [100–102, 127] to solve the problems of sparse data in large-scale datasets. One future direction is to introduce active learning for the process of job selection and worker matching. The objective of sparse crowdsensing is then to dynamically recruit workers to actively collect an optimal subset of more valuable data to speed up the sampling process.

7.7 Crowdsourcing 2.0 The next generation of crowdsourcing seamlessly combines multiple advantageous technologies together and provides more powerful functionalities to solve many complex problems in the real world.

Crowdsourcing as a Future Collaborative Computing Paradigm

25

The emergence of AI, IoT, Blockchain, and advanced sensing and edge computing techniques bring new opportunities to crowdsourcing. Many cross-platform novel applications have been developed. Third-Eye [61] monitors PM2.5 pollution by analyzing the outdoor images taken by citizens’ phones [133] This system integrates AI, Crowdsourcing, and Image processing together to protect the environment. LiFS [124] uses radio signals, sensors, and crowdsourcing to build an indoor floor plan and future provide localization service. As new technologies have been integrated into crowdsourcing, there are more issues to be tackled. Unlike traditional crowdsourcing, the next generation employs both human intelligence and machine computing capability. The workers consist of not only human beings but also a variety of types of sensors. How to efficiently manage those heterogeneous workers, how to allocate and schedule tasks, and how to coordinate and orchestrate them in an appropriate manner become open questions. In addition, the existing crowdsourcing applications usually address specific problems independently, and therefore compatibility inevitably becomes a crucial issue. CrowdOS [63] is an initial attempt at it, but the problem of compatibility is far from sufficiently explored, especially since other technologies are integrating into crowdsourcing.

8 Conclusion In this chapter, we gave a systematic review of crowdsourcing, its definition, history, applications, key challenges, and future research directions. We took a view of human capability and knowledge in crowdsourcing as the Human Processing Unit (HPU) that complements the widespread applications that resort mainly to computers, i.e., the Central Processing Unit (CPU). Our future seems to be moving in the direction of a combination of both HPU and CPU. For example, the current state of the art in medical imaging is a kind of CPU-assisted HPU application, where the CPU plays an important role, but the final decision is still made by the doctors (special HPUs). Through studying the history of chess and Go games between humans and machines, it seems that the future challenges lie in finding methods to combine human intelligence (HPU) and machine intelligence (CPU) to reach new heights. Perhaps, we can take the view of humans racing with machines side by side, rather than humans racing against machines. Acknowledgments The work of Jie Wu was supported in part by NSF grants CPS 2128378, CNS 2107014, CNS 2150152, CNS 1824440, CNS 1828363, and CNS 1757533. The work of Chao Song was supported in part by the National Key R&D Program under grants No. 2021YFB3101302 and No. 2021YFB3101303, the NSFC under grant 62020106013, the S & T Achievements Transformation Demonstration Project of Sichuan No. 2018CC0094, and the Fundamental Research Funds for the Central Universities No. ZYGX2019J075. This chapter is expanded based on Jie Wu’s keynotes at IEEE IPCCC 2014, IEEE APDCM 2017, and EAI CollaborateCom 2021.

26

J. Wu et al.

References 1. https://en.wikipedia.org/wiki/Amazon_Mechanical_ 2. M. Al-Rubaie, J.M. Chang, Privacy-preserving machine learning: threats and solutions. IEEE Secur. Priv. 17(2), 49–58 (2019) 3. A. Amato, A.D. Sappa, A. Fornés, F. Lumbreras, J. Lladós, Divide and conquer: atomizing and parallelizing a task in a mobile crowdsourcing platform, in Proceedings of the 2nd ACM International Workshop on Crowdsourcing for Multimedia (CrowdMM ’13) (2013) 4. Amazon Mechanical Turk. http://www.mturk.com/ [Online] 5. V. Ambati, S. Vogel, J.G. Carbonell, Collaborative workflow for crowdsourcing translation, in Proceedings of the ACM 2012 Conference on Computer Supported Cooperative Work (2012) 6. Y. Amsterdamer, Y. Grossman, T. Milo, P. Senellart, Crowd mining, in Proceedings of the 2013 ACM SIGMOD International Conference on Management of Data (SIGMOD ’13) (2013) 7. E.E. Arolas, F.G.L. de Guevara, Towards an integrated crowdsourcing definition. J. Inf. Sci. 38, 189–200 (2012) 8. A. Badanidiyuru, R. Kleinberg, A. Slivkins, Bandits with knapsacks: dynamic procurement for crowdsourcing, in The 3rd Workshop on Social Computing and User Generated Content, co-located with ACM EC, vol. 9 (Citeseer, Princeton, 2013), p. 2 9. A. Badanidiyuru, R. Kleinberg, A. Slivkins, Bandits with knapsacks. J. ACM (JACM) 65(3), 1–55 (2018) 10. M.S. Bernstein, J. Teevan, S.T. Dumais, D.J. Liebling, E. Horvitz, Direct answers for search queries in the long tail, in Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (2012) 11. J.P. Bigham, C. Jayant, H. Ji, G. Little, A. Miller, R. Miller, R. Miller, A. Tatarowicz, B.A. White, S. White, T. Yeh, Vizwiz: nearly real-time answers to visual questions, in Proceedings of the 23rd annual ACM Symposium on User Interface Software and Technology (2010) 12. D.C. Brabham, Crowdsourcing as a model for problem solving. Convergence 14, 75–90 (2008) 13. D.C. Brabham, Corwdsourcing (the MIT Press Essential Knowledge Series) (The MIT Press, New York, 2013) 14. D.C. Brabham, Crowdsourcing (MIT Press, New York, 2013) 15. L. Bradeško, M. Witbrock, J. Starc, Z. Herga, M. Grobelnik, D. Mladeni´c, Curious cat– mobile, context-aware conversational crowdsourcing knowledge acquisition. ACM Trans. Inf. Syst. 35(4), 1–46 (2017) 16. C.D. Brummitt, S. Chatterjee, P.S. Dey, D. Sivakoff, Jigsaw percolation: what social networks can collaboratively solve a puzzle?, in The Annals of Applied Probability (2015), pp. 2013– 2038 17. C. Cao, J. Tu, Z. Liu, L. Chen, H.V. Jagadish, Tuning crowdsourced human computation, in 2017 IEEE 33rd International Conference on Data Engineering (ICDE) (2017), pp. 1021– 1032 18. W. Chang, J. Wu, Progressive or conservative: Rationally allocate cooperative work in mobile social networks. IEEE Trans. Parallel Distrib. Syst. 26(7), 2020–2035 (2014) 19. W. Chang, J. Wu, Reliability enhanced social crowdsourcing, in 2015 IEEE Global Communications Conference (GLOBECOM) (IEEE, New York, 2015), pp. 1–6 20. W. Chen, Y. Wang, Y. Yuan, Combinatorial multi-armed bandit: general framework and applications, in International Conference on Machine Learning (PMLR, Singapore, 2013), pp. 151–159 21. Z. Chen, R. Fu, Z. Zhao, Z. Liu, L. Xia, L. Chen, P. Cheng, C.C. Cao, Y. Tong, C.J. Zhang, gMission: a general spatial crowdsourcing platform. Proceedings of the VLDB Endowment 7(13), 1629–1632 (2014) 22. A.I. Chittilappilly, L. Chen, S. Amer-Yahia, A survey of general-purpose crowdsourcing techniques. IEEE Trans. Knowl. Data Eng. 28, 2246–2266 (2016)

Crowdsourcing as a Future Collaborative Computing Paradigm

27

23. S. Cooper, F. Khatib, A. Treuille, J. Barbero, J. Lee, M. Beenen, A. Leaver-Fay, D. Baker, Z. Popovic, F. players, Predicting protein structures with a multiplayer online game. Nature 466, 756–760 (2010) 24. S.B. Davidson, S. Khanna, T. Milo, S. Roy, Using the crowd for top-k and group-by queries, in Proceedings of the 16th International Conference on Database Theory (ICDT ’13) (2013) 25. J.C. de Winter, M. Kyriakidis, D. Dodou, R. Happee, Using crowdflower to study the relationship between self-reported violations and traffic accidents. Procedia Manufacturing 3, 2518–2525 (2015) 26. S. Doroudi, E. Kamar, E. Brunskill, E. Horvitz, Toward a learning science for complex crowdsourcing tasks, in Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems (2016), pp. 2623–2634 27. Y. Duan, G. Gao, M. Xiao, J. Wu, A privacy-preserving order dispatch scheme for ride-hailing services, in 2019 IEEE 16th International Conference on Mobile Ad Hoc and Sensor Systems (MASS) (IEEE, New York, 2019), pp. 118–126 28. C. Eickhoff, A. de Vries, How crowdsourcable is your task, in Proceedings of the Workshop on Crowdsourcing for Search and Data Mining (CSDM) at the Fourth ACM International Conference on Web Search and Data Mining (WSDM) (2011), pp. 11–14 29. J. Fan, G. Li, B.C. Ooi, K.l. Tan, J. Feng, iCrowd: An adaptive crowdsourcing framework, in Proceedings of the 2015 ACM SIGMOD International Conference on Management of Data (2015), pp. 1015–1030 30. A. Feng, M.J. Franklin, D. Kossmann, T. Kraska, S. Madden, S. Ramesh, A. Wang, R. Xin, CrowdDB: Query processing with the VLDB crowd. Proc. VLDB Endow. 4, 1387–1390 (2011) 31. M.J. Franklin, D. Kossmann, T. Kraska, S. Ramesh, R. Xin, CrowdDB: answering queries with crowdsourcing, in Proceedings of the 2011 ACM SIGMOD International Conference on Management of Data (SIGMOD ’11) (2011) 32. U. Gadiraju, B. Fetahu, R. Kawase, Training workers for improving performance in crowdsourcing microtasks, in European Conference on Technology Enhanced Learning (Springer, Berlin, 2015), pp. 100–114 33. G. Gao, H. Huang, M. Xiao, J. Wu, Y.E. Sun, Y. Du, Budgeted unknown worker recruitment for heterogeneous crowdsensing using CMAB. IEEE Trans. Mob. Comput. 21(11), 3895– 3911 (2021) 34. G. Gao, J. Wu, Z. Yan, M. Xiao, G. Chen, Unknown worker recruitment with budget and covering constraints for mobile crowdsensing, in 2019 IEEE 25th International Conference on Parallel and Distributed Systems (ICPADS) (IEEE, New York, 2019), pp. 539–547 35. G. Gao, J. Wu, M. Xiao, G. Chen, Combinatorial multi-armed bandit based unknown worker recruitment in heterogeneous crowdsensing, in IEEE INFOCOM 2020-IEEE Conference on Computer Communications (IEEE, New York, 2020), pp. 179–188 36. G. Gao, H. Huang, M. Xiao, J. Wu, Y.E. Sun, S. Zhang, Auction-based combinatorial multiarmed bandit mechanisms with strategic arms, in IEEE INFOCOM 2021-IEEE Conference on Computer Communications (IEEE, New York, 2021), pp. 1–10 37. P. Germain, A. Lacasse, F. Laviolette, M. Marchand, J.F. Roy, Risk bounds for the majority vote: From a PAC-Bayesian analysis to a learning algorithm. arXiv preprint arXiv:1503.08329 (2015) 38. J. Gu, C. Song, W. Jiang, L. Shi, H. Gong, M. Liu, Pedestrian flow prediction with business events, in 2019 15th International Conference on Mobile Ad-Hoc and Sensor Networks (MSN) (IEEE, New York, 2019), pp. 43–48 39. J. Gu, C. Song, W. Jiang, X. Wang, M. Liu, Enhancing personalized trip recommendation with attractive routes, in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34 (2020), pp. 662–669 40. S. Guo, A.G. Parameswaran, H. Garcia-Molina, So who won?: dynamic max discovery with the crowd, in Proceedings of the 2012 ACM SIGMOD International Conference on Management of Data (2012)

28

J. Wu et al.

41. U.U. Hassan, E. Curry, A multi-armed bandit approach to online spatial task assignment, in 2014 IEEE 11th International Conference on Ubiquitous Intelligence and Computing and 2014 IEEE 11th International Conference on Autonomic and Trusted Computing and 2014 IEEE 14th International Conference on Scalable Computing and Communications and Its Associated Workshops (IEEE, New York, 2014), pp. 212–219 42. M. Hirth, T. Hoßfeld, P. Tran-Gia, Analyzing costs and accuracy of validation mechanisms for crowdsourcing platforms. Math. Comput. Model. 57(11–12), 2918–2932 (2013) 43. J.J. Horton, L.B. Chilton, The labor economics of paid crowdsourcing, in Proceedings of the 11th ACM Conference on Electronic Commerce (2010), pp. 209–218 44. J. Howe, The rise of crowdsourcing. https://www.wired.com/2006/06/crowds/ (2006) 45. Hybrid machine/human computing arrangement. U.S. Patent (2001) 46. J. Jiang, B. An, Y. Jiang, C. Zhang, Z. Bu, J. Cao, Group-oriented task allocation for crowdsourcing in social networks. IEEE Trans. Syst. Man Cybern. Syst. 51(7), 4417–4432 (2019) 47. Y. Jin, M. Carman, Y. Zhu, W. Buntine, Distinguishing question subjectivity from difficulty for improved crowdsourcing, in Asian Conference on Machine Learning (PMLR, Singapore, 2018), pp. 192–207 48. Y. Jin, M. Carman, Y. Zhu, Y. Xiang, A technical survey on statistical modelling and design methods for crowdsourcing quality control. Artif. Intell. 287, 103351 (2020) 49. J.Y. Jung, S. Qiu, A. Bozzon, U. Gadiraju, Great chain of agents: The role of metaphorical representation of agents in conversational crowdsourcing, in CHI Conference on Human Factors in Computing Systems (2022), pp. 1–22 50. G. Kasparov, The chess master and the computer. The New York Review of Books 57(2), 16–19 (2010) 51. A. Kittur, E.H. Chi, B. Suh, Crowdsourcing user studies with Mechanical Turk, in Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI) (2008) 52. A. Kittur, B. Smus, S. Khamkar, R.E. Kraut, CrowdForge: Crowdsourcing complex work, in Proceedings of the 24th Annual ACM Symposium on User Interface Software and Technology (2011), pp. 43–52 53. A. Kulkarni, M. Can, B. Hartmann, Collaboratively crowdsourcing workflows with turkomatic, in Proceedings of the ACM 2012 Conference on Computer Supported Cooperative Work (2012), pp. 1003–1012 54. K.R. Lakhani, D.A. Garvin, E. Lonstein, Topcoder (a): Developing software through crowdsourcing, in Harvard Business School General Management Unit Case (610–032) (2010) 55. W. Lasecki, C. Miller, A. Sadilek, A. Abumoussa, D. Borrello, R. Kushalnagar, J. Bigham, Real-time captioning by groups of non-experts, in Proceedings of the 25th Annual ACM Symposium on User Interface Software and Technology (2012), pp. 23–34 56. M. Lease, J. Hullman, J. Bigham, M. Bernstein, J. Kim, W. Lasecki, S. Bakhshi, T. Mitra, R. Miller, Mechanical Turk is not anonymous. Available at SSRN 2228728 (2013) 57. J.M. Leimeister, M.J. Huber, U. Bretschneider, H. Krcmar, Leveraging crowdsourcing: Activation-supporting components for it-based ideas competition. J. Manag. Inf. Syst. 26, 197–224 (2009) 58. G. Li, J. Wang, Y. Zheng, M.J. Franklin, Crowdsourced data management: A survey. IEEE Trans. Knowl. Data Eng. 28(9), 2296–2319 (2016) 59. W.Y.B. Lim, N.C. Luong, D.T. Hoang, Y. Jiao, Y.C. Liang, Q. Yang, D. Niyato, C. Miao, Federated learning in mobile edge networks: a comprehensive survey. IEEE Commun. Surv. Tutorials 22(3), 2031–2063 (2020) 60. C.J. Lintott, K. Schawinski, A. Slosar, K. Land, S. Bamford, D. Thomas, M.J. Raddick, R.C. Nichol, A. Szalay, D. Andreescu, et al. Galaxy zoo: morphologies derived from visual inspection of galaxies from the Sloan digital sky survey. Mon. Not. R. Astron. Soc. 389(3), 1179–1189 (2008) 61. L. Liu, W. Liu, Y. Zheng, H. Ma, C. Zhang, Third-eye: a mobilephone-enabled crowdsensing system for air quality monitoring. Proc. ACM Interact. Mobile Wearable Ubiquitous Technol. 2(1), 1–26 (2018)

Crowdsourcing as a Future Collaborative Computing Paradigm

29

62. T. Liu, Y. Zhu, Y. Yang, F. Ye, alc2 : when active learning meets compressive crowdsensing for urban air pollution monitoring. IEEE Internet Things J. 6(6), 9427–9438 (2019) 63. Y. Liu, Z. Yu, B. Guo, Q. Han, J. Su, J. Liao, CrowdOS: a ubiquitous operating system for crowdsourcing and mobile crowd sensing. IEEE Transactions on Mobile Computing 21(3), 878–894 (2020) 64. J. Liu, M. Li, W.C. Tang, S. Islam, A cyber physical system crowdsourcing inference method based on tempering: an advancement in artificial intelligence algorithms. Wirel. Commun. Mob. Comput. 2021, 1–11 (2021) 65. I. Lotosh, T. Milo, S. Novgorodov, CrowdPlanr: planning made easy with crowd, in Proceedings of the 2013 IEEE 29th International Conference on Data Engineering (ICDE) (2013), pp. 1344–1347 66. A. Lynch, Crowdsourcing is not new—the history of crowd-sourcing (1714 to 2010) (2015). http://blog.designcrowd.com/article/202/crowdsourcing-is-not-new--the-history-ofcrowdsourcing-1714-to-2010 67. K. Mao, L. Capra, M. Harman, Y. Jia, A survey of the use of crowdsourcing in software engineering. J. Syst. Softw. 126, 57–84 (2017) 68. A. Marcus, E. Wu, D.R. Karger, S. Madden, R. Miller, Human-powered sorts and joins. Proc. VLDB Endow. 5, 13–24 (2011) 69. A. Moreno, J.L. de la Rosa, B.K. Szymanski, J.M. Barcenas, Reward system for completing FAQs, in CCIA (2009), pp. 361–370 70. A. Morishima, N. Shinagawa, T. Mitsuishi, H. Aoki, S. Fukusumi, CyLog/Crowd4U: a declarative platform for complex data-centric crowdsourcing. Proc. VLDB Endow. 5(12), 1918–1921 (2012) 71. A. Morishima, S. Amer-Yahia, S.B. Roy, Crowd4u: An initiative for constructing an open academic crowdsourcing network, in Second AAAI Conference on Human Computation and Crowdsourcing (2014) 72. M. Negri, L. Bentivogli, Y. Mehdad, D. Giampiccolo, A. Marchetti, Divide and conquer: crowdsourcing the creation of cross-lingual textual entailment corpora, in Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing (2011), pp. 670–679 73. A. Nguyen, M. Halpern, B. Wallace, M. Lease, Probabilistic modeling for crowdsourcing partially-subjective ratings, in Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, vol. 4 (2016), pp. 149–158 74. J. Noronha, E. Hysen, H. Zhang, K.Z. Gajos, Platemate: crowdsourcing nutritional analysis from food photographs, in Proceedings of the 24th Annual ACM Symposium on User Interface Software and Technology (2011), pp. 1–12 75. https://www.nerdwallet.com/article/finance/how-to-make-money-on-upwork [Online] 76. https://digital.hbs.edu/platform-digit/submission/crowdflower-powering-the-human-sideof-artificial-intelligence/ [Online] 77. M. Pantic, A.S. Pentland, A. Nijholt, T.S. Huang, Human computing and machine understanding of human behavior: a survey, in Proceedings of the 8th International Conference on Multimodal Interfaces (ICMI’06) (2006) 78. A. Parameswaran, A.D. Sarma, H. Garcia-Molina, N. Polyzotis, J. Widom, Human-assisted graph search: it’s okay to ask questions. arXiv preprint arXiv:1103.3102 (2011) 79. A.G. Parameswaran, H. Garcia-Molina, H. Park, N. Polyzotis, A. Ramesh, J. Widom, Crowdscreen: algorithms for filtering data with humans. Proceedings of the 2012 ACM SIGMOD International Conference on Management of Data (2012) 80. A.G. Parameswaran, N. Polyzotis, Answering queries using humans, algorithms and databases, in CIDR (2011) 81. F. Pesapane, M. Codari, F. Sardanelli, Artificial intelligence in medical imaging: threat or opportunity? radiologists again at the forefront of innovation in medicine. European Radiology Experimental 2, 1–10 (2018) 82. S. Qiu, U. Gadiraju, A. Bozzon, Estimating conversational styles in conversational microtask crowdsourcing. Proc. ACM. Hum. Comput. Interact. 4(CSCW1), 1–23 (2020)

30

J. Wu et al.

83. S. Qiu, U. Gadiraju, A. Bozzon, Improving worker engagement through conversational microtask crowdsourcing, in Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems (2020), pp. 1–12 84. S. Qiu, A. Bozzon, M.V. Birk, U. Gadiraju, Using worker avatars to improve microtask crowdsourcing. Proc. ACM. Hum. Comput. Interact. 5(CSCW2), 1–28 (2021) 85. J. Ross, L.C. Irani, M.S. Silberman, A. Zaldivar, B. Tomlinson, Who are the crowdworkers?: shifting demographics in Mechanical Turk, in CHI ’10 Extended Abstracts on Human Factors in Computing Systems (2010) 86. Y. Roth, 11 of the 12 best global brands use creative crowdsourcing [online]. http://yannigroth. com/2012/03/23/xx-of-the-100-best-global-brands-use-creative-crowdsourcing/ (2012) 87. M. Roughan, Y. Zhang, W. Willinger, L. Qiu, Spatio-temporal compressive sensing and internet traffic matrices (extended version). IEEE/ACM Trans. Networking 20(3), 662–676 (2011) 88. P. Shah, D. Hakkani-Tur, B. Liu, G. Tür, Bootstrapping a neural conversational agent with dialogue self-play, crowdsourcing and on-line reinforcement learning, in Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, vol. 3 (Industry Papers) (2018), pp. 41–51 89. G. Smith, H. Rudge-Pickard, Longitude problem. http://crazysquirrel.com/computing/software/vrml/essay/longitude_problem.jspx (2015) 90. X. Tang, The role of artificial intelligence in medical imaging research. BJR Open 2(1), 20190031 (2020) 91. M. Tsvetkova, T. Yasseri, E.T. Meyer, J.B. Pickering, V. Engen, P.W. Walland, M. Lüders, A. Følstad, G.N. Bravos, Understanding human-machine networks. ACM Comput. Surv. (CSUR) 50, 1–35 (2017) 92. K. Tuite, N. Snavely, D.y. Hsiao, N. Tabing, Z. Popovic, Photocity: training experts at large-scale image acquisition through a competitive game, in Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (2011), pp. 1383–1392 93. M. Venanzi, J. Guiver, G. Kazai, P. Kohli, M. Shokouhi, Community-based Bayesian aggregation models for crowdsourcing, in Proceedings of the 23rd International Conference on World Wide Web (2014), pp. 155–164 94. L. Von Ahn, B. Maurer, C. McMillen, D. Abraham, M. Blum, recaptcha: Human-based character recognition via web security measures. Science 321(5895), 1465–1468 (2008) 95. L. von Ahn, Games with a purpose (GWAP). http://www.gwap.com/ (2008) 96. J. Wang, P.G. Ipeirotis, F. Provost, Managing crowdsourcing workers, in The 2011 Winter Conference on Business Intelligence Citeseer, Princeton (2011), pp. 10–12 97. J. Wang, G. Li, T. Kraska, M.J. Franklin, J. Feng, Leveraging transitive relations for crowdsourced joins. ArXiv abs/1408.6916 (2013) 98. W. Wang, Z. He, P. Shi, W. Wu, Y. Jiang, B. An, Z. Hao, B. Chen, Strategic social team crowdsourcing: forming a team of truthful workers for crowdsourcing in social networks. IEEE Trans. Mob. Comput. 18(6), 1419–1432 (2018) 99. X. Wang, X. Qin, M.B. Hosseini, R. Slavin, T.D. Breaux, J. Niu, Guileak: tracing privacy policy claims on user input data for android applications, in Proceedings of the 40th International Conference on Software Engineering (2018), pp. 37–47 100. E. Wang, M. Zhang, X. Cheng, Y. Yang, W. Liu, H. Yu, L. Wang, J. Zhang, Deep learningenabled sparse industrial crowdsensing and prediction. IEEE Trans. Industr. Inform. 17(9), 6170–6181 (2020) 101. E. Wang, M. Zhang, Y. Yang, Y. Xu, J. Wu, Exploiting outlier value effects in sparse urban crowdsensing, in Proceedings of the 2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS) (IEEE, New York, 2021), pp. 1–10 102. E. Wang, M. Zhang, Y. Xu, H. Xiong, Y. Yang, Spatiotemporal fracture data inference in sparse urban crowdsensing, in IEEE INFOCOM 2022-IEEE Conference on Computer Communications (IEEE, New York, 2022), pp. 1499–1508 103. Crowd4u projects. http://crowd4u.org/en/projects [Online]

Crowdsourcing as a Future Collaborative Computing Paradigm

31

104. Crowdflower lawsuit could change crowd labor industry forever. https://tinywork.wordpress. com/2013/07/29/crowdflower/ [Online] 105. Crowdflowers. http://www.crowdflower.com/ [Online] 106. Didi chuxing. https://www.didiglobal.com/ [Online] 107. Digital China. http://www.digitalchina.com/en/ [Online] 108. Do you know what the difference between captcha and recaptcha is. https://datadome.co/botmanagement-protection/captcha-vs-recaptcha-whats-the-difference/ [Online] 109. From “Toyoda” to “Toyota”. http://www.toyota-global.com/showroom/emblem/history/ [Online] 110. The long history of crowdsourcing—and why you’re just now hearing about it. http://www. crowd-source.com/blog/2013/08/the-long-history-of-crowdsourcing-and-why-youre-justnow-hearing-about-it/ [Online] 111. Quora. http://www.quora.com/ [Online] 112. Stack overflow. http://stackoverflow.com [Online] 113. Tohme. https://makeabilitylab.cs.washington.edu/project/tohme/ [Online] 114. Topcoder. https://www.topcoder.com [Online] 115. Upwork. https://www.upwork.com [Online] 116. Upwork—Wikipedia. https://en.wikipedia.org/wiki/Upwork [Online] 117. Waze. https://www.waze.com/ [Online] 118. Wiki: Crowdsourcing. https://en.wikipedia.org/wiki/Crowdsourcing [Online] 119. Wiki galaxy zoo. https://en.wikipedia.org/wiki/Galaxy_Zoo [Online] 120. Wiki: Smart city. https://en.wikipedia.org/wiki/Smart_city [Online] 121. wikihow. https://www.wikihow.com/ [Online] 122. Wikipedia. https://www.wikipedia.org/ [Online] 123. Fusioncomp. Available: http://www.kc.tsukuba.ac.jp/~fusioncomp/index.html (2009) [Online] 124. C. Wu, Z. Yang, Y. Liu, Smartphones based crowdsourcing for indoor localization. IEEE Trans. Mob. Comput. 14(2), 444–457 (2014) 125. M. Xiao, J. Wu, L. Huang, Y. Wang, C. Liu, Multi-task assignment for crowdsensing in mobile social networks, in Proceedings of the 2015 IEEE Conference on Computer Communications (INFOCOM) (IEEE, New York, 2015), pp. 2227–2235 126. M. Xiao, B. An, J. Wang, G. Gao, S. Zhang, J. Wu, CMAB-based reverse auction for unknown worker recruitment in mobile crowdsensing. IEEE Trans. Mob. Comput. 21(10), 3502–3518 (2021) 127. K. Xie, X. Li, X. Wang, G. Xie, J. Wen, D. Zhang, Active sparse mobile crowd sensing based on matrix completion, in Proceedings of the 2019 International Conference on Management of Data (2019), pp. 195–210 128. G. Xintong, W. Hongzhi, Y. Song, G. Hong, Brief survey of crowdsourcing for data mining. Expert Syst. Appl. 41(17), 7987–7994 (2014) 129. X. Xu, Q. Liu, X. Zhang, J. Zhang, L. Qi, W. Dou, A blockchain-powered crowdsourcing method with privacy preservation in mobile environment. IEEE Trans. Comput. Social Syst. 6(6), 1407–1419 (2019) 130. T. Yan, V. Kumar, D. Ganesan, Crowdsearch: exploiting crowds for accurate real-time image search on mobile phones, in Proceedings of the 8th International Conference on Mobile Systems, Applications, and Services (2010), pp. 77–90 131. X.S. Yang, R. Cheng, L. Mo, B. Kao, D.W.L. Cheung, On incentive-based tagging, in Proceedings of the 2013 IEEE 29th International Conference on Data Engineering (ICDE) (2013), pp. 685–696 132. Y. Yang, Q. Bai, Q. Liu, Modeling random guessing and task difficulty for truth inference in crowdsourcing, in Proceedings of the 18th International Conference on Autonomous Agents and Multiagent Systems (AAMAS), vol. 19 (2019), pp. 2288–2290 133. Z. Yu, H. Ma, B. Guo, Z. Yang, Crowdsensing 2.0. Commun. ACM 64, 76–80 (2021) 134. C. Zhang, L. Chen, H.V. Jagadish, C.C. Cao, Reducing uncertainty of schema matching via crowdsourcing. Proc. VLDB Endow. 6, 757–768 (2013)

32

J. Wu et al.

135. S. Zhu, S. Kane, J. Feng, A. Sears, A crowdsourcing quality control model for tasks distributed in parallel, in CHI’12 Extended Abstracts on Human Factors in Computing Systems (2012), pp. 2501–2506 136. Y. Zhu, Z. Li, H. Zhu, M. Li, Q. Zhang, A compressive sensing approach to urban traffic estimation with probe vehicles. IEEE Trans. Mob. Comput. 12(11), 2289–2302 (2012) 137. Zooniverse. https://www.zooniverse.org/projects [Online] 138. Zooniverse: scientific research through crowdsourcing. https://digital.hbs.edu/platform-digit/ submission/zooniverse-scientific-research-through-crowdsourcing/ [Online]

Urban Mobility-Driven Crowdsensing: Recent Advances in Machine Learning Designs and Ubiquitous Applications Suining He and Kang G. Shin

1 Introduction Urban mobility-driven crowdsensing (UMCS) has been drawing considerable attention in recent years [70, 78]. With the rise of mobile and Internet of Things (IoTs) devices [105, 107], mobile crowdsourcing emerged as a typical form of UMCS extensively due mainly to pervasive proliferation of smartphones [78]. With increasing integration of sensors of various mobility systems, such as taxicabs, ridesharing vehicles, public transportation platforms, and emerging unmanned aerial vehicle (UAV), the UMCS has been broadened and advanced in the recent decades. Leveraging these prevalent mobile sensing devices and platforms and “crowds as sensors,” UMCS enables a myriad of urban applications, as illustrated in Fig. 1. A UMCS, like a cyber-physical system (CPS), can leverage the distribution of the tasks and mobility of participants in a city and steers them toward the sensing tasks via strategizing task allocations and incentive payments quantified based on a certain performance requirement (e.g., signal coverage, data quality). For example, city government can conduct timely noise monitoring by leveraging the human crowds’ feedback and smartphone recordings and decides the corresponding measures, such as road network planning and factory management to control the sources and the impacts of the noise. The crowdsensed air quality metrics by mobile devices can serve as an input for a variety of more proactive venting and purification control within a building’s heating, ventilation, and air conditioning (HVAC) system or for the city air quality monitoring and environmental response departments to provide S. He () University of Connecticut, Storrs, CT, USA e-mail: [email protected] K. G. Shin The University of Michigan, Ann Arbor, MI, USA e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 J. Wu, E. Wang (eds.), Mobile Crowdsourcing, Wireless Networks, https://doi.org/10.1007/978-3-031-32397-3_2

33

34

S. He and K. G. Shin

Fig. 1 Ideas of UMCS, which will enable the environment/data monitoring and data collection through the mobility platforms

timely responses. Furthermore, more adaptive urban traffic control can be realized by harvesting the crowds’ Global Positioning System (GPS) location sharing (with individual location privacy preservation). In summary, the potential of crowds and pervasiveness of mobile devices have made the urban mobility-driven crowdsensing increasingly popular and important. We have witnessed the introduction of various urban and mobile crowdsensing systems and platforms [8], such as Crowdsensing Map [12], Waze [89], and Gigwalk [22, 95]. These recent advances in UMCS platforms have led to the creation of numerous urban and mobile crowdsensing applications. Note that the success of UMCS hinges on the interactions across the UMCS platform (or the platform owner) and the crowdsensing participants (or UMCS users), as well as the urban mobility environments (e.g., transportation network infrastructures, mobility platforms, such as public transportation systems, and built environments). To improve the effectiveness of crowdsensing, numerous researchers have recently investigated how to make use of the interactions of participants’ mobility patterns and the crowdsensing platforms as witnessed from various proposals [103, 119]. Furthermore, the prevalence of smartphones, equipped with various radio and environmental sensors, has enabled myriads of interesting mobile applications, of which the crowdsensing has recently become very popular [58, 129]. Driven by the existing efforts, recent progresses, and the remaining technical concerns, this book chapter, as illustrated in Fig. 2, will review the recent advances in response to the following two important aspects within the UMCS designs: • Advancing ML Designs for UMCS: Thanks to urban big data and advances in parallel computing, ML, including deep learning, has been attracting attentions

Urban Mobility-Driven Crowdsensing: Recent Advances in Machine Learning. . .

35

Fig. 2 An overview of the focuses of this book chapter. This chapter focuses on (a) the machine learning advances and (b) expanding the user cases in the UMCS

from various research and industrial sectors. Traditional urban-scale or large-area crowdsensing is often costly, especially for metropolitan areas or large shopping malls. On the other hand, the distribution of crowdsourcing participants over spatial and temporal spaces is uneven and dynamic, leading to sparse and skewed coverage. In order to reduce the total data collection cost, one may need to devise various deep learning approaches and infer the missing and unexplored signal values by leveraging the spatio-temporal correlations of the signals. Furthermore, the quality of the crowdsourced signals is essential for the UMCS platform. Thus, besides the cost concern, deep learning-based approaches have been taken for estimating and maintaining data quality. In addition, such quality estimation also determines the cycles of crowdsourcing, preventing over- and under-sampling. To overcome the quality issues, thanks to the recent advances in urban big data, various ML approaches, including deep learning, have been proposed and studied. This book chapter will further focus on the recent advances in deep learning approaches for UMCS, with the crowdsourced signal map construction and crowd mobility learning as two typical cases, and illustrate the future directions. • Expanding Ubiquitous Use Cases for UMCS: Crowdsensing, particularly the UMCS paradigm, has emerged as an excellent data source and foundation to address various urban monitoring purposes, enabling a ubiquitous, distributed, collaborative, inexpensive, and accurate manner for a myriad of data collection tasks. Even though there already exist relevant technologies, how to expand further use cases of the UMCS remains to be challenging. To this end, this book chapter, from the scope of smart cities, will re-think about the UMCS from two geo-spatial perspectives, i.e., the indoor and urban (outdoor) UMCS applications, since smart cities are expected to revolutionize our view of the world. With the use cases of indoor crowd detection and urban mobility reconfiguration, we envision the UMCS platforms will achieve a very high level of integration, coordination, and cooperation between crowds and various

36

S. He and K. G. Shin

systems (e.g., mobility-aware CPSs), enabling a greater degree of the ambient intelligence. The rest of this book chapter is organized as follows. We first provide an overview of UMCS system designs in Sect. 2. We then review the recent advances in machine (deep) learning for UMCS in Sect. 3, followed by the two interesting use cases of emerging ubiquitous UMCS in Sect. 4. We finally conclude this book chapter in Sect. 5.

2 An Overview of Urban Mobility-Driven Crowdsensing Figure 3 first provides a general overview of the pipeline of a mobility-driven crowdsensing system (UMCS) as follows: 1. Mobility-Driven Task Publishing: The UMCS platform publishes M sensing tasks located in different zones of the site of interests (such as cities or the building’s indoor environments). Let .M denote the set of these published tasks .Ti , i.e.,.M = {T1 , . . . , TM }. Each UMCS is often labeled with its location (e.g., GPS coordinate) or the zone (e.g., if the city map is discretized). Traditionally, each UMCS task should be executed within a time interval (the effective period of the task, i.e., with the beginning and the end timestamps), particularly for the spatio-temporarily varying tasks (e.g., air quality monitoring). This part of UMCS often involves the research questions of (a) how to improve the sensing quality and coverage and (b) how to leverage the spatial and temporal correlations of the mobility systems (such as ride-sharing systems) to enhance the UMCS performance. Fig. 3 An overview of the pipeline of urban mobility-driven crowdsensing (UMCS)

Urban Mobility-Driven Crowdsensing: Recent Advances in Machine Learning. . .

37

2. Crowdsensing Task Requesting: We may consider that N mobile users are willing to participate in this UMCS. These UMCS users (or participants) can be human users with their mobile devices (such as smartphones), vehicles, and other related mobile sensors (e.g., cameras, hand-held air quality sensors). Each user may submit her/his location (with certain privacy-preserving mechanism) and available period to request the tasks from the UMCS platform. UMCS platforms often come with task bidding in order to provide effective incentivization, since relocating from one location to another for data collection can be costly (e.g., transportation transit costs, motorized vehicle gasoline consumption, mobile device battery consumption, and other human labor). Each user can also make her/his bidding for those tasks based on their preference of the tasks or their own spatial or temporal availability, including the tasks they want to perform and the desired payments. 3. Crowdsensing Task Allocation: Based on the requests of mobile users, the UMCS platform allocates these UMCS tasks to the N mobile users (say, to maximize the total sensing utility under the limited budget), while considering the potential incentive of users. A UMCS user (participant), based on his or her preference or availability as well as other UMCS platform designs (e.g., the participant’s sensing reputations), might be allocated with one or more tasks .Ti ’s or no task. Based on the well-studied game theory framework, different auctionbased incentive designs can be further applied in the UMCS task allocation [11]. By considering diverse properties of bidding, the multi-dimensional auction mechanism [87] and the incentive design based on double auction [47] can be further adopted to stimulate both UMCS task requesters and the UMCS participants. Interested readers can refer to these related articles in [78]. 4. Mobility-Driven Crowdsensing Task Execution and Payments: Upon receiving the task allocation, each UMCS user or participant accomplishes her/his allocated task. One may further consider measuring and updating the reliability of the crowdsourced results based on the participants’ UMCS reputations and the uncertainty of user mobility (e.g., spatial and temporal mobility ranges) in the mean time for future task allocation use. After the UMCS participants finish the tasks, they will obtain the sensing payment from the UMCS platform, sometimes called as the UMCS participant’s income. Once the crowdsourced data are harvested, further data analytics can be conducted with various machine learning and deep learning techniques. In the following section (Sect. 3.1), we will discuss these recent advances in machine learning and deep learning, particularly in the cases of signal map crowdsourcing.

3 Advancing Machine Learning Designs for UMCS In this section, we will first overview in Sect. 3.1 the machine learning design advances particularly in the case of crowdsensing signal reconstruction, followed

38

S. He and K. G. Shin

by the recent advances in the crowd mobility learning in Sect. 3.2. For each part, we will examine the contributions and importance of deep learning approaches studied.

3.1 Machine Learning Advances in Crowdsensed Signal Reconstruction Among various UMCS applications, signal map construction has been attracting much attention from both industry and academia due to its importance in the urban and indoor site spectrum monitoring [106, 124, 129], location-based service (LBS) [27, 50, 98], and the wireless network construction [58]. The recent endeavors in the self-driving car development have also stimulated the needs for largescale urban signal map collection and data construction (e.g., LiDAR signals, camera recordings). For instance, for location-based service deployment [20], the gamification integration with the online learning was proposed in [50] to motivate the users in crowdsourcing Wi-Fi signals. Despite the prior advances, many of these existing UMCS studies for signal map construction did not consider inference of the potentially missing signals to reduce the overall sensing accuracy and cost. In addition, the gamification designs in their mobility might not necessarily motivate the UMCS participants to fill the missing signals given the complex user mobility. Furthermore, one key challenge lies in how to jointly account for the interaction of crowdsourcing payment, coverage, and signal quality, thus achieving better crowdsensing performance. UMCS designs should also carefully consider whether the UMCS design is also amendable to those emerging solutions in the radio frequency (RF) signal map construction to further improve their crowdsourcing quality and deployability [52, 58]. Driven by these needs and gaps, machine learning (including the deep learning techniques) has been considered for the signal map construction. Thanks to the recent advances of big data and parallel computing, the ML techniques have demonstrated their capability of (a) learning the spatio-temporal correlations for missing data inference and (b) capturing the spatio-temporal dynamics of UMCS user mobility [32]. Therefore, one may observe the recently proliferating efforts in studying machine and deep learning approaches for UMCS-based signal map construction. We first briefly overview the conventional ML approaches for signal map construction. ML-based algorithms have been considered to assist learning dynamic wireless signals in the complicated environments. To capture the inherent spatial correlations between signals, Sun et al. [86] considered the Gaussian process regression (GPR) model to predict the signal map distributions of the received signal strengths based on the limited training data. Some more recent studies consider compressive sensing [17] for the allocation of crowdsourcing tasks [90]. Awareness of cost and density has also been incorporated in [102] and [25], respectively. However, they did not consider the corresponding incentive design to improve performance [105, 106]. Compressive sensing was studied in [68] for air quality

Urban Mobility-Driven Crowdsensing: Recent Advances in Machine Learning. . .

39

monitoring, which, however, did not consider the correlations among sample points to refine the quality. One research direction related to signal map construction is the sparse location estimation. Specifically, the location of a mobile device can be considered as a sparse recovery problem, aiming at recovering from a small number of signal measurements (i.e., a signal map) based on the compressive sensing theory. For instance, Feng et al. [20] proposed a dual-stage location estimator, with a coarse stage classifying which cluster the targe device should belong to, and a fine-grained stage further leverages the compressive sensing to recover the location estimation. These pioneering studies further motivate the usage of compressive sensing in reconstructing the radio map based on partial RSSI measurements during the offline phase [51, 106]. More recent studies have shown that learning the sparsity within the crowdsensed data will help improve the structural designs of UMCS. The work in [29] designed several approaches to jointly address the challenges in signal sparsity, sensing quality, and UMCS user incentivization. In order to reduce the sensing area (cost) needed for task allocation, the work in [29] proposed BCCS, i.e., Bayesian Compressive CrowdSensing) [44], to estimate the unexplored/missing values. At each sensing time step t, the relationship between the i-th sparsely crowdsourced signal sample, .zit , and those within the entire signal map of the target site (a city or the indoor environment), .s t , is considered as zit = ψ i s t + i ,

.

(1)

where .ψ i is the i-th row (vector) of .()M×N . We let .e = [1 , . . . , M ]T be a 1D vector consisting of M noise elements .i ’s. Then, the relationship between the crowdsourced samples and the signal map can be modeled as .

 t   z M×1 = ()M×N s t N ×1 + (e)M×1 ,

(2)

where the projection matrix is formally given by ⎤ ⎡ ⎤ ψ(1, 1) . . . ψ(1, N) ψ1 ⎥ ⎢ ⎢ ⎥ .. .. .. = ⎣ ... ⎦ = ⎣ ⎦. . . . ψM ψ(M, 1) . . . ψ(M, N) ⎡

()M×N

.

(3)

Unlike previous compressive sensing approaches, this approach not only predicts the signals but also further provides a confidence interval for quality estimation through the Bayesian modeling. Thus, the BCS-based approach provides flexibility in signal inference and the subsequent incentive design. Enabled by the BCS framework, the work in [29] estimates the quality of crowdsensing and determines the incentive distribution map at a lower cost with a signal map crowdsensing framework. BCCS is designed to iteratively determine signal map, task, and incentive distribution. We further illustrate its process in Fig. 4. More specifically, the work in BCCS provides the following two novel designs:

40

S. He and K. G. Shin

Fig. 4 Ideas of BCCS for the signal map crowdsensing. The framework consists of Bayesian compressive sensing for reconstructing the signal map and the probabilistic inference of participation which will help calculate the payment

(a) Joint Bayesian Compressive Sensing: The researchers in [29] have leveraged the latent correlation across different measurement points, i.e., spatial, signal, and temporal dimensions. The BCCS can hence accurately reconstruct from the sparse signals with substantially reduced data collection cost. (b) Probabilistic Incentive Design: BCCS further takes into account the unknown relations between the monetary reward and crowdsensing process. Specifically, BCCS depicts the user participation in a stochastic manner without explicit user cost information and adapts to the participation dynamics with enhanced flexibility. Extensive experimental evaluation based on urban mesh network and indoor wireless local area network datasets have further validated the effectiveness and applicability of BCCS. Compared to the state-of-the-art approaches in learning- and regression-based approaches, BCCS is shown to significantly reduce the payment and iteration costs, while maintaining higher missing value estimation accuracy. BCCS is considered a general UMCS framework that can be applied in various emerging crowdsourced signal map construction scenarios [58, 94, 130] beyond utilizing the wireless signals. Another strength is shown in the stochastic modeling. Compared to the existing traditional compressive sensing techniques [25, 77, 90, 91, 102] which provide only a single point estimate, the BCCS framework augments it with a full posterior probability density, showing the confidence level. These error estimates can be further used to improve crowdsensing and steer the sensing cycles, and the approach is experimentally compared with state-of-theart algorithms [74, 90, 111, 112] in missing value inference, which validated the accuracy and deployment efficiency.

Urban Mobility-Driven Crowdsensing: Recent Advances in Machine Learning. . .

41

More recent efforts have been focused on the deep learning techniques due to the explosion of urban mobility data. Jung et al. [49] employed an unsupervised signal map calibration approach that includes a hybridized global–local optimization method. Specifically, the global search algorithm builds the initial model without the reference location (i.e., the location point in a signal map) information. Afterward, a local optimization algorithm further estimates locations. To further reduce the data collection costs, methods such as semi-supervised learning techniques with some labeled crowdsourced fingerprints, including manifold-based learning [85, 131] and Gaussian process regression [42, 97], have been proposed. In addition, Sorour et al. [85] locate real-time collected signal values using manifold alignment. The semi-supervised transfer learning method exploits the inherent spatial correlation of wireless values to reduce calibrating load. With some labeled reference locations, the semi-supervised algorithms achieve signal reconstruction accuracy with lower data collection costs. Furthermore, semi-supervised deep learning approaches based on generative adversarial networks (GANs) [13] and convolutional autoencoders [82] can be further considered to enhance the model learnability in complex signal environments. Driven by these efforts in deep learning model design, Li et al. [59] designed a fine-grained signal map crowdsensing approach using deep spatio-temporal reconstruction networks. This approach leveraged the spatial–temporal residual block and external factor fusion module, where the three-dimensional convolutional layers are considered to learn the signal maps within the time series. A more recent work by Zhao et al. [125] considered reconstructing and updating the signal map in the case of partially measured signal maps through a generative adversarial network (GAN)-based active signal map reconstruction method. The GAN mainly consists of two parts, a generator and a discriminator. The main function of the generator is to generate a fake sample from the input random noise, and the goal of the discriminator is to distinguish the real signal samples from the fake samples generated by the abovementioned generator. This way, the GAN approach can be adopted for the complex signal map crowdsensing given the noisy crowdsourced samples. A key challenge with the existing deep learning approaches for signal map crowdsensing lies in that the recovered signal map often lacks high-frequency details due to the designs of conventional loss functions (e.g., mean squared error or MSE). Despite the advances in model learnability, the signal map reconstruction accuracy of these deep learning approaches might have a significant drop when the measurement environments get significantly changed. In particular, while crowdsourcing can update the radio map online, how to handle mobile device power consumption, sensor estimation noises, and quality concerns of the harvested data remains to be challenging for developing an effective UMCS system based deep learning, which is worth further exploration for the future studies.

42

S. He and K. G. Shin

3.2 Machine Learning Advances in Understanding Crowd Mobility Distributions In addition to signal map construction, machine learning, particularly deep learning, has been widely considered for understanding the mobility patterns of the crowd participants. This is essential to model the distributions of the crowd participants in UMCS, thus enabling more proactive and accurate crowd task allocation, and enhance the performance of the UMCS platform. With the growing need of handling more complex city systems [127], there have been numerous efforts in urban traffic and mobility analytics, enabled by advent of big mobile data science [101, 114, 121] and Internet of Things (IoTs) [29, 72, 105]. These traffic analytics techniques have been further extended to the UMCS scenarios to support large-scale crowdsensing scenarios that are driven by increasing connectivity and exploding data in ubiquitous computing. Among these prior efforts, the mobility modeling in smart transportation [45, 93, 99] has recently attracted significant attention [57, 61]. These efforts have investigated the various conventional time series and statistical feature learning analyses for mobility traffic prediction [7, 21]. For instance, instead of considering location-to-location correlations, the authors of [43] studied different feature learning algorithms for prediction. Predicting the aggregated crowd flows at different locations via grouping them into clusters has been explored [10, 43, 57]. Deep learning, thanks to the big data and advancing computing parallelism, has emerged as a more versatile approach toward more promising traffic analytics [93] rather than implementing “shallow” machine learning model structures in the prior studies [60, 88, 120]. For example, the sequence learning approaches, represented by recurrent neural networks (RNNs) and the long-short-term memory (LSTM) techniques [39], have been shown to effectively learn the scalar-based traffic sequence. Convolutional neural network (CNN) has been further explored for traffic monitoring [71], which similarly models the mobility patterns at the neighborhood regions into scalars. In these studies, CNN locally captures, or pools, the spatial dependencies of geo-spatial zones, thanks to its prior success in processing images or photos [55]. Deep residual network [121] and fusion of CNN with LSTM or RNN [71] have been considered to predict aggregated flows for each zone. However, the image-based formulation may not be easily extended to fine-grained prediction of flows at each individual location. Different from the image-based processing, the advances of geometric signal processing [5] have further enabled the deep graph learning for the non-Euclidean data [100, 104]. Graph data in many real-world applications [41, 100], with variable numbers of both un-ordered nodes and neighbors for each node, make conventional operations like convolutions difficult to apply. To enable graph convolution, various theoretical foundations have been established, including those on spectral graph theory [6], spatial-based aggregation [24], and pooling modules [36]. Despite the differences in notations and approximations, their basic idea all tries to propagate and aggregate the neighbor feature information of nodes in a graph iteratively until convergence.

Urban Mobility-Driven Crowdsensing: Recent Advances in Machine Learning. . .

43

Though advances have been achieved, the existence of close or distant points of interests (PoIs) and complex zone-to-zone ride correlations have been overlooked by the scalar and region-based learning of RNNs and CNNs. The integration of both still could not fully capture the holistic picture of metropolitan traffic [110] due to their scalar-based nature. A more recent work [31] adopts the capsule networks [79] in a novel spatio-temporal manner. This work proposed CAPrice in order to model the crowd mobility patterns for the important UMCS tasks such as ride-sourcing dispatching. CAPrice [31] adopts a novel spatio-temporal ride prediction scheme based on deep capsule neural network, which accurately forecasts future demands/supplies via structural and vectorized capsules—structured groups of neurons [79]. A capsule neural network [37] can derive both semantic and interpretable representations from input images. Compared with conventional CNNs, the neurons inside a capsule are activated for various physical properties of the input images for better instantiation of the objects of interest [38]. The proposed approach in CAPrice takes advantage of this strength in the model and considered formulating the shared taxi or ride-sharing trips into heatmap frames. The neurons in each capsule produce a vector, taking into account essential spatial hierarchies between simple and complex objects in an input image. Considering input mobility distributions. as images, it captures the inherent correlations between pixels (i.e., the city zones) by a novel vectorization structure. A dynamic routing mechanism [79] is then applied in their studies to enhance the prediction accuracy. Compared to scalarbased conventional CNNs, the capsule-based approach has been shown to retrieve more spatial knowledge between locations, leading to better prediction accuracy. Given the accurate demand/supply predictions, CAPrice formulates a joint optimization framework, anticipating prices and subsidies toward incoming riderequests and thus incentivizing drivers more responsively to customers than previous greedy surge-chasing formulations for the UMCS-based applications. The formulation of spatial equilibrium in the UMCS vehicle (re)distribution and long-term ride-request patterns (similar to dynamic pricing [1]) jointly optimizes the distributions of incentive-compatible prices and subsidies for the coming rides. This way, CAPrice can handle the demand–supply imbalances via more responsive driver distribution flows between zones, realizing a more effective UMCS platform design. The UMCS applications can leverage the predictive modeling, such as the capsule network, to comprehensively capture the mobility of the participants, such as the inherent pick-up/drop-off relationship among city zones, to provide more informed decisions about the crowdsourcing participants’ spatial distributions. Integrated with spatio-temporal ride distributions and external ride factors, existing UMCS applications can benefit from the fusion of the deep mobility learning with the crowd incentivization. In more recent UMCS studies, the graph convolutional neural network (GCNN) [53] has attracted attention in formatting datasets as networks (e.g., knowledge graphs and social networks). Based on the spectral graph theory [5], the operation of spectral convolutions on graphs [53] can be formally given as the multiplication of an input signal .x ∈ RN with a graph filter .gθ in the Fourier domain, i.e.,

44

S. He and K. G. Shin

.

gθ  x =

P

p=1

θp Ep ET x =

P

θp Lp x,

(4)

p=1

where .E represents the matrix consisting of eigenvectors from the graph Laplacian −1 −1 T .L, i.e.,.L = IN − D 2 AD 2 = EE . Here, .D is the degree matrix, .A is the adjacency matrix, . is a diagonal matrix of .L’s eigenvalues, .Lp represents its p-th power, and .ET x is the graph Fourier transform. In the context of urban mobility learning, one may consider modeling the traffic data as input signal .x for graph learning. Recently, they have been extended to urban traffic applications, investigating speed prediction for road segments and vehicle flows in [7, 63, 113]. These advances will further gain mobility predictability for the UMCS applications. For instance, GBikes in [33] investigates comprehensive spatio-temporal features via mobility network data analytics, aiming to enable more fine-grained traffic prediction model designs. GBikes differentiates the correlations of nearby stations and quantifies multiple different levels of temporal correlations. To this end, GBikes provides the designs of spatio-temporal graph attention mechanisms that can efficiently capture inter-location flows. GCNN therefore does not rely on sophisticated sequence matching via LSTM/RNN [75] and complicated image convolutions. GBikes formalizes the bike station network (stations as nodes and trips as edges) into a graph with attention upon each station’s neighborhood structure. By incorporating spatio-temporal and multi-level features as well as comprehensive external factors, GBikes captures the complex bike-flow patterns. Station neighbors with stronger correlations are further identified and discriminated by our attention mechanism, leading to fine-grained correlation modeling and accurate bike-flow prediction. Specifically, in the neighborhood aggregation process, the conventional graph convolution often considers assigning a weight upon two neighboring location nodes based on only their degrees. The graph attention in GBikes introduces an additional network structure between neighboring nodes, as illustrated in Fig. 5, and thus more important or correlated neighboring station nodes are assigned with “stronger attentions” and larger weights than others. This way, the

Fig. 5 Illustration of modeling the location networks as a graph for crowd mobility learning

Urban Mobility-Driven Crowdsensing: Recent Advances in Machine Learning. . .

45

locations that are more correlated spatio-temporally can be further differentiated, yielding better flow prediction. The researchers in [33] have conducted comprehensive data analytics for bike station networks from multiple metropolitan cities to design and derive data-driven components and parameters. They studied the spatio-temporal factors, such as spatial station-to-station connections, multi-level temporal inflow and outflow trip correlations, points of interest (POIs), and other external factors, and then derived the corresponding component designs for GBikes. Extensive data analytics and experimental studies have been conducted on over .1.13 × 107 bike trips from three metropolitan bike sharing systems in New York City, Chicago, and Los Angeles. GBikes is shown to outperform state of the arts in terms of prediction accuracy, effectiveness, and robustness given environmental variation. GBikes has also demonstrated fine-grained prediction with short time intervals. Such benefits and strengths can further boost the proactiveness of the existing UMCS applications. In summary, the recent deep learning advances will gain more mobility insights for the emerging UMCS systems on how to account for the uncertainty and dynamics of the UMCS participants (including their transportation modalities) to enhance the performance of UMCS, with important algorithmic implications for future UMCS applications.

4 Expanding Ubiquitous Use Cases for UMCS After discussing the aforementioned advances in machine/deep learning for UMCS, we further discuss the following emerging new use cases for UMCS. In particular, we will first discuss the interesting use cases of indoor crowd detection and group identification in Sect. 4.1. Afterward, we further present the crowdsourced information fusion for urban mobility system reconfiguration in Sect. 4.2.

4.1 Indoor Crowd Detection and Group Identification To aid the future UMCS deployment, here we particularly focus on crowd mobility sensing and learning, i.e., understanding and capturing the human crowd movement, which is the key enabler for UMCS platforms to determine the crowd distribution, and further enhance the efficiency, effective, coverage, and fairness of task allocation. Accurate crowd detection and group identification can also benefit the city planners and other facility management departments in monitoring the urban events [9], analyzing epidemics [115], and other social recommendation based on location-based services [62]. The recent COVID-19 pandemic has also shed the bright light upon the importance of understanding crowd mobility and distribution for the sake of public health. With the abovementioned social and business values

46

S. He and K. G. Shin

(particularly after the outbreak of COVID-19), the global crowd analytics market was valued at $912.68 million in 2020 and is projected to hit $5.7 billion by 2030. The resulting crowd analytics market has grown at a Compound Annual Growth Rate (CAGR) of .24.3% by 2021 [14], with an estimated future CAGR of more than 20% from 2021 to 2030 [15]. The advent of smart cities accompanied by increasing pervasiveness of Internet of Things (IoTs) provides unprecedented capabilities and opportunities to monitor, model, and comprehend the mobility of urban crowds, benefiting both the smart city planners and residents. With the proliferation of big crowd data, IoTs, and deep learning, myriad of deep mobility modeling approaches have been studied for the ubiquitous and mobile computing applications. Lin et al. [64] investigated a context-aware framework in order to find the long-range spatial features and accounted for various location-based attributes to forecast the crowd flow. Jiang et al. [46] proposed an online crowd mobility system that extracts the deep trend from a momentary and short-term observations to predict the future mobility. Zhang et al. [123] introduced a deep neural network architecture based on residual neural networks [26] to analyze the crowd mobility. Among the various mobility patterns explored for urban and social sensing, this chapter focuses on finding the individuals in a certain site of interest moving together on similar paths, namely the crowd flows. Detecting the existence of crowd flows is key to many emerging UMCS-related applications [18, 76, 81, 83, 122], including event surveillance, urban planning, social analysis, recommendation, and consequent commercial promotions. Crowd status can be obtained from pairingbased [65, 73] and location-based sensing techniques [28]. Trajectory mining for group/community discovery [54] has also been studied. Recently, researchers started associating signal modalities to infer the mobility, social, or demographic patterns of crowds [18, 122]. Kjærgaard et al. [54] considered temporal user clustering with different sensor readings. GruMon [81] detects groups mainly based on users’ temporal movement correlation. Conventional indoor crowd mobility detection and identification studies often require location estimation and subsequent trajectory mining. Despite their success in vehicle networking or macro migration tracking, few of them can be fully deployable in the complex crowded indoor environments, where dedicated infrastructures (say, GPS, CCTV/camera, or wireless probing transceivers) for localizing devices are likely non-existent or provide poor accuracy (due to crowds or other reasons). Beyond these prior coarse-grained estimates, a fine-grained augmentation is needed for more pervasive deployment. In some studies, mutual proximity between users can be obtained from device pairing (say, Bluetooth) but may cause privacy risks, especially when they are discoverable by other parties. Furthermore, urban crowd flows are highly dynamic due to many opportunistically encountered users. While different signal modalities and their combination have been considered in mobility analytics, few of them considered fine-grained and hybrid signal feature designs for crowd flows and provided spatio-temporally adaptive models for their fast identification. In particular, a scalable and efficient

Urban Mobility-Driven Crowdsensing: Recent Advances in Machine Learning. . .

47

Fig. 6 Illustration of co-flow detection for crowd mobility analytics. By analyzing the similarities of the signals of different users, the proposed work [30] considered a graph-based approach to estimate the crowds. The detected crowd information can be further utilized for UMCS applications

mechanism is required for urban or spacious indoor settings like large malls or airports. Motivated by ambiance of urban/indoor WLAN infrastructures and geomagnetic anomalies, the researchers in [30] propose CFid, a crowd flow identification system via fine-grained spatio-temporal signal fusion, with the following design features. By leveraging the spatial diversity (particularly along a certain walking path), we associate the Wi-Fi and magnetic features measured from individuals’ smartphones with their sequential/temporal co-presences or co-flow, as illustrated in Fig. 6, without explicitly calibrating, pairing devices, or tracing the locations. CFid advances from the related studies in the following aspects. While most pilot studies focus on single correlation measure in terms of group mobility or signal modalities [48, 81], CFid adopts several comprehensive metrics jointly on spatio-temporal features to detect crowds more effectively. Furthermore, instead of the computationally-expensive static clustering [3, 40] and supervised learning [54, 81] that requires a priori training, CFid adopts the fast graph streaming and clustering framework without extensive model or parameter calibration. WiFi and geomagnetism sensing, due to their ubiquity, have triggered a myriad of mobile apps [40, 80], including location-based service [48, 84] and smartphone sensing [28]. However, few of these studies systematically investigated their fusion potential for crowd flow study. To fill these gaps, the CFid approach is built on several novel signal processing and crowd-related feature extraction techniques to unleash their potentials for fast and accurate crowd flow analytics. The resulting crowd detection results can be further fed to UMCS systems to enable more ubiquitous crowdsourcing applications. Specifically, in CFid, the closeness or spatio-temporal similarities between people in the crowds, who are the device carriers or users, can be efficiently identified by online comparison of fine-grained signal sequences between users, hence enabling fast detection without extensive localization of devices. On the signal patterns derived and fine-grained similarity measures, the proposed work in

48

S. He and K. G. Shin

CFid takes into account the crowd flows as the graph stream, where the individuals as vertices or nodes in a graph are dynamically connected via correlations of their signals. The stream of the generated edges can then be fed and processed efficiently. As these signals can be measured from inertial phones and can be easily sanitized and crowdsourced to a central hub or server, CFid can mitigate individuals’ privacy concerns by regulating pairing or communication with unknown peer devices. By deriving spatio-temporal features from inertially measured smartphone signals (i.e., the Wi-Fi and geomagnetic signals), CFid is amendable to these studies or applications and can serve as a plug-in to these UMCS applications for their more adaptive and pervasive deployment. In summary, prior studies, including CFid, can provide important use case studies on how to further integrate the indoor social sensing for UMCS applications, and the insights can be further integrated with other indoor UMCS applications such as indoor air quality monitoring and landmark crowdsourcing.

4.2 Urban Mobility Reconfiguration with Crowdsourced Information Fusion Another important application of UMCS lies in the urban mobility applications, particularly how to leverage the crowdsourced inputs for reconfiguration of the dynamic urban mobility systems. With the advent of smart cities/communities and IoTs, the urban sharing economy, including the bike sharing, has been evolving very rapidly. These applications often originate from the UMCS design principles. In particular, bike sharing service (BSS) has emerged as a popular and revolutionary platform that changes the people’s urban life. Bike sharing enables the first-/lastmile urban travel to be more economic, greener and healthier than traditional gasoline-engine-powered vehicle riding. City transportation also benefits from an additional network of bike stations connected by the trips with less hassle of traffic planning. The emerging UMCS studies and use cases also align with the trends of emerging urban computing applications. Urban computing [127] aims to improve social life quality under the trend of speedy urbanization. With faster computing, smarter IoTs, and more sensing data, many urban transportation problems have been redefined intelligently and efficiently. UMCS should carefully consider integrating the novel cross-domain knowledge fusion technique [126], unleashing the datadriven and crowdsourcing power to reimagine traditional UMCS problems, such as site (re)configuration for emerging smart mobility systems [16, 19, 35, 109]. Thanks to such an expansion, the global bike sharing market is expected to grow at a CAGR of 21% during 2018–2022 [23]. Experiencing the initial deployment successes and receiving positive feedbacks, many BSS providers have begun expanding their bike sharing service networks. Driven by such a growing need, Divvy bicycle sharing program in Chicago, IL is adding 10,500 new bikes and 175 additional

Urban Mobility-Driven Crowdsensing: Recent Advances in Machine Learning. . .

49

stations over the next 3 years from 2019. Meanwhile, Citi Bike in New York City will embrace another 4000 bikes, 13 stations in the busiest areas, and 2500 docks since 2019. On the other hand, there exist BSS network shrinkages (at a micro- or macro-scale) for financial, event, seasonal, or meteorological reasons. Improving the distributions through UMCS designs will also benefit other applications in the recent boom of urban computing and planning, including the site placement of gas stations [116], ambulance points [127], and electric vehicle charging docks [56], which have been investigated to improve their social and business values. UMCS research provides a new view in expanding the BSS network. One might observe from the existing BSS platforms on crowdsourcing the human suggestions (e.g., station distributions) for expanding and updating the BSS network. We note that such a (re)configuration can be done monthly, seasonally, or annually subject to the urbanization process, profit, cost, and the service provider’s own customization. Recent popularity of BSS has triggered many interesting studies, such as mobility and demand prediction [69, 96, 108, 117], station re-balancing [67], lane planning [2], and trip recommendation and station deployment [66, 69]. However, few of state-of-the-art studies have considered optimizing the (re)configuration of existing BSS network with crowdsourced knowledge, which has been overlooked in the important spatial–temporal modeling for real-time bike demand prediction (including dynamic geographical, meteorological, or seasonal factors) [67, 69, 108]. How to fuse long-term batched station usage [92, 128] with aggregated crowdsourced feedbacks, for periodic network (re)configurations, has not been thoroughly studied in these prior works. In summary, how to reconfigure the BSS network, by integrating the crowdsourced feedback to enhance the performance of the bike sharing stations, remains challenging in the following four perspectives. • The first challenge lies in the data heterogeneity of crowdsourced information inputs. Crowdsourced feedbacks often provide local and fragmented suggestions due to each individual’s limited geographic scope or personal interest or preference (say, close to their home residence), while BSS network (re)configuration largely relies on global knowledge of user mobility and station-to-station dynamics. How to incorporate the local suggestions or comments together is important and should thus be considered carefully for the related UMCS designs. • The second challenge lies in the user side. As all stations are “linked” by users’ trips, their trip tendency might be discouraged by over-crowded or inadequate BSS network placement and ignorance of popular station–station pairs for users’ commute may discourage cyclists, thus lowering bike usage and platform profit. • The third challenges will come from the platform side. Since the web crowds are enabled with large freedom to label locations they want, how to address such naturally noisy/biased crowdsourced inputs is challenging which should be considered by a joint fusion formulation. • The fourth challenge resides within the complexity of BSS reconfiguration. Many external factors may influence the success of (re)configuration [117, 128], including human-built facilities (quality/availability), natural environments (like

50

S. He and K. G. Shin

topography, season or weather [108]), socio-economic or psychological considerations (say, social norms or habits), and utility (cost and travel time). Though it is very challenging to design a complete model, incorporating historical spatialtemporal usages, large-scale crowdsourced preferences, and refined cost metric would be a good way to accommodate these factors. Driven by the above challenges, in contrast to recent approaches to BSS deployment [69, 118], the researchers in [34] propose a generic optimization framework that accommodates both network expansion and reduction using data-driven designs and novel semidefinite programming [4]. CBikes provides a flexible formulation fusing crowdsourced knowledge with historical usage statistics jointly and accounts for mobility interactions of BSS users and the stations, thus adapting much better to complex station correlations. CBikes is comprised of four consecutive layers for computing bike station (re)configuration: input, design, core, and action layers. Specifically, at the input layer, the historical and estimated bike usages with respect to each station, the crowdsourced feedback of station expansion/shrinkage suggestions, as well as predefined costs are collected and delivered to a central server hosted by the city planners. These datasets will be preprocessed and then stored into databases. We note that other practical geographic design concerns or constraints, including the number of service bikes and accessible station deployment areas, can be also inputted by the BSS service provider, processed, and stored into its database. At the design layer, CBikes forms the joint objective functions and integrates map information and station geographic distances into constraints. The focus of CBikes design layer is to develop a generic optimization framework, given the above primary and secondary information. Then, CBikes formulates a joint optimization framework, transforms the input crowdsourced inputs, inter-station trip tendency, and the bike usage as shown in Fig. 7, and solves the joint optimization problem at the core layer, optimizing station sites with respect to predefined map grids. Driven by the results of the action layer, the service provider may (re)place stations and resize their docks. In case results are not satisfactory, the parameters can be tuned interactively for another optimization trial. In summary, these prior studies including CBikes have developed a novel insight for the emerging UMCS systems on how to reconfigure the UMCS platform by fusing the valuable information from crowdsourced human inputs with other existing mobility data and optimization approaches, with practical data-driven and large-scale mobility use cases for future UMCS applications.

Urban Mobility-Driven Crowdsensing: Recent Advances in Machine Learning. . .

51

Fig. 7 Illustration of crowdsourced information fusion for BSS reconfiguration. The framework takes in the station suggestions from the crowdsourced inputs, the inter-station trip tendency, and the historical bike usage and generates the joint difference heatmap characterizing the affinity of the BSS network with the urban city map. The resulting station locations can be used to relocate or update the stations

5 Conclusion In this chapter, we have reviewed the recent advances in the Urban Mobilitydriven CrowdSensing (UMCS) that has been widely adopted for various urban and ubiquitous applications, including ubiquitous event monitoring, urban planning, and smart transportation system. In particular, this chapter has examined and analyzed the recent advances and application of UMCS machine learning algorithms and emerging use cases in indoor and urban environments. In particular, we have first overviewed the recent advances of ML techniques and algorithms for crowdsensing signal reconstruction and mobility learning to understand the importance of signal learning and mobility characterization for UMCS. We have reviewed the emerging applications for UMCS, including the indoor crowd detection and urban mobility system reconfiguration. For each category, we have identified the strengths and weaknesses of the related studies and summarized future research directions, which will serve as a guideline for new researchers and practitioners in this emerging and interesting research field.

References 1. E.M. Azevedo, E.G. Weyl, Matching markets in the digital age. Science 352(6289), 1056– 1057 (2016) 2. J. Bao, T. He, S. Ruan, Y. Li, Y. Zheng, Planning bike lanes based on sharing-bikes’ trajectories, in Proceeding of the ACM KDD (2017), pp. 1377–1386

52

S. He and K. G. Shin

3. S.P. Borgatti, M.G. Everett, A graph-theoretic perspective on centrality. Soc. Networks 28(4), 466–484 (2006) 4. S. Boyd, L. Vandenberghe, Convex Optimization (Cambridge University Press, Cambridge, 2004) 5. M.M. Bronstein, J. Bruna, Y. LeCun, A. Szlam, P. Vandergheynst, Geometric deep learning: going beyond Euclidean data. IEEE Signal Process. Mag. 34(4), 18–42 (2017) 6. J. Bruna, W. Zaremba, A. Szlam, Y. LeCun, Spectral networks and locally connected networks on graphs, in Proceedings of the ICLR (2013) 7. D. Chai, L. Wang, Q. Yang, Bike flow prediction with multi-graph convolutional networks, in Proceedings of the ACM SIGSPATIAL (2018), pp. 397–400 8. X. Chen, X. Wu, X.-Y. Li, Y. He, Y. Liu, Privacy-preserving high-quality map generation with participatory sensing, in Proceedings of the IEEE INFOCOM (2014) 9. L. Chen, J. Jakubowicz, D. Yang, D. Zhang, G. Pan, Fine-grained urban event detection and characterization based on tensor cofactorization. IEEE Trans. Hum.-Mach. Syst. 47(3), 380– 391 (2016) 10. L. Chen, D. Zhang, L. Wang, D. Yang, et al. Dynamic cluster-based over-demand prediction in bike sharing systems, in Proceedings of the ACM UbiComp (2016), pp. 841–852 11. M.H. Cheung, F. Hou, J. Huang, Make a difference: diversity-driven social mobile crowdsensing, in Proceedings of the IEEE INFOCOM (2017) 12. V. Coric, M. Gruteser, Crowdsensing maps of on-street parking spaces, in Proceedings of the2013 IEEE International Conference on Distributed Computing in Sensor Systems (IEEE, New York, 2013), pp. 115–122 13. A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, A.A. Bharath, Generative adversarial networks: an overview. IEEE Signal Process. Mag. 35(1), 53–65 (2018) 14. Crowd analytics market worth SPSSlashDollar1,142.5 million by 2021 (2018). https://www. marketsandmarkets.com/PressReleases/crowd-analytics.asp 15. Crowd Analytics Market Statistics—Forecast—2030 (2022). https://www. alliedmarketresearch.com/crowd-analytics-market 16. P. DeMaio, Bike-sharing: history, impacts, models of provision, and future. J. Public Transp. 12(4), 3 (2009) 17. D.L. Donoho, Compressed sensing. IEEE Trans. Inf. Theor. 52(4), 1289–1306 (2006) 18. H. Du, Z. Yu, F. Yi, Z. Wang, Q. Han, B. Guo, Recognition of group mobility level and group structure with mobile devices. IEEE Trans. Mob. Comput. 17(4), 884–897 (2018) 19. Z. Fang, L. Huang, A. Wierman, Prices and subsidies in the sharing economy, in Proceedings of the WWW (2017), pp. 53–62 20. C. Feng, W.S.A. Au, S. Valaee, Z. Tan, Received-signal-strength-based indoor positioning using compressive sensing. IEEE Trans. Mob. Comput. 11(12), 1983–1993 (2011) 21. J. Froehlich, J. Neumann, N. Oliver, Sensing and predicting the pulse of the city through shared bicycling, in Proceedings of the IJCAI (2009), pp. 1420–1426 22. Gigwalk app (2017). http://www.gigwalk.com/ 23. Global bike-sharing market 2018–2022—21% cagr projection over the next five years— technavio (2018). https://www.businesswire.com/news/home/20181226005076/en/GlobalBike-sharing-Market-2018-2022-21-CAGR-Projection 24. W. Hamilton, Z. Ying, J. Leskovec, Inductive representation learning on large graphs, in Proceedings of the NIPS (2017), pp. 1024–1034 25. X. Hao, L. Xu, N. D. Lane, X. Liu, T. Moscibroda, Density-aware compressive crowdsensing, in Proceedings of the ACM/IEEE IPSN (2017), pp. 29–39 26. K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in Proceedings of the IEEE CVPR (2016), pp. 770–778 27. S. He, W. Lin, S.H.G. Chan, Indoor localization and automatic fingerprint update with altered AP signals. IEEE Trans. Mob. Comput. 16(7), 1897–1910 (2017) 28. S. He, K.G. Shin, Geomagnetism for smartphone-based indoor localization: challenges, advances, and comparisons. ACM Comput. Surv. (CSUR) 50(6), 97:1–97:37 (2018)

Urban Mobility-Driven Crowdsensing: Recent Advances in Machine Learning. . .

53

29. S. He, K.G. Shin, Steering crowdsourced signal map construction via bayesian compressive sensing, in IEEE INFOCOM 2018—IEEE Conference on Computer Communications (IEEE Press, New York, 2018), pp. 1016–1024 30. S. He, K.G. Shin, Crowd-flow graph construction and identification with spatio-temporal signal feature fusion, in IEEE INFOCOM 2019—IEEE Conference on Computer Communications (2019), pp. 757–765 31. S. He, K.G. Shin, Spatio-temporal adaptive pricing for balancing mobility-on-demand networks. ACM Trans. Intell. Syst. Technol. 10(4), 1–28 (2019) 32. S. He, K.G. Shin, Spatio-temporal capsule-based reinforcement learning for mobility-ondemand network coordination, in Proceedings of the WWW (2019), pp. 2806–2813 33. S. He, K.G. Shin, Towards fine-grained flow forecasting: a graph attention approach for bike sharing systems, in Proceedings of The Web Conference 2020 (WWW ’20) (Association for Computing Machinery, New York, 2020), pp. 88–98 34. S. He, K.G. Shin, Information fusion for (re)configuring bike station networks with crowdsourcing. IEEE Trans. Knowl. Data Eng. 34(2), 736–752 (2022) 35. T. He, J. Bao, S. Ruan, R. Li, Y. Li, H. He, Y. Zheng, Interactive bike lane planning using sharing bikes’ trajectories. IEEE Trans. Knowl. Data Eng. (2019), pp. 1–1 36. M. Henaff, J. Bruna, Y. LeCun, Deep convolutional networks on graph-structured data. arXiv preprint arXiv:1506.05163 (2015) 37. G.E. Hinton, Z. Ghahramani, Y.W. Teh, Learning to parse images, in Proceedings of the NIPS (2000), pp. 463–469 38. G.E. Hinton, A. Krizhevsky, S.D. Wang, Transforming auto-encoders, in Proceedings of the ICANN (2011), pp. 44–51 39. S. Hochreiter, J. Schmidhuber, Long short-term memory. Neural Comput. 9(8), 1735–1780 (1997) 40. H. Hong, C. Luo, M.C. Chan, SocialProbe: understanding social interaction through passive WiFi monitoring, in Proceedings of the MobiQuitous (2016), pp. 94–103 41. J. Hu, C. Guo, B. Yang, C.S. Jensen, Stochastic weight completion for road networks using graph convolutional networks, in Proceedings of the 2019 IEEE 35th International Conference on Data Engineering (ICDE) (2019), pp. 1274–1285 42. B. Huang, Z. Xu, B. Jia, G. Mao, An online radio map update scheme for wifi fingerprintbased localization. IEEE Internet Things J. 6(4), 6909–6918 (2019) 43. P. Hulot, D. Aloise, S.D. Jena, Towards station-level demand prediction for effective rebalancing in bike-sharing systems, in Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2018), pp. 378–386 44. S. Ji, Y. Xue, L. Carin, Bayesian compressive sensing. IEEE Trans. Signal Process. 56(6), 2346–2356 (2008) 45. S. Ji, Y. Zheng, Z. Wang, T. Li, A deep reinforcement learning-enabled dynamic redeployment system for mobile ambulances. Proc. ACM Interact. Mobile Wearable Ubiquitous Technol. 3(1), 15:1–15:20 (2019) 46. R. Jiang, Z. Cai, Z. Wang, C. Yang, Z. Fan, Q. Chen, X. Song, R. Shibasaki, Predicting citywide crowd dynamics at big events: a deep learning system. ACM Trans. Intell. Syst. Technol. 13(2), 1–24 (2022) 47. H. Jin, L. Su, K. Nahrstedt, CENTURION: incentivizing multi-requester mobile crowd sensing, in IEEE INFOCOM 2017-IEEE Conference on Computer Communications (INFOCOM) (IEEE, New York, 2017) 48. J. Jun, Y. Gu, L. Cheng, B. Lu, J. Sun, T. Zhu, J. Niu, Social-Loc: improving indoor localization with social sensing, in Proceedings of the 11th ACM Conference on Embedded Networked Sensor Systems (2013) 49. S.-H. Jung, B.-C. Moon, D. Han, Unsupervised learning for crowdsourced indoor localization in wireless networks. IEEE Trans. Mob. Comput. 15(11), 2892–2906 (2015)

54

S. He and K. G. Shin

50. R. Kawajiri, M. Shimosaka, H. Kashima, Steered crowdsensing: incentive design towards quality-oriented place-centric crowdsensing, in Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing (ACM UbiComp) (2014), pp. 691– 701 51. M.S. Khan, J. Kim, E.H. Lee, S.M. Kim, A received signal strength based localization approach for multiple target nodes via bayesian compressive sensing, in International Multitopic Conference (INMIC) (2019), pp. 1–6 52. Y. Kim, Y. Chon, H. Cha, Mobile crowdsensing framework for a large-scale Wi-Fi fingerprinting system. IEEE Pervasive Comput. 15(3), 58–67 (2016) 53. T.N. Kipf, M. Welling, Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907 (2016) 54. M.B. Kjærgaard, M. Wirz, D. Roggen, G. Tröster, Detecting pedestrian flocks by fusion of multi-modal sensors in mobile phones, in Proceedings of the 2012 ACM Conference on Ubiquitous Computing (ACM UbiComp) (2012), pp. 240–249 55. A. Krizhevsky, I. Sutskever, G.E. Hinton, ImageNet classification with deep convolutional neural networks, in Proceedings of the NIPS (2012), pp. 1097–1105 56. Y. Li, J. Luo, C.Y. Chow, K.L. Chan, Y. Ding, F. Zhang, Growing the charging station network for electric vehicles with trajectory data analytics. In 2015 IEEE 31st International Conference on Data Engineering (IEEE ICDE) (2015), pp. 1376–1387 57. Y. Li, Y. Zheng, H. Zhang, L. Chen, Traffic prediction in a bike-sharing system, in Proceedings of the 23rd SIGSPATIAL International Conference on Advances in Geographic Information Systems (ACM SIGSPATIAL) (2015), pp. 33:1–33:10 58. Z. Li, A. Nika, X. Zhang, Y. Zhu, Y. Yao, B.Y. Zhao, H. Zheng, Identifying value in crowdsourced wireless signal measurements, in Proceedings of the 26th International Conference on World Wide Web (2017), pp. 607–616 59. K. Li, J. Chen, B. Yu, Z. Shen, C. Li, S. He. Supreme: fine-grained radio map reconstruction via spatial-temporal fusion network, in Proceedings of the ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN) (IEEE, New York, 2020), pp. 1–12 60. Y. Liang, Z. Jiang, Y. Zheng, Inferring traffic cascading patterns, in Proceedings of the 25th ACM Sigspatial International Conference on Advances in Geographic Information Systems (ACM SIGSPATIAL) (2017), pp. 2:1–2:10 61. Y. Liang, S. Ke, J. Zhang, X. Yi, Y. Zheng, GeoMAN: Multi-level attention networks for geo-sensory time series prediction, in Proceedings of the IJCAI (2018), pp. 3428–3434 62. K.H. Lim, J. Chan, S. Karunasekera, C. Leckie, Tour recommendation and trip planning using location-based social media: a survey. Knowl. Inf. Syst. 60(3), 1247–1275 (2019) 63. L. Lin, Z. He, S. Peeta, Predicting station-level hourly demand in a large-scale bike-sharing network: a graph convolutional neural network approach. Transportation Research Part C: Emerging Technologies 97, 258–276 (2018) 64. Z. Lin, J. Feng, Z. Lu, Y. Li, D. Jin, Deepstn+: Context-aware spatial-temporal neural network for crowd flow prediction in metropolis, in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 33 (2019), pp. 1020–1027 65. S. Liu, Y. Jiang, A. Striegel, Face-to-face proximity estimation using Bluetooth on smartphones. IEEE Trans. Mob. Comput. 13(4), 811–823 (2014) 66. J. Liu, Q. Li, M. Qu, W. Chen, J. Yang, H. Xiong, H. Zhong, Y. Fu, Station site optimization in bike sharing systems, in 2015 IEEE International Conference on Data Mining (IEEE ICDM) (2015), pp. 883–888 67. J. Liu, L. Sun, W. Chen, H. Xiong, Rebalancing bike sharing systems: a multi-source data smart optimization, in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (ACM KDD) (2016), pp. 1005–1014 68. T. Liu, Y. Zhu, Y. Yang, F. Ye, Incentive design for air pollution monitoring based on compressive crowdsensing, in Proceedings of the 2016 IEEE Global Communications Conference (GLOBECOM) (2016), pp. 1–6

Urban Mobility-Driven Crowdsensing: Recent Advances in Machine Learning. . .

55

69. J. Liu, L. Sun, Q. Li, J. Ming, Y. Liu, H. Xiong, Functional zone based hierarchical demand prediction for bike system expansion, in Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (ACM KDD) (2017), pp. 957–966 70. S. Liu, Z. Zheng, F. Wu, S. Tang, G. Chen, Context-aware data quality estimation in mobile crowdsensing, in IEEE INFOCOM 2017-IEEE Conference on Computer Communications (IEEE INFOCOM) (IEEE, New York, 2017), pp. 1–9 71. X. Ma, Z. Dai, Z. He, J. Ma, Y. Wang, Y. Wang, Learning traffic as images: a deep convolutional neural network for large-scale transportation network speed prediction. Sensors 17(4), 818 (2017) 72. C. Meng, X. Yi, L. Su, J. Gao, Y. Zheng, City-wide traffic volume inference with loop detector data and taxi trajectories, in Proceedings of the 25th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems (ACM SIGSPATIAL) (2017), pp. 1:1–1:10 73. A. Montanari, S. Nawaz, C. Mascolo, K. Sailer, A study of Bluetooth low energy performance for human proximity detection in the workplace, in 2017 IEEE International Conference on Pervasive Computing and Communications (PerCom) (2017) 74. S. Nikitaki, G. Tsagkatakis, P. Tsakalides, Efficient multi-channel signal strength based localization via matrix completion and Bayesian sparse learning. IEEE Trans. Mob. Comput. 14(11), 2244–2256 (2015) 75. Y. Pan, R.C. Zheng, J. Zhang, X. Yao, Predicting bike sharing demand using recurrent neural networks. Procedia Comput. Sci. 147, 562–566 (2019) 76. K.K. Rachuri, C. Mascolo, M. Musolesi, P.J. Rentfrow, SociableSense: exploring the tradeoffs of adaptive sampling and computation offloading for social sensing, in Proceedings of the 17th Annual International Conference on Mobile Computing and Networking (2011) 77. R.K. Rana, C.T. Chou, S.S. Kanhere, N. Bulusu, W. Hu, Ear-phone: an end-to-end participatory urban noise mapping system, in Proceedings of the 9th ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN) (2010), pp. 105–116 78. F. Restuccia, S.K. Das, J. Payton, Incentive mechanisms for participatory sensing: survey and research challenges. ACM Trans. Sens. Netw. (TOSN) 12(2), 13:1–13:40 (2016) 79. S. Sabour, N. Frosst, G.E. Hinton, Dynamic routing between capsules, in Proceedings of the NIPS (2017), pp. 3859–3869 80. P. Sapiezynski, A. Stopczynski, D.K. Wind, J. Leskovec, S. Lehmann, Inferring personto-person proximity using WiFi signals. Proc. ACM Interact. Mobile Wearable Ubiquitous Technol. (IMWUT) 1(2), 24:1–24:20 (2017) 81. R. Sen, Y. Lee, K. Jayarajah, A. Misra, R.K. Balan, GruMon: fast and accurate group monitoring for heterogeneous urban spaces, in Proceedings of the 12th ACM Conference on Embedded Network Sensor Systems (2014), pp. 46–60 82. M.S. Seyfio˘glu, A.M. Özbayo˘glu, S.Z. Gürbüz, Deep convolutional autoencoder for radarbased classification of similar aided and unaided human activities. IEEE Trans. Aerosp. Electron. Syst. 54(4), 1709–1723 (2018) 83. H. Shao, S. Yao, Y. Zhao, C. Zhang, J. Han, L. Kaplan, L. Su, T. Abdelzaher, A constrained maximum likelihood estimator for unguided social sensing, in IEEE INFOCOM 2018-IEEE Conference on Computer Communications (INFOCOM) (2018), pp. 2429–2437 84. Y. Shu, K.G. Shin, T. He, J. Chen, Last-mile navigation using smartphones, in Proceedings of the 21st Annual International Conference on Mobile Computing and Networking (2015) 85. S. Sorour, Y. Lostanlen, S. Valaee, K. Majeed, Joint indoor localization and radio map construction with limited deployment load. IEEE Trans. Mob. Comput. 14(5), 1031–1043 (2014) 86. W. Sun, M. Xue, H. Yu, H. Tang, A. Lin, Augmentation of fingerprints for indoor wifi localization based on gaussian process regression. IEEE Trans. Veh. Technol. 67(11), 10896– 10905 (2018) 87. M. Tang, S. Wang, L. Gao, J. Huang, L. Sun, MOMD: a multi-object multi-dimensional auction for crowdsourced mobile video streaming, in IEEE INFOCOM 2017-IEEE Conference on Computer Communications (2017)

56

S. He and K. G. Shin

88. Y. Tong, Y. Chen, Z. Zhou, L. Chen, J. Wang, Q. Yang, J. Ye, W. Lv, The simpler the better: a unified approach to predicting original taxi demands based on large-scale online platforms, in Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD) (2017), pp. 1653–1662 89. S. Vasserman, M. Feldman, A. Hassidim, Implementing the wisdom of waze, in Proceedings of the Twenty-Fourth International Joint Conference on Artificial Intelligence (2015) 90. L. Wang, D. Zhang, A. Pathak, C. Chen, H. Xiong, D. Yang, Y. Wang, CCS-TA: qualityguaranteed online task allocation in compressive crowdsensing, in Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing (2015), pp. 683–694 91. L. Wang, D. Zhang, Y. Wang, C. Chen, X. Han, A. M’hamed, Sparse mobile crowdsensing: challenges and opportunities. IEEE Commun. Mag. 54(7), 161–167 (2016) 92. X. Wang, G. Lindsey, J.E. Schoner, A. Harrison, Modeling bike share station activity: effects of nearby businesses and jobs on trips to and from stations. J. Urban Plann. Dev. 142(1), 04015001 (2016) 93. D. Wang, W. Cao, J. Li, J. Ye, DeepSD: supply-demand prediction for online car-hailing services using deep neural networks, in 2017 IEEE 33rd International Conference on Data Engineering (ICDE) (2017) 94. J. Wang, N. Tan, J. Luo, S.J. Pan, Woloc: Wifi-only outdoor localization using crowdsensed hotspot labels, in IEEE INFOCOM 2017-IEEE Conference on Computer Communications (IEEE, New York, 2017), pp. 1–9 95. J. Wang, Y. Wang, D. Zhang, F. Wang, H. Xiong, C. Chen, Q. Lv, Z. Qiu, Multi-Task allocation in mobile crowd sensing with individual task quality assurance. IEEE Trans. Mob. Comput. 17(9), 2101–2113 (2018) 96. S. Wang, T. He, D. Zhang, Y. Shu, Y. Liu, Y. Gu, C. Liu, H. Lee, S.H. Son, BRAVO: improving the rebalancing operation in bike sharing with rebalancing range prediction. Proc. ACM Interact. Mobile Wearable Ubiquitous Technol. (IMWUT) 2(1), 44:1–44:22 (2018) 97. X. Wang, X. Wang, S. Mao, J. Zhang, S.C.G. Periaswamy, J. Patton, Indoor radio map construction and localization with deep gaussian processes. IEEE Internet Things J. 7(11), 11238–11249 (2020) 98. C. Wu, Z. Yang, C. Xiao, C. Yang, Y. Liu, M. Liu, Static power of mobile devices: selfupdating radio maps for wireless indoor localization, in 2015 IEEE Conference on Computer Communications (INFOCOM) (2015), pp. 2497–2505 99. G. Wu, Y. Li, J. Bao, Y. Zheng, J. Ye, J. Luo, Human-centric urban transit evaluation and planning, in 2018 IEEE International Conference on Data Mining (ICDM) (2018), pp. 547– 556 100. Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, P.S. Yu, A comprehensive survey on graph neural networks. arXiv preprint arXiv:1901.00596 (2019) 101. X. Xie, F. Zhang, D. Zhang, PrivateHunt: multi-source data-driven dispatching in for-hire vehicle systems. Proc. ACM Interact. Mobile Wearable Ubiquitous Technol. 2(1), 45:1–45:26 (2018) 102. L. Xu, X. Hao, N. D. Lane, X. Liu, T. Moscibroda, More with less: lowering user burden in mobile crowdsourcing through compressive sensing, in Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing (2015), pp. 659–670 103. J. Xu, S. Wang, N. Zhang, F. Yang, X. Shen, Reward or Penalty: Aligning Incentives of Stakeholders in Crowdsourcing. IEEE Trans. Mob. Comput. 18(4), 974–985 (2019) 104. S. Yan, Y. Xiong, D. Lin, Spatial temporal graph convolutional networks for skeleton-based action recognition, in Proceedings of the AAAI Conference on Artificial Intelligence (2018), pp. 7444–7452 105. D. Yang, G. Xue, X. Fang, J. Tang, Crowdsourcing to smartphones: incentive mechanism design for mobile phone sensing, in Proceedings of the 18th Annual International Conference on Mobile Computing and Networking (2012), pp. 173–184

Urban Mobility-Driven Crowdsensing: Recent Advances in Machine Learning. . .

57

106. B. Yang, S. He, S.-H.G. Chan, Updating wireless signal map with Bayesian compressive sensing, in Proceedings of the 19th ACM International Conference on Modeling, Analysis and Simulation of Wireless and Mobile Systems (2016), pp. 310–317 107. D. Yang, G. Xue, X. Fang, J. Tang, Incentive mechanisms for crowdsensing: Crowdsourcing with smartphones. IEEE/ACM Trans. Networking 24(3), 1732–1744 (2016) 108. Z. Yang, J. Hu, Y. Shu, P. Cheng, J. Chen, T. Moscibroda, Mobility modeling and prediction in bike-sharing systems, in Proceedings of the 14th Annual International Conference on Mobile Systems, Applications, and Services (2016), pp. 165–178 109. Z. Yang, J. Chen, J. Hu, Y. Shu, P. Cheng, Mobility modeling and data-driven closed-loop prediction in bike-sharing systems. IEEE Trans. Intell. Transp. Syst. 20(12), 4488–4499 (2019) 110. H. Yao, F. Wu, J. Ke, X. Tang, Y. Jia, S. Lu, P. Gong, J. Ye, Z. Li, Deep multi-view spatialtemporal network for taxi demand prediction, in Proceedings of the AAAI Conference on Artificial Intelligence (2018), pp. 2588–2595 111. H.C. Yen, C.C. Wang, Cross-device Wi-Fi map fusion with Gaussian processes. IEEE Trans. Mob. Comput. 16(1), 44–57 (2017) 112. F. Yin, F. Gunnarsson, Distributed recursive Gaussian Processes for RSS map applied to target tracking. IEEE J. Sel. Top. Sign. Proces. 11(3), 492–503 (2017) 113. B. Yu, H. Yin, Z. Zhu, Spatio-temporal graph convolutional networks: a deep learning framework for traffic forecasting, in Proceedings of the IJCAI (2018), pp. 3634–3640 114. Y. Yuan, D. Zhang, F. Miao, J.A. Stankovic, T. He, G. Pappas, S. Lin, Dynamic integration of heterogeneous transportation modes under disruptive events, in 2018 ACM/IEEE 9th International Conference on Cyber-Physical Systems (ICCPS) (2018), pp. 65–76 115. D. Zeng, Z. Cao, D.B. Neill, Artificial intelligence–enabled public health surveillance– from local detection to global epidemic monitoring and control, in Artificial Intelligence in Medicine (Elsevier, Amsterdam, 2021), pp. 437–453 116. F. Zhang, N.J. Yuan, D. Wilkie, Y. Zheng, X. Xie, Sensing the pulse of urban refueling behavior: a perspective from taxi mobility. ACM Trans. Intell. Syst. Technol. 6(3), 1–23 (2015) 117. J. Zhang, X. Pan, M. Li, P.S. Yu, Bicycle-sharing system analysis and trip prediction, in 2016 17th IEEE International Conference on Mobile Data Management (MDM), vol. 1 (2016), pp. 174–179 118. J. Zhang, X. Pan, M. Li, P.S. Yu, Bicycle-sharing systems expansion: station re-deployment through crowd planning, in Proceedings of the 24th ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems (2016) 119. X. Zhang, Z. Yang, W. Sun, Y. Liu, S. Tang, K. Xing, X. Mao, Incentives for mobile crowd sensing: a survey. IEEE Commun. Surv. Tutorials 18(1), 54–67 (2016) 120. D. Zhang, T. He, F. Zhang, Real-time human mobility modeling with multi-view learning. ACM Trans. Intell. Syst. Technol. 9(3), 1–25 (2017) 121. J. Zhang, Y. Zheng, D. Qi, Deep spatio-temporal residual networks for citywide crowd flows prediction, in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 31 (2017) 122. H. Zhang, W. Du, P. Zhou, M. Li, P. Mohapatra, An acoustic-based encounter profiling system. IEEE Trans. Mob. Comput. 17(8), 1750–1763 (2018) 123. J. Zhang, Y. Zheng, D. Qi, R. Li, X. Yi, T. Li, Predicting citywide crowd flows using deep spatio-temporal residual networks. Artif. Intell. 259, 147–166 (2018) 124. Y. Zhao, W. Li, J. Wu, S. Lu, Quantized conflict graphs for wireless network optimization, in 2015 IEEE Conference on Computer Communications (INFOCOM) (2015), pp. 2218–2226 125. Y. Zhao, C. Liu, K. Zhu, S. Zhang, J. Wu, GSMAC: GAN-based signal map construction with active crowdsourcing. IEEE Trans. Mob. Comput. 22(4), 2190–2204 (2023) 126. Y. Zheng, Methodologies for cross-domain data fusion: an overview. IEEE Trans. Big Data 1(1), 16–34 (2015) 127. Y. Zheng, L. Capra, O. Wolfson, H. Yang. Urban computing: concepts, methodologies, and applications. ACM Trans. Intell. Syst. Technol. (TIST) 5(3), 38:1–38:55 (2014)

58

S. He and K. G. Shin

128. X. Zhou, Understanding spatiotemporal patterns of biking behavior by analyzing massive bike sharing data in Chicago. PloS one 10, 1–20 (2015) 129. X. Zhou, Z. Zhang, G. Wang, X. Yu, B.Y. Zhao, H. Zheng, Practical conflict graphs in the wild. IEEE/ACM Trans. Networking 23(3), 824–835 (2015) 130. B. Zhou, M. Elbadry, R. Gao, F. Ye, BatMapper: acoustic sensing based indoor floor plan construction using smartphones, in Proceedings of the 15th Annual International Conference on Mobile Systems, Applications, and Services (2017), pp. 42–55 131. M. Zhou, Y. Tang, Z. Tian, L. Xie, W. Nie, Robust neighborhood graphing for semi-supervised indoor localization with light-loaded location fingerprinting. IEEE Internet Things J. 5(5), 3378–3387 (2018)

Part II

Key Technical Components: User Recruitment and Incentive Mechanisms

Unknown Worker Recruitment in Mobile Crowdsourcing Mingjun Xiao, Yin Xu, and He Sun

1 Introduction Recent years have witnessed the pervasiveness of smart mobile devices (e.g., smartphones) and wireless communication network (e.g., 5G) in day-to-day life. These smartphones often have powerful capabilities of storage, sensing, computation, and communication. To take full advantage of these mobile computing resources, a new paradigm, called Mobile CrowdSourcing (MCS), is advocated to coordinate a crowd of mobile users equipped with smartphones to cooperatively accomplish some largescale complex tasks, having attracted lots of attention. Nowadays, a great number of MCS applications and systems have been developed, such as information collection (e.g., Waze and OSM), intelligent transportation (e.g., Uber and DiDi), food delivery (e.g., Seamless), micro-tasks (e.g., gMission), and so on [1]. A typical MCS system is composed of a cloud platform, task requesters, and a crowd of workers. Each requester can outsource traditionally costly or timeconsuming tasks to mobile users (a.k.a., workers) via the platform. In general, owing to the various configurations of workers’ smart devices (e.g., storage capacity, camera pixel, CPU frequency, etc.) and their behaviors (e.g., work habit, skill level, work experience, etc.), the completion qualities of diverse workers might be different, even for the same task. Therefore, recruiting suitable workers to maximize the total completion quality as much as possible is one of the most important issues in MCS systems. So far, a wide spectrum of studies has been devoted to designing worker recruitment or task allocation mechanisms [2, 3]. The majority of existing works operate under the assumption that the platform possesses the prior knowledge of workers’ qualities. Only a few studies have investigated the

M. Xiao () · Y. Xu · H. Sun University of Science and Technology of China, Hefei, China e-mail: [email protected]; [email protected]; [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 J. Wu, E. Wang (eds.), Mobile Crowdsourcing, Wireless Networks, https://doi.org/10.1007/978-3-031-32397-3_3

61

62

M. Xiao et al.

circumstance that there is no prior knowledge about workers’ qualities, i.e., the socalled Unknown Worker Recruitment (UWR) problem. For instance, Wu et al. [4] designed a worker selection mechanism based on a modified Thompson sampling strategy. Xiao et al. [5] proposed an incentive mechanism to determine a recruiting strategy so as to maximize the total sensing quality. Yang et al. [6] presented user selection algorithms to find the most informative contributors with unknown costs. Nonetheless, these researchers either ignore the platform’s budget constraint or do not take the workers’ privacy into consideration. In this chapter, we focus on the issue of recruiting unknown workers for MCS systems. Since the qualities of workers are non-public in advance (generally following some unknown distributions), the MCS platform needs to estimate workers’ quality distributions through online learning, i.e., the so-called exploration process. On the other hand, since the platform can acquire a higher valuation of task results (e.g., collected data) from the recruited high-quality workers, it might directly recruit the best group of workers according to the learned knowledge of qualities so as to maximize the total quality, a.k.a., the exploitation process. Hence, there are three major challenges in the UWR problem to be addressed urgently. The first challenge is how to determine an optimal trade-off between the exploration and exploitation processes so that the platform can maximize the total completion quality of tasks. Second, since workers might often spend diverse costs in performing tasks with different importance, the platform needs to determine the best worker group so as to ensure that more important tasks would be accomplished under a given recruitment budget, which involves a weighted set coverage problem with limited budgets. The third challenge is the privacy-preserving issue. Generally, there is competition among workers in MCS systems, and thus each worker is not reluctant to reveal its qualities to others, due to the risk of disclosing some sensitive private information. Thus, we must protect each worker’s individual information from being revealed, making the UMR problem more challenging. To tackle the above challenges, we model the unknown worker recruitment problem as a multi-armed bandit game, where each worker is regarded as an arm, its quality is seen as the corresponding reward, and then recruiting workers is equivalent to pulling arms. Under this model, we propose three unknown worker recruitment schemes in a progressive manner: the UWR scheme, the budget-limited UWR scheme, and the privacy-preserving UWR scheme. More specifically, we extend the widely used Upper Confidence Bound (UCB) strategy to design a general recruitment scheme after a fixed time threshold. Moreover, we adopt the greedy strategy to repeatedly select the worker that has the largest estimated quality value. Based on the UWR scheme, we take the budget constraint and the privacy demand into consideration and then further design the budget-limited UWR scheme and the privacy-preserving UWR scheme. In addition, we analyze the bounds on the regret of each proposed scheme. The major contributions are summarized as follows: 1. We introduce the UWR issues into MCS and propose a UWR scheme based on multi-armed bandit modeling, where an extended UCB bandit strategy is

Unknown Worker Recruitment in Mobile Crowdsourcing

63

designed to balance the exploration and exploitation processes. Furthermore, we also extend UWR scheme to support the budget constraint. 2. We derive the regret bounds of the proposed UWR schemes through the theoretical analysis, both of which achieve the sub-linear regret performances. 3. We propose a privacy-preserving UWR scheme to protect the privacy from being disclosed by using the differential privacy technique. Meanwhile, we also analyze the regret performance of this scheme. The remainder of this chapter is organized as follows. In Sect. 2, we introduce the related works, followed by the system model and the whole workflow in Sect. 3. The detailed UWR design, the budget-limited UWR scheme, and the corresponding theoretical analysis are elaborated in Sect. 4. In Sect. 5, we introduce the privacypreserving UWR scheme in detail. Finally, the conclusion is presented in Sect. 6.

2 Related Work Worker recruitment or task assignment is one of the most important issues in MCS, having attracted great attention from both academia and industry. Here, we present the relevant research about unknown worker recruitment and privacy-preserving unknown worker recruitment schemes as follows. First, many works investigate MCS worker recruitment issues in which workers are generally assumed to have been known by the platform in advance. For example, Ho et al. [7] formalize an online MCS task assignment problem and propose twophase exploration–exploitation offline algorithm to maximize the total benefit that the requester obtains from completed tasks. Xiao et al. [8] propose a deadlinesensitive worker recruitment algorithm for probabilistically collaborative mobile crowdsourcing, which can achieve a logarithmic approximation ratio. Wang et al. [9] propose a personalized task-oriented MCS worker recruitment mechanism in which workers’ preferences and task-worker fitness are taken into account. Wang et al. [10] propose a greedy worker recruitment scheme for self-organized MCS systems to maximize communication probability. Liu et al. [11] propose a group buyingbased MCS auction mechanism where requesters with similar data demands can form a group to share the payment. Hu et al. [12, 13] propose a framework based on reinforcement learning to make a good sequence of participant selection decisions for each sensing slot. Moreover, the mechanism is proved to be computationally efficient and good economic properties. Second, since workers’ information is usually unknown a priori in practical MCS applications, many efforts have also been devoted to designing unknown worker recruitment issues. For example, Wang et al. [14] consider recruiting unknown workers with not only a strong objective ability but also a good collaboration likelihood. They design a graph theory-based algorithm to find the optimal group of unknown workers. Moreover, a multi-round worker recruitment strategy based on the combinatorial multi-armed bandit model is proposed to balance exploration and

64

M. Xiao et al.

exploitation. Pu et al. [15] propose an optimal unknown worker recruitment scheme for self-organized MCS, called Crowdlet, to maximize the expected sum of service quality through the dynamic programming principle. Wang et al. [2] propose an MCS unknown worker recruitment scheme by leveraging the influence propagation on social networks to maximize the coverage. Abououf et al. [16] propose a novel recruitment system in MCS based on behavioral profiling. Subsequently, a groupbased genetic algorithm is proposed to maximize the task completion quality. Third, some complex unknown worker recruitment schemes are also developed for various MCS systems with different constraints. For instance, Wang et al. [17] propose a new worker recruitment game in socially aware MCS and design a random diffusion model where the invitations are propagated to the social neighbors and receivers can decide to accept or not, which is formulated as a combinatorial optimization problem with an incentive budget to be solved. Gao et al. [18] propose a novel extended UCB-based algorithm maximizing the sensing quality to solve the unknown worker recruitment problem under a limited budget. Moreover, an extended algorithm is also proposed to solve the case that workers’ costs are also unknown. Xiao et al. [19] propose an incentive mechanism based on the combinatorial multi-armed bandit and reverse auction to solve the unknown worker recruitment problem in MCS, whose goal is to maximize the total sensing under a limited budget while ensuring the truthfulness and individual rationality of workers. For recruiting the workers with a quality update and alternativeness, an adaptive incentive mechanism is proposed, which can achieve a better result. Chen et al. in [20] model the unknown worker recruitment problem as a Bayesian Markov decision process and use a dynamic program to obtain the optimal allocation policy for a given budget. Furthermore, an approximate policy, called optimistic knowledge gradient, is proposed to tackle the computational challenge. Gao et al. [21] model the qualities of workers through two factors, i.e., bias and variance, and propose a multi-armed bandit algorithm based on UCB together with a weighted data aggregation scheme to estimate a more accurate ground truth of a sensing task. Moreover, a budget allocation is designed to achieve the global optimization. Gao et al. [22] propose to devise a fairness-aware unknown worker recruitment algorithm to guarantee a minimum selection fraction for each registered worker. Wang et al. [23] take coverage and workload balancing into consideration and propose two approximate methods to find the optimal solutions in 1D and 2D scenarios, respectively. Zhang et al. [24] study a reliable task assignment problem for MCS in a large worker market. Worker confidence is used to represent the reliability of completing the task and two optimization problems, i.e., maximum reliability assignment under a budget and minimum cost assignment under a task reliability requirement are addressed. Fourth, since privacy is also an important issue in MCS, many researchers propose different privacy-preserving worker recruitment schemes to protect diverse privacy information. Xiao et al. in [25] propose a basic user recruitment protocol to achieve a satisfactory sensing quality for each task while minimizing the number of used workers, based on which a secure user recruitment protocol is proposed by using secret sharing schemes to ensure security. Moreover, this protocol is

Unknown Worker Recruitment in Mobile Crowdsourcing

65

further extended to the scenarios with homogeneous and heterogeneous recruitment costs in [26]. Li et al. [27] propose a privacy-preserving participant recruitment scheme for mobile crowdsourcing to maximize the spatial coverage of sensing range while preserving the participants’ location privacy under a budget constraint. Furthermore, a fault-aware crowdsourcing framework is developed to improve the robustness of the recruitment strategy. Xiao et al. [28] develop a .(t, p)collusion resistant scheme to tackle the scenarios, where workers can strategically form collusion coalitions and manipulate their bids together, achieving desirable properties, e.g., p-truthfulness and p-individual rationality. Han et al. [29] allow participants to choose their desired privacy level and formulate the corresponding privacy-preserving recruitment process as an optimization problem to be solved by using encryption techniques. Wang et al. [30] propose a location authenticationbased secure participant recruitment (LA-SPR) scheme for vehicular MCS, which preserves the location and sensitive information from being revealed to the server or requesters. However, all of these works assume that workers’ information is known by the platform a priori. Additionally, there are also other worker recruitment works in MCS that address the dual challenges of privacy preservation and unknown worker scenarios. For example, Wang et al. [31] propose a personalized privacy-preserving task allocation framework for mobile crowdsensing to allocate tasks while preserving location privacy. Specifically, a probabilistic winner selection mechanism and a Vickery payment determination mechanism are proposed to minimize the total distance with the obfuscated information from workers and determine the appropriate payment for each winner. Li et al. [32] design a scalable grouping-based MCS participants selection scheme by using Lagrange polynomial interpolation and study how to protect the bid privacy from being revealed. Yan et al. [33] investigate both of the task bidding and assignment while preserving location privacy, with no assumption that the platform is trustworthy. A probability cost-efficient worker selection mechanism is proposed to determine winners and a probability individualrationality critical payment mechanism is proposed to determine payments for winners, which satisfies the truthfulness and probability-individual rationality.

3 System Model and Workflow In this section, we first introduce the MCS system model and then present the workflow of the model.

3.1 System Model We consider a typical MCS system, which consists of a platform, a requester (which can be extended to multiple requesters easily), and a crowd of unknown workers. A requester can outsource some location-relative crowdsourcing tasks to mobile

66

M. Xiao et al.

workers via the platform periodically before a deadline. The tasks, the platform, and workers are defined as follows: Definition 1 (Tasks and Round) A requester wants to collect some locationsensitive data (e.g., air quality, traffic flow, urban noise) for a period of time. The whole data collection is divided into many tasks, denoted by .M = {1, ..., j, ..., M}. Each task .j ∈ M is specified by a set of attributes .[wj , τj , locj ]. Since the requester may pay different attention to various tasks, we use .wj to denote the weight of the j -th task. .τj and .locj represent the deadline and the location of task j , respectively. Moreover, the MCS system evolves in a time-slotted fashion and the whole process is divided into T rounds, denoted by .t ∈ {1, 2, · · · , T }. Definition 2 (The Platform) As an intermediate agent, the platform builds a bridge between the requester and workers. Specifically, the platform accepts the crowdsourcing tasks .M from the requester, recruits some suitable workers who are willing to perform these tasks, aggregates the data collected by workers, and sends the statistics to the requester. Definition 3 (Unknown Workers) In the MCS systems, there are N workers with unknown qualities, denoted by .N = {1, ..., i, ..., N}. If a recruited worker .i ∈ N is assigned the task j , he/she will travel to the requested location .locj and will upload the collected data to the platform before the deadline .τj . In order to accomplish task j , the worker i will incur a cost .ci,j . We set a threshold .cth to judge the capability of workers. If .ci,j > cth , worker i is not suitable to accomplish task j ; otherwise, the worker is capable of performing task j . t Let a normalized non-negative random variable .qi,j ∈ (0, 1] represent the data quality of worker i completing task j in the t-th round. Actually, the values t of qualities .{qi,j | i ∈ N, j ∈ M} follow an independent and identically t ). Since the unknown distribution with an unknown expectation .qi , i.e., .qi  E(qi,j distribution and the expected quality are unknown, we call these workers unknown workers or workers for short. Note that we assume that the expected quality .qi is fixed since it is mainly determined by the configurations of workers’ smart devices (e.g., storage capacity, camera pixel, CPU frequency, etc.) and their behaviors (e.g., t work habit, skill level, work experience, etc.). Moreover, the actual quality .qi,j t might be influenced might deviate from the expected quality. This is because .qi,j by some exogenous factors (e.g., weather, daily routine, personal willingness, etc.). Additionally, we summarize the commonly used notations throughout the chapter in Table 1.

3.2 System Workflow As illustrated in Fig. 1, the whole workflow of the MCS system includes the following main procedures:

Unknown Worker Recruitment in Mobile Crowdsourcing

67

Table 1 Description of major notations Variable

.qi,j

Description The indexes of worker, task, and round, respectively The number of workers, the number of tasks, the number of rounds The number and set of recruited workers in period t The set of workers who have capability to perform task j The set of tasks that worker i can perform The cost of worker i on task j and the threshold of workers’ capability The privacy budget and the monetary budget B The weight of task j and the expected quality of worker i The quality of worker i in terms of task j in the t-th round

.q¯i,j

t

The learned quality of worker i in terms of task j in the t-th round

.qˆi,j

The CUCB-based quality of worker i in terms of task j in round t.

.i, j, t .N, M, T .K, W

t

.Nj .Mi .ci,j , cth ., B .wj , qi

t

t

Fig. 1 Illustration of the main procedures in MCS

1. The requester publishes multiple location-related tasks to the platform, which will be conducted round by round until T . Meanwhile, the requester can ask for some demands (e.g., the deadline of tasks, the number of each task’s selected workers, the amount of collected data, etc.). 2. The platform broadcasts these tasks and the corresponding demands to mobile workers. 3. Each worker needs to evaluate the cost of completing each task and then submits the evaluation results to the platform. Here, we assume that all workers will truthfully report their costs. This is reasonable because many state-of-the-art auction mechanisms [28, 34, 35] can be introduced to guarantee the truthfulness. 4. According to the objective function, the platform determines a recruitment strategy, whereby the platform recruits suitable workers with high qualities as much as possible and then informs these selected workers.

68

M. Xiao et al.

5. Each recruited worker will move to some specified locations and carry out the corresponding tasks to acquire enough data. Note that each selected worker will perform tasks of which he/she is capable. 6. After receiving the collected data from workers, the platform aggregates these data. If the requester prefers data statistics rather than the original chaotic data, the platform can also provide a data analysis service. Meanwhile, the quality values of recruited workers can be measured. For example, the platform may take the photographing angle, the photo clarity, and some other factors into consideration when determining the qualities of workers in terms of the photos. A profile is used to record the learned qualities of each worker, which will be updated in each round. Finally, the platform will return the aggregated data to the requester.

4 Unknown Worker Recruitment Scheme In this section, we propose an Unknown Worker Recruitment (UWR) scheme. Specifically, we first model the unknown worker recruitment problem as a Combinatorial Multi-Armed Bandit (CMAB) game. Then, we extend the classical Upper Confidence Bound (UCB) mechanism to design a selection criterion, followed by the detailed algorithm and the theoretical analysis. Based on the UWR scheme, we further consider the circumstance that the platform has a limited monetary budget to recruit workers and design an extended UWR scheme to support the budget constraint.

4.1 Modeling and Formulation Selecting workers with the largest qualities under the circumstance that their qualities are unknown as priority is a critical issue in MCS systems. Essentially, it is an online learning and decision-making process, and thus we employ the CMAB model to deal with this problem. Basically, CMAB is made up of a slot machine with multiple arms [36–38]. Pulling an arm can gain a reward which follows an unknown distribution. Based on a bandit policy, a player will pull a set of arms (called a super arm) together round by round to maximize the cumulative revenue. As illustrated in Fig. 2, we model the unknown worker recruitment problem as a CMAB game. More specifically, the platform is regarded as the player, each worker plays the role of an arm, its data quality is treated as the corresponding reward, and recruiting K workers is equivalent to pulling a super arm. In each round, the platform needs to learn the qualities of K recruited workers. Here, .K ∈ {1, 2, ..., N } is a preset value determined by the platform according to task demands. Specially, .K = 1 is a MAB problem.

Unknown Worker Recruitment in Mobile Crowdsourcing

69

Fig. 2 The CMAB model and the recruitment model

Based on the above modeling, the unknown worker recruitment problem is transformed into a bandit policy determination problem, which aims to maximize the total task completion quality. Then, we define the bandit policy: Definition 4 (Bandit Policy) A bandit policy . is a sequence of arm-pulling decisions, denoted by .( 1 , · · · ,  t , · · · ,  T ), which can be represented as an indicator vector. Here, . t = (ψ1t , ψ2t , · · · , ψNt ) ∈ {0, 1}N and T is the number of total round. Moreover, .ψit = 1 means that the worker i will be selected, while t .ψ = 0 indicates that it will not be recruited in the t-th round. i Let .W t be the set of recruited workers in the t-th round, and we use .Nj to denote the set of workers who have capability to perform task j (i.e., .ci,j ≤ cth , ∀i ∈ Nj ). Moreover, we define the final completion quality of a task according to .W t in round t, denoted by .Ujt (W t ),  Ujt (W t ) =

.

W t ∩ Ntj = ∅

0,

t | i ∈ W t ∩ Nt }, W t ∩ Nt = ∅ max{qi,j j j

(1)

Our objective is to determine .{W 1 , · · · , W t , · · · }, such that the total expected weighted completion quality of all tasks is maximized under a bandit policy .. Thus, the recruitment problem can be formulated as follows: P1 :

.

Maximize : E[

M

T t=1

Subj ect to : | W t |=

j =1

N

i=1

wj · Ujt (W t )].

ψit = K, ∀t ∈ [1, T ].

ψit ∈ {0, 1}, ∀i ∈ N, ∀t ∈ [1, T ] Here, Eqs. (3) and (4) indicate the quantity constraint in each round.

(2) (3) (4)

70

M. Xiao et al.

4.2 Algorithm Design We design an extended UCB-based bandit policy to select unknown workers for our K-armed CMAB problem, which can balance the trade-off between exploration (i.e., trying some sub-optimal workers to find the potential optimal workers) and exploitation (i.e., recruiting the current best workers based on the learned results). The detailed design is expounded as follows. When a worker i is recruited in the t-th round (i.e., .ψit = 1), worker i will perform the tasks .Mi = {j | ci,j ≤ cth } and upload collected data. Thus, the number of times of worker i’s quality being learned by the platform in the t-th round is actually .| Mi |. Based on this, we first introduce .nti for .i ∈ N, t ∈ [1, T ] to record the number of times that i’s quality has been learned. Then, we introduce the notation .q¯it to record the average learned quality value for the worker i until the t-th round. After the recruited workers upload data, the values of .nti and .q¯it at the end of t are updated:  nti =

nt−1 i +| Mi |, .

nt−1 i ,

if

ψit =1

if ψit =0

⎧ t−1 t−1  t q¯ ni + j ∈Mi qi,j ⎪ ⎪ ⎨ i , if ψit =1 t−1 t ni + | Mi | q¯i = ⎪ ⎪ ⎩ t−1 q¯i , if ψit =0

(5)

Now, we design a Combinatorial UCB (CUCB)-based quality value .qˆit to recruit workers. .qˆit is comprised of two parts: the empirical quality (indicating the learned knowledge from observed data qualities) and a confidence bound (indicating the uncertainty of empiricism), i.e., .

qˆit = q¯it + γit ,

γit =



(K + 1) ln(

N

t t j =1 nj )/ni .

(6)

Here, .γit is the upper confidence bound, which takes a simple heuristic principle called optimism in the face of uncertainty. Meanwhile, when the number of times t t .n becomes large, we can find that .γ decreases rapidly. This implies that workers i i who are selected less may get more chances to be recruited in the next round. In this chapter, the values of .nti , .q¯it , and .qˆit make up the worker profiles recorded by the platform. Next, considering the maximum weighted completion quality problem in our MCS system, we introduce the CUCB-based quality function. When a task is conducted by multiple workers, we use the maximum quality value of these workers to denote the final result of this task in a round. According to the CUCB-based quality values .{qˆit−1 | i ∈ N} revealed in the first .t − 1 rounds, we compute the CUCB-based quality function for the solution .W t , i.e., .F t (W t , qˆit−1 ) = M t−1 · I{ψit = 1, i ∈ Nj }}. Here, .I{·} is an indicator function, i.e., j =1 wj · max{qˆi .I{true} = 1 and .I{f alse} = 0. Based on the above CUCB-based quality function, we introduce a greedy armpulling strategy to recruit unknown workers in each round. In the initialization period, the platform will recruit all workers to explore their initial qualities, to initialize the values of .nti , .q¯it , and .qˆit . In each subsequent round, the set .W t is firstly

Unknown Worker Recruitment in Mobile Crowdsourcing

71

set to be empty, and then we add workers into .W t until .| W t |= K. The selection criterion is the ratio of the marginal value of the function .F t (W t , qˆit−1 ) and costs, i.e., .

i = argmaxi  ∈N\W t

F t (W t ∪i  ,qˆit−1 )−F t (W t ,qˆit−1 )  j ∈M ci  ,j

(7)

i

After the recruitment process in each round (i.e., .|W t | = K), each selected worker i needs to accomplish the task set .Mi and upload the corresponding collected data. Based on these data, the platform can obtain the specific quality t | j ∈ M }) and then updates the worker profiles according to information (i.e., .{qi,j i Eqs. (5) and (6). According to the above solution, we illustrate the unknown worker recruitment scheme pseudocode in Algorithm 1. In the initialization phase (i.e., .t = 1), the platform will select all workers to get the initial workers’ qualities (Steps 1–4), which will be used for initializing several parameters (Step 5). In steps 6–16, the platform conducts the recruitment process round by round until .t = T . Specifically, the platform selects the worker according to the CUCB-based quality values and the designed selection criteria (i.e., Eq. (7))) and adds the worker into the set .W t (Steps 9–10). That is, the largest ratio of the marginal CUCB-based quality function value and costs is selected from the set .N\W t . After that, the recruited workers will accomplish their corresponding tasks and send the collected data to the platform in step 13. In Steps 14–15, the platform evaluates the qualities of selected workers and updates the worker profiles (i.e., .nti , .q¯it , and .qˆit ). Finally, the scheme returns the set of recruited workers in each round. Moreover, the computation complexity of the algorithm is dominated by in the exploitation and exploration phase, which is denoted by .O(T KNM).

4.3 Theoretical Analysis When knowing the quality distribution of each worker, i.e., .qi for .∀i ∈ N, the platform can output an approximately optimal solution (denoted by .W ∗ ) by recruiting the workers with the high ratios of marginal weighted quality value and costs in each round. We let .α be the ratio of the approximately optimal solution and the optimal one, which meets .α ≥ 1/2 according to the existing studies [39]. Based on this, it is no longer fair to directly compare our unknown worker recruitment results with the optimal solution. Therefore, we introduce the concept of .α-approximation regret [6, 40], which is the difference for the total completion qualities achieved by the approximately optimal solution and our scheme. That is, .

 W t , qˆit−1

   T t−1 t ≤ Tt=1 F (W ∗ , qi ) − E t=1 F W , qˆi

Rα = α ·

T

t=1 F (W

∗, q ) − E i



T t=1 F

(8)

72

M. Xiao et al.

Algorithm 1 Unknown Worker Recruitment (UWR) scheme Require: N, M, {wj | j ∈ M}, K, {ci,j | i ∈ N, j ∈ M} Ensure: , W t ⊆ W for ∀t ∈ [1, T ] 1: Initialization phase: 2: Recruit all workers when t = 1, i.e., 3: for each worker i ∈ N do ψi1 = 1; perform tasks; upload collected data; 1 (j ∈ M ); 4: The platform evaluates the quality of each worker qi,j i 5: for each worker i ∈ N do Update n1i , q¯i1 , γi1 , qˆi1 ; 6: Exploration and exploitation: 7: while each round t < T do 8: t ⇐ t + 1; W t = ∅. 9: while | W t |< K do F t (W t ∪i  ,qˆit−1 )−F t (W t ,qˆit−1 )  j ∈Mi ci  ,j W t = W t ∪ i.

Determine i = argmaxi  ∈N\W t

10:

11: Add worker i into W t , i.e., 12: end while 13: for each worker i ∈ W t do ψit = 1; perform tasks; upload collected data; t , i ∈ Wt, j ∈ M ; 14: Obtain the qualities qi,j i 15: for each worker i ∈ N do Update nti , q¯it , γit , qˆit ; 16: end while 17: Return W t , , ∀t ∈ [1, T ].

 t ∗ Here, .F t (W ∗ , qi ) = M j =1 wj · max{qi · I{ψi = 1, i ∈ Nj ∩ W }}. Then, we define the largest/smallest possible difference of the selection criteria among non ∗ .α-optimal workers (i.e., .W = W ): .

max =

F (W ∗ ,qi )  i∈W ∗ Ci

F (W  ,qi )  ; min i∈W  Ci W =  W

− min  ∗

=

F (W ∗ ,qi )  i∈W ∗ Ci

F (W  ,qi )  ; W  =W ∗ i∈W  Ci

− max

 where .Ci = j ∈Mi ci,j . Let .Cmin = min {Ci | i ∈ N}, Cmax = max {Ci | i ∈ N}. Besides, we define the largest possible difference of the quality values when .W  = W ∗ , i.e., .∇max = F (W ∗ , qi )−minW  =W ∗ F (W  , qi ). Next, we introduce the counter t .χ of worker i after the initialization phase. In each round, if the .α-optimal workers i are recruited (i.e., .W t = W ∗ ), .χit will remain unchange; if a non-.α-optimal set of workers is selected (i.e., .W t = W ∗ ), we can find .i = argmini  ∈W t χit−1 and let  t−1 t t .χ = χ + 1. Since there must be one element in .χ increased by 1 when the i i i non-.α-optimal workers are recruited, the sum of the counter .χit is equal to the total number of non-.α-optimal sets of workers. Before analyzing the .α-approximation regret of the proposed scheme, we need to analyze the bound of the expected counter, shown in the following lemma. Lemma 1 (Chernoff–Hoeffding Bound) [36] Suppose that .X1 , X2 , · · · , Xn are n random variables with common range .[0, 1], satisfying .E [Xt | X1 , · · · , Xt−1 ] = μ for .∀t ∈ [1, n]. Let .Sn = X1 + · · · + Xn . Then, .∀a ≥ 0, −2a P [S . n ≥ nμ + a] ≤ e

2 /n

, P [Sn ≤ nμ − a] ≤ e−2a

2 /n

.

(9)

Unknown Worker Recruitment in Mobile Crowdsourcing

73

Lemma 2 The expected counter .χiT has an upper bound for any worker .i ∈ N in the T -th round, that is, .

  E χiT ≤

4K 2 (K+1) ln(N MT ) (min )2 (Cmin )2

+1+

Kπ 2 3 .

(10)

Proof In the t-th round, we use .Iit to denote the indicator that .χit is incremented. According to the update rule of the counter, we can get the following results: T     I Iit = 1 = ζ + I Iit = 1, χit ≥ ζ t=2   T  F (W ∗ , qˆit ) t F (W t+1 , qˆit )  ≤ζ+ ≥ , χ ≥ ζ I i t=1 i∈W ∗ i∈W t+1 Ci Ci

=

T

.χi

T

t=2

=ζ+

T t=1

T  ≤ ζ+ I t=1

≤ζ+

I

 i∈W

ϑ t+1 t+1 i



qˆ t qˆit  ≥ ϑ ∗ i , χit ≥ ζ ∗ i i∈W Ci Ci



K K   ϑit+1 t ϑi∗ t max qˆs ∗ (i) qˆs(i) ≥ min t t t t C C 1≤ns ∗ (1) ≤···≤ns ∗ (K) ≤t ζ ≤ns(1) ≤···≤ns(K) ≤t i∗ i i=1

i=1

t T  

t 

···

t=1 nt =ζ s(1)

t 



t 

nts(K) =ζ nts ∗ (1) =1 nts ∗ (K) =1

 K K   ϑ t+1 ϑi∗ t i t qˆs ∗ (i) qˆs(i) ≥ I Ci ∗ Ci i=1

(11)

i=1

where .s(i) and .s ∗ (i) denote the i -th element in .W t+1 and .W ∗ , respectively. .ϑit+1 =   t+1 } · wj . Next, we prove the probability of the event j ∈Mi I i = argmaxi  ∈W t+1 {qˆi

K

.

ϑit+1 t i=1 Ci (q¯s(i)

K

ϑi∗ i=1 Ci ∗

t ) ≥ + γs(i)

(q¯st ∗ (i) + γst∗ (i) ). According to the proof by con-

tradiction, we have that at least one of the following equations must hold: (i) K ϑit+1 K ϑi∗ K ϑit+1 t ϑi∗ t t t i=1 Ci (qs(i) + γs(i) ), i=1 Ci ∗ q¯s ∗ (i) ≤ i=1 Ci ∗ (qs ∗ (i) − γs ∗ (i) ), (ii) . i=1 Ci q¯s(i) ≥

K

.





ϑ∗

ϑ t+1

K t i i and (iii) . K i=1 Ci ∗ qs ∗ (i) < i=1 Ci (qs(i) + 2γs(i) ). Based on Lemma 1, we can prove the upper bounds for (i) and (ii) and get

.

 K K   ϑ∗ ϑi∗ i t t (qs ∗ (i)−γs ∗ (i) ) Pr q¯ ∗ ≤ Ci ∗ Ci ∗ s (i) i=1

i=1

≤

K  i=1

 Pr

ϑi∗ Ci ∗

q¯st ∗ (i)≤

 ϑi∗ (qs ∗ (i)−γst∗ (i) ) Ci ∗

  ≤ K · exp −2nts ∗ (i) (K + 1) ln  

i ∈N

nts ∗ (i  )

≤ K · exp(−2(K + 1) ln(N t)) ≤ K · t −2(K+1). K  K  ϑ t+1  ϑit+1 i t t Pr (qs(i)+γs(i) ) q¯ ≥ Ci Ci s(i) i=1

i=1



/nts ∗ (i)



(12)

74

M. Xiao et al.

≤

K 

 Pr

i=1

 ϑit+1 ϑit+1 t t q¯ ≥ (qs(i)+γs(i) ) Ci s(i) Ci

   ≤ K · exp −2nts(i) (K + 1) ln 

i ∈N

nts(i  )



/nts(i) ≤ K · t −2(K+1)

(13)

If conditions (i) and (ii) are false, we can infer that (iii) is true due to the contradiction. Therefore, we need to pick a suitable .ζ to make (iii) impossible. K .

K ϑ t+1 K ϑ t+1 ϑi∗ i i t qs(i) − 2 γs(i) qs ∗ (i) − i=1 Ci i=1 Ci ∗ i=1 Ci     t  j ∈M wj  (K + 1) ln( i  ∈N ns(i  ) ) ≥ min − 2K Cmin nts(i)  K ≥ min − 2 Cmin

(K + 1) ln(N MT ) 4K 2 (K + 1) ln(N MT ) , ≥0⇒ζ ≥ ζ (min )2 (Cmin )2 2

ln(N MT ) Now, we can continue Eq. (11)) and get .E[χiT ] ≤  4K((K+1) + 2 2 min ) (Cmin )

4K 2 (K+1) ln(N MT ) (min )2 (Cmin )2

+1+

Kπ 2 3

T

t=1 2Kt

(14) −2

≤  

.

Based on Lemma 2, we further derive the upper bound on the .α-approximation regret as follows. Theorem 1 At the end of the round T , the upper bound on the .α-approximation regret .Rα is .O(N K 3 ln(N MT )). Proof According to Lemma 2 and the definition of the .α-approximation regret in Eq. (8)), we can get Rα . ≤

T t=1

 ≤N

∗

F (W , qi ) − E



T t=1

  N  t−1 t T ≤ F W , qˆi χi ∇max

Kπ 4K 2 (K + 1) ln(N MT ) +1+ 3 (min )2 (Cmin )2

i=1

2

= O(N K 3 ln(N MT ))

The proof is now completed.

(15)  

4.4 Extansion: Budget-Limited UWR Scheme In real scenarios, the platform needs to compensate for the costs of workers and might have a limited monetary budget B. Under such a circumstance, the platform cannot determine the number of rounds T . That is, the whole recruitment process will terminate until the budget is exhausted. When we consider the budget constraint, our unknown worker recruitment problem will be reformulated as

Unknown Worker Recruitment in Mobile Crowdsourcing

75

follows: P2 :

.

Maximize : E



wj · Ujt (W t ) j =1

t≥1

Subj ect to : | W t |=



M N i=1

(16)

.

ψit = K, ∀t > 1.

(17)

ψit ∈ {0, 1}, ∀i ∈ N, ∀t ≥ 1.    ci,j ≤ B t t≥1

(18) (19)

j ∈Mi

i∈W

The main differences between the problem P1 and the problem P2 are (i) the total cost of recruited workers cannot exceed a fixed monetary budget B and (ii) since the recruitment cost in each round is uncertain, the stopping round (i.e., the total round T ) is indeterminate. We use .T(B) to represent the stopping round under the budget constraint B. Different from the scheme designed for solving P1, we change the loop condition (i.e., Step 7 in Algorithm 1) into the budget constraint. Due to the limited space, we omit the detailed algorithm for P2. Considering the budget constraint, we also need to re-analyze the .αapproximation regret and obtain the following theorem: Theorem 2 For any budget .B > 0, the expected .α-approximation regret is tightened to .O(NK 3 ln(B + NK 2 ln(MN K 2 ))). T(B)

Proof According to Lemma 2, we can have .E[χi T(B) .E[χ ] i

Kπ 2 3 ,

] ≤

4K 2 (K+1) ln(N MT(B)) (min )2 (Cmin )2

+

denoted by ≤ θ1 ln(T(B)) + θ2 for brevity. Then, we derive 1+  the stopping round of .α-optimal solution: .T∗ (B) =  CB∗ , where .C ∗ = i∈W ∗ Ci . Meanwhile, we get .B/C ∗ − 1 ≤ T∗ (B) ≤ B/C ∗ . Next, we derive the upper bound on .T(B) and have .

T(B) ≤ T∗ (B) + T



i ∈W / ∗

T(B)

ni

Cmax ≤ T∗ (B) +

T(B)

N Cmax E[χi KCmin

]

(20)

Based on the inequality .ln(x) < x − 1 for .∀x > 0, we obtain 

.

KCmin T(B) ln 2NCmax θ1



⇒ ln(T(B)) ≤ T(B)

≤

KCmin T(B) −1 2NCmax θ1

  2NCmax θ1 KCmin T(B) − 1 + ln KCmin 2NCmax θ1 T(B)

Owing to .E[χi ] ≤ θ1 ln(T(B)) + θ2 , we substitute .E[χi into Eq. (20)), and we further get .

2B 2NCmax T(B) ≤ ∗ + KCmin C

(21)

] and Eq. (21))

   2B 2NCmax θ1 θ2 − θ1 + θ1 ln = ∗ + θ3 (22) KCmin C

76



M. Xiao et al.

Moreover, we derive the lower bound on .T(B). According to .T(B) ≥ T∗ (B − T(B)

ni can attain i ∈W / ∗

Cmax ) ≥

T(B) . ≥

T(B)

B−N Cmax E[χi C∗

B−N Cmax [θ1 ln(T(B))+θ2 ] C∗

]

− 1 and .ln(T(B)) ≤ ln(2B/C ∗ + θ3 ), we

−1≥

B−N Cmax [θ1 ln(2B/C ∗ +θ3 )+θ2 ] C∗

−1

(23)

Now, we have derived the upper bound and the lower bound of .T(B). Based on this, we begin the analysis of the .α-approximation regret. .Rα (B)

≤

T∗ (B) t=1

F (W ∗ , qi ) − E



T(B) t=1

  F W t , qˆit−1

   T(B) BF ∗ t−1 ∗ ∗ t − T(B)F + T(B)F − E F W , q ˆ i t=1 C∗  T(B) ≤ F ∗ (B/C ∗ − T(B)) + χ ∇max i∈N i    B − N Cmax [θ1 ln(2B/C ∗ + θ3 ) + θ2 ] − 1 ≤ F ∗ B/C ∗ − C∗ ≤

+ N ∇max (θ1 ln(2B/C ∗ + θ3 ) + θ2 ) = O(N K 3 ln(B + N K 2 ln(MN K 2 )))

where .F ∗ = F (W ∗ , qi ). We complete the proof of the theorem.

(24)  

5 Privacy-Preserving Unknown Worker Recruitment Scheme As mentioned in Sect. 3, many algorithms have been proposed to solve unknown worker recruitment problems, whereby the platform can learn the knowledge on workers’ qualities from the collected data [28, 34–36, 38]. However, the workers in real applications are generally unwilling to disclose their qualities since they might contain some sensitive private information [41]. Therefore, it is crucial to preserve each worker’s quality from being revealed to other workers during the unknown worker recruitment process. In this section, we focus on the privacy-preserving unknown worker recruitment issue. First, we model the privacy-preserving unknown worker recruitment problem as a Differentially Private Multi-Armed Bandit (DP-MAB) game. Then, by extending the .δ-First and Upper Confidence Bound (UCB) mechanisms, we propose a budget feasible Differentially Private .δ-First (DPF) bandit algorithm and a budgetfeasible Differentially Private UCB-based (DPU) bandit algorithm.

Unknown Worker Recruitment in Mobile Crowdsourcing

77

5.1 DP-MAB Model By taking privacy and budget into consideration, we model the unknown worker recruitment as a DP-MAB game, where the role map between problem and model is introduced in Sect. 3. In addition, the rewards are sensitive information that needs to be protected by a differentially private mechanism. The objective of the platform is to maximize the accumulative reward by pulling the arms sequentially according to the budge-feasible policy, while preserving the privacy of the rewards. The budgetfeasible policy is defined as follows: Definition 5 (Budget-Feasible Bandit Policy) A bandit policy . is a sequence of maps: .{1 , 2 , ..., i , ...}, each of which specifies the arm that the platform 1:t−1 1:t−1 will pull under the historical records, i.e., .at = t (qi,j ), where .qi,j = t−1 ). Moreover, the total cost is no larger than the given budget, i.e., (q 1 , q 2 , ..., qi,j i,j t i,j . t ci ≤ Bi , where .Bi is the budget of task i.

We adopt the differential privacy mechanism to preserve the rewards of each arm from being disclosed, defined as follows: Definition 6 (.-Differential Privacy [42]) A bandit policy . is .-differentially 1:t−1 1:t−1 and .qi,j that private if and only if over all time slots, for all sequences .qi,j differ in at most one time slot, and for any subset .S ∈ N, we have       1:t−1 1:t−1 ∈ S ≤ e × P t qi,j ∈S P t qi,j

.

(25)

Here, . is a very small constant that the policy provides, indicating the privacy level.

5.2 Problem Formulation Under the DP-MAB model, a bandit policy helps the platform recruit workers. This policy needs to (i) satisfy .-differential privacy in the whole process, (ii) maximize the expected accumulative reward for each task, and (iii) conduct the policy under t and .z the budget .Bi , ∀i ∈ M. Let .zi,j i,j,t + denote the number of times that worker j has been pulled from time slot 1 to time slot t for task i and from time slot t to the end, respectively. We use .ri and .E[ri ] to denote the accumulative reward of task i and its expectation value. .E[r] can be calculated as follows: E[ri ] =

.

 j ∈N

qi,j E[zi,j,1+ ]

(26)

78

M. Xiao et al.

Therefore, the unknown worker recruitment problem can be formulated as follows: P3 :

.

Maximize : E[r].  Subj ect to :

(27)

j ∈N

ci,j zi,j,1+ ≤ Bi.

(28)

Eq. (25) holds

(29)

5.3 The DPF Algorithm Under DP-MAB model, the DPF algorithm adopts a budget-feasible .δ-First bandit policy to determine workers, where the .δ ∈ [0, 1] ratio of .Bi is invented for learning the workers’ quality (i.e., exploration phases) and .(1 − δ) ratio of .Bi is used to select the best worker (i.e., exploitation phases). The DPF algorithm utilizes a hybrid differentially private mechanism to protect the privacy of workers’ quality information and prevent its disclosure. In each time slot, the platform will determine an arm and pull it, whose quality will be added to the accumulative reward. When the accumulative reward of each worker updates, this mechanism will generate a Laplace noise for each incremental value. Let the rewards contributed by the worker .j ∈ N for task i be denoted as .qi,j,1+ = {qi,j,1 , qi,j,2 , ..., qi,j,t , ...}. Based on the hybrid differentially private mechanism, we introduce a function .H(·) to add Laplace noise, which maps a series of rewards to a disturbed accumulative reward. Laplace distribution with scale .λ and mean zero can be denoted by .Lap(λ), whose 1 − |x| probability density function is denoted by .f (x)Lap(λ) = 2λ e λ . Therefore, for the input .qi,j,1:t = {qi,j,1 , qi,j,2 , ..., qi,j,t }, .Hj (·) can be calculated by Hj (qi,j,1:t ) =

.

t l=1

 qi,j,l + Lap

2N 



 + (k + 1)Lap

2N logt 

 (30)

where k is the number of 1’s in the binary expression of t and the k Laplace noises are added at time slot t. The perturbed accumulative reward of worker j can be denoted by .rˆi,j,t = Hj (qi,j,1:t ). Therefore, for each arm of task i, the platform can calculate perturbed accumulative reward. In this way, the true value of .qi,j,t can be protected from being revealed. In the DP-MAB model, the whole process is divided into the exploration and the exploitation phases, in which .δB is for exploration and .(1 − δ)B is for exploitation. In the exploration phase, since there is no prior knowledge in workers’ qualities, the platform estimates the mean reward by recruiting the workers indiscriminately and equally one by one. W.L.O.G., we assume that the costs of all workers (arms) in .N for task i satisfy .ci,1 ≤ ci,2 ≤ ... ≤ ci,N . Therefore, within the budget .δB, the end  time slot .τ of the exploration phase can be calculated by .τ = argmax tl=1 ci,al . t

Then, we can compute the total number of times that worker j has been pulled which satisfies .zi,j,τ ≥  NB  and the perturbed accumulative reward is .rˆi,j,τ = H(qi,j,1:τ ).

j =1 ci,j

Unknown Worker Recruitment in Mobile Crowdsourcing

79

In the exploitation phase, the platform selects the workers of each task according rˆ to their mean quality . zi,j,τ . In order to maximize the expected accumulative reward i,j,τ within budget .(1 − δ)B, we model the arm-pulling problem in this phase as the Knapsack problem, where each arm and its estimated mean reward are deemed as an item and its value, the cost of pulling the arm corresponds to the weight of the item, and the budget .(1 − δ)B is regarded as the capacity of the knapsack. Thus, the problem can be formulated as follows: maximize :

N

.

subj ect to :

i=1

N l=1

rˆi,j,τ zi,j,(τ +1)+. zi,j,τ

(31)

ci,l zi,j,(τ +1)+

(32)

Since the knapsack problem is a well-known NP-hard problem, we solve it by a greedy strategy. First, the platform computes the value per weight for each item, rˆ i.e., . zi,ji,jci,j , called density. Then, the platform sorts these arms in the ascending order of their densities. Next, the platform pulls arms with the highest density within the budget .(1 − δ)B. The detailed DPF algorithm is shown in Algorithm 2.

5.4 Performance Analysis of DPF Algorithm In this subsection, we prove the security and analyze the regret performance of DPF. Theorem 3 The DPF algorithm is .-differentially private. Proof Consider an arbitrary time slot t and a pair of reward sequences .qi,j,1:t  and .qi,j,1:t with at most one different reward vector, i.e., .∃j ≤ t, .qi,j,1:t =    = {qi,1,j , ..., qi,N,j }. Therefore, for any {qi , 1, j, ..., qi,N,j } is tampered to .qi,j,1:t  task i and worker j , .qi,j,1:t and .qi,N,j differ in at most one reward record. Let . = t  |. We can get . ≤ 1 and .| t q  maxl∈[1,t] |qi,j,l − qi,j,l l=1 i,j,l − l=1 qi,j,l | ≤  ≤ 1. Therefore, we can get .

P{Hj (qi,j,1:t ) = ri,j }  P{Hj (qi,j,1:t ) = ri,j } =

=

2N logt 2N t q P{ri,j − l=1 )} i,j,l = Lap(  ) + (k − 1)Lap(  2N logt 2N t q P{ri,j − l=1 )} i,j,l = Lap(  ) + (k − 1)Lap( 

t t q k−1 f (ri,j − l=1 i,j,l )|Lap 2N · [f (ri,j − l=1 qi,j,l )|Lap( 2Nlogt ) ] 



t q  )| t  k−1 f (ri,j − l=1 i,j,l Lap 2N · [f (ri,j − l=1 qi,j,l )|Lap( 2Nlogt ) ] 

≤e

 k−1 t t  2N (1+ logt )|l=1 qi,j,l −l=1 qi,j,l |



≤e

 N

≤e

 N

80

M. Xiao et al.

Algorithm 2 The DPF algorithm Require: .N, {qi,j,t |i ∈ M, j ∈ N, t ∈ T}, .B, , δ Ensure: r Initialization: t=0;.∀i ∈ M, ∀j ∈ N : zi,j,t = 0; 2: Exploration phase:  .t = t + 1; Let .Bt = δB be the residual budget; Let .t = t; 4: while .Bt ≥ ci,1 do while .j = t  mod N and .Bt ≥ ci,j do 6: .at = j ; Pull the .at th arm;  .∀j ∈ N : zi,j  ,t = zi,j  ,t−1 ; 8: .zi,j  ,t = zi,j  ,t−1 + 1;  .∀j ∈ N : rˆi,j  ,t = Hj  (qi,j  ,1:t );   10: .Bt+1 = Bt − cat ; .t = t + 1; .t = t + 1; end while  12: .t = 1; end while 14: Exploitation phase: Let .Bt = (1 − δ)B be the residual budget; 16: Let .N = N be the workers that have not been recruited; while .Bt ≥ ci,1 do rˆ

while .at = argmaxj ∈N ci,j,t do i,j Pull the .at th arm;  20: .∀j ∈ N : zi,j  ,t = zi,j  ,t−1 ; .zi,j  ,t = zi,j  ,t−1 + 1;  .∀j ∈ N : rˆi,j  ,t = Hj  (qi,j  ,1:t ); 22: .Bt+1 = Bt − cat ; .t = t + 1; end while.N = N − at ; 24: end  while .r = j ∈N rˆi,j,t−1 ; 18:

Since k is the number of 1’s in binary expression of t, .k − 1 ≤ logt. Therefore, as for each task, the hybrid mechanism can guarantee each worker’s reward sequence is . N -differentially private. Now, we can get N P{(qi,j,1:t ) = a} j =1 P{Hi (qi,j,1:t ) = ri,j } ≤ N ≤ e .   P{(qi,j,1:t ) = a} P{H (q ) = r } i i,j j =1 i,j,1:t

(33)

Therefore, the DPF algorithm is .-differentially private.   Next, in order to analyze the upper bound on regret of the DPF algorithm, we introduce a lemma in [42, 43]) and Chernoff–Hoeffding bound in Sect. 3 will be used in the analysis. Lemma 3 Consider the accumulative reward .ri,j,t of worker j and the perturbed √

accumulative reward .rˆi,j,t . Denote .vi,t = −b (b > 0), we have .0 < γ ≤ t

8 4  log( γ )(logt

+ 1). Thus, .∀t ∈ T and

Unknown Worker Recruitment in Mobile Crowdsourcing

81

P{|ˆri,j,t − ri,j,t | ≥ vt } ≤ γ

.

where, .|ˆri,j,t − ri,j,t | is the sum of added Laplace noises to .ri,j,t . And .vt indicates an upper bound on the total Laplace noises with a high probability. Lemma 4 Denote the total accumulative reward at time slot t as .ri,j for task i, ∗ . Let .τ (τ < t) be the end and denote the optimal total accumulative reward as .ri,t time slot of exploration phase. Then .∀η > 0 and .0 < γ ≤ t −b (b > 0), with the probability as least .1 − (e− E

.



∗ ri,t



2η2 t

∗ ] − E[r ] satisfies + γ ), the expected regret .E[ri,t i,t

N c 4(η + vi,t )l=1 i,j − E[ri,t ] ≤ 2 + δBσ + ci,j∗ ri,j  i,j ci,j

where .j∗ = argmaxj ∈N z

ri,j  i,j ci,j

, .i0 = argminj ∈N z



 1 −1 δ

and .σ =

ri,j∗ zi,j∗ ci,j∗

(34) −z

r

i,j0  i,j0 ci,j

. 0

Proof First, according to the DPF algorithm, the total accumulative reward in the exploration phase satisfies ri,j0 .E[ri,T ] ≥ δB zi,j0

! ≥ δB

ri,j0 −1 zi,j0

(35)

Second, according to the greedy arm-pulling strategy in the exploitation phase, the expected accumulative reward .ri,T,t satisfies " E[ri,T,t ] ≥

.

(1 − δ)B ci,jˆ∗

#

ri,jˆ∗ zi,jˆ∗

≥ (1 − δ)B

ri,jˆ∗ zi,jˆ∗ ci,jˆ∗

−1

(36)

rˆ

t q where .jˆ∗ = argmaxj ∈N zˆ i,jc . Next, since .ri,j,t = l=1 i,j,l denotes the sum of i,j i,j qualities that worker j actually contributes to the platform for task i, we can get according to Lemma 1:

$  $ $ rˆi,j,t − ri,j,t $ vi,t $ $ = P{|ri,j,t ˆ − ri,j,t | ≥ vi,t } ≤ γ .P $≥ z $ zi,j,t i,j,t

(37) r

At the same time, the expected value of accumulative reward denoted by .zi,j,t zi,j i,j r

i,j t q i.e., .E[ri,j,t ] = E[l=1 i,j,l ] = zi,j,t zi,j . Then, we can get according to Lemma 3

$    $ 2 $ ri,j,t ri,j $$ ri,j η − z2η i,j,t ≥ = P r ≤ e − z P $$ − i,j,t i,j,t zi,j $ zi,t zi,j zi,j,t

.

(38)

82

M. Xiao et al.

Combining Eqs. (37) and (38), we can get % & $  $ 2 $ rˆi,j ri,j $$ vi,t + η − z2η $ i,j,t ≥1− e .P +γ $z − z $ ≤ z i,j i,j,t i,j Note that, .t ≥ zi,j,t ≥ zi,j,T ≥ 

δB N c  l=i i,l

≥

δB N c . 2l=1 i,l

(39)

We can get

$  $   2 $ rˆi,j ri,j $$ vi,t + η − 2ηt $ ≤ − + γ .P ≥ 1 − e $z N c zi,j $ δB/2l=1 i,j i,l

(40)

Finally, according to Eqs. (35) and (36), we have ri,j∗ − E[ri,T ] − E[ri,T:t ] zi,j∗ ci,j∗ % & ri,jˆ∗ ri,j∗ ≤ δσ B + (1 − δ)B − +2 zi,jˆ∗ ci,j∗ zi,j∗ ci,j∗

∗ E[ri,t ] − E[ri,t ] ≤ B

.

(41)

Furthermore, we can get according to Eq. (40)) $  $   2 $ rˆi,j∗ $ r v + η i,j i,t ∗ $ − 2ηt $ ≤ − +γ } .P ≥1− e $z N c zi,j∗ $ δB/2l=1 i,j∗ i,l

(42)

$ $    $ rˆ ˆ 2 ri,jˆ∗ $$ vi,t + η $ i,j∗ − 2ηt − + γ } .P $ ≥ 1 − e $≤ N c $ zi,jˆ zi,jˆ∗ $ δB/2l=1 i,l ∗

(43)

ri,jˆ ∗ zi,jˆ ci,j∗ ,

r

Since . zi,j i,jc∗i,j ≤ ∗

∗

∗

we can get according to Eqs. (41) and (43)

∗ E[ri,t ] − E[ri,t ] ≤ 2 + δBσ +

.

2(1−δ)B(η+vi,t )

N c ci,j∗ δB/2l=1 i,l   N c 4(η + vi,t )l=1 1 i,j −1 ≤ 2 + δBσ + ci,j∗ δ

with the probability of no less than .1−(e−

2η2 t

(44)

+γ ). Therefore, the lemma holds.

 

Unknown Worker Recruitment in Mobile Crowdsourcing

83

5.5 DPU Algorithm Based on the traditional UCB bandit algorithm, we extend it to be budget-feasible by considering the costs and budget of pulling arms simultaneously. Meanwhile, we apply the same hybrid differentially private mechanism to protect the privacy of workers’ quality. Besides using the estimated average qualities, the upper confidence bounds of the estimated average qualities are used in the DPU algorithm. The traditional UCB policy computes a UCB index for each arm, containing the current average reward and an upper bound of the corresponding confidence. In each time slot, the arm with the maximal UCB index value will be pulled. Here, we define a novel concept, called differentially private UCB index, taking privacy into consideration. Compared with the UCB index, the differentially private UCB index adds a Laplace noise and selects multiple best arms within the budget constraint rather than the maximum, which can also be formalized as a series of knapsack problems. First, we assume that the platform has calculated the current accumulative reward of task i contributed by each worker .j ∈ N and has perturbed the value to get .rˆi,j,t by using a hybrid differentially private mechanism. The differentially private UCB index is defined as follows: Definition 7 (Differentially Private UCB Index) The differentially private UCB index of the j -th arm for task i, denoted by .Ii,j,t , indicates the perturbed expected average reward, i.e., the estimated quality and the size of the corresponding confidence interval, satisfying Ti,j,t

.

rˆi,j,t + = zi,j,t



2lnt vi,t + zi,j,t zi,j,t

(45)

vi,t is an upper bound of confidence for the perturbed accumulative + zi,j,t where . z2lnt i,j,t reward. Like the DPF algorithm, the platform seeks for the optimal policy within the budget, which is also modeled as a knapsack problem. Different from DPF algorithm, the differentially private UCB index is seen as the value of the item. The problem is formulated as follows: maximize :

.

subj ect to :

N l=1

N

l=1

zi,j,t + Ii,j,t−1.

(46)

ci,j zi,j,t+ ≤ Bi,t

(47)

where .Bi,t is the remaining budget at time slot t for task i. Finally, the platform solves the above problem to produce a solution .{zi,j,t+,...,zi,N,t+ } by using a greedy strategy and selects an arm to be pulled with the following probability:

84

M. Xiao et al.

Algorithm 3 The DPU algorithm Require: .N, {qi,j,t |i ∈ M, j ∈ N, t ∈ T}, B, , δ, ci,j0 = minj ci,j Ensure: r Initialization: t=0;.∀i ∈ M, ∀j ∈ N : zi,j,t = 0; .t = t + 1; Let .Bt = δB be the residual budget; 3: while .Bt ≥ ci,j0 do if .t ≤ N then .at = t; Pull the .at th arm; .∀j ∈ N : zi,j,t = zi,j,t−1 ; .zi,at ,t = zi,at ,t−1 + 1; 6: else for each .i ∈ N do 9: Compute .Ii,j,t−1 according to Definition 7.1; end for Solving the problem shown in Eqs. 25–26 to get .{zi,1,t + , ..., zi,N,t + }; 12: Pull the .at th arm with probability shown in Eq,27; .∀i ∈ N : zi,j,t = zi,j,t−1 zi,at ,t = zi,at ,t−1 + 1; end if .∀i ∈ N : rˆi,j,t = H(qi,j,1:t ); 15: .Bt+1 = Bt − cat ; .t = t + 1; end while 18: .r = i∈N rˆi,j,t−1 ;

N P{ai,t = i} = zi,j,t + /l=1 zi,j,t +

(48)

.

The detailed DPU algorithm is shown as Algorithm 3.

5.6 Performance Analysis of DPU algorithm Since the DPU algorithm adopts the same hybrid differentially private mechanism as DPF algorithm, we can get Theorem 4 without any proof: Theorem 4 The DPU algorithm is .-differentially private. In order to analyze the regret performance of DPU algorithm, we assume that the DPU algorithm terminates at time slot T . Thus, we can calculate the probability of pulling an arm: Lemma 5 .∀k ∈ N, ∀0 < t ≤ T , we can get P{ai,t = k|T } ≤ P{jˆi,t = k|T } +

.



ci,j∗ ci,j0

2

1 T −t +1

I where .jˆi,t = argmaxj i,j,t−1 ci ,j , ci,j∗ = maxj ci,j , and .ci,j0 = minj ci,j .

(49)

Unknown Worker Recruitment in Mobile Crowdsourcing

85

Proof According to the greedy selection strategy, DPU will first select the .jˆi,t -th B arm as most . c i,t  times and the residual budget is as most .cjˆi,t . Therefore, we have jˆi,t

cjˆ

j =jˆi,t zi,j,t+ ≤ ci,ji,t and .jN=1 zi,j,t + ≥ 0 largest cost. We can get

.

.

j =jˆi,t zi,j,t + jN=1 zi,j,t +

≤

cjˆi,t ci,j0

Bi,t ci,j∗

/

when DPU only selects the arm with

Bi,t ≤ ci,j∗



ci,j∗ ci,j0

2

ci,j0 Bi,t

(50)

In the process, DPU can still pull .T − t + 1 arms from time slot t, which means ci,j 1 Bi,t ≥ cai,t + cai,t , +...+, cai,T ≥ (T − t + 1)ci,j0 . Then, we can get . Bi,t0 ≤ T −t+1 and

.

.

j =jˆi,t zi,j,t + jN=1 zi,j,t +

 ≤

ci,j∗ ci,j0

2

1 T −t +1

Since DPU pulls the k-th arm with probability .P{ai,t = k} =

zi,j,t + N z l=1 i,j,t +

, we have

P{ai,t = k|Bi,t }

.

= P{ai,t = k, jˆi,t = k|Bi,t } + P{ai,t = k, jˆi,t = k|Bi,t } ≤

zjˆi,t P{jˆi,t N j =1 zi,j,t +

≤ P{jˆi,t = k|Bi,t } +

= k|Bi,t } +

j =jˆi,t zi,j,t + jN=1 zi,j,t +

j =jˆi,t zi,j,t + jN=1 zi,j,t +

P{jˆi,t = k|Bi,t }

≤ P{jˆi,t = k|Bi,t } +



ci,j∗ ci,j0

2

1 T −t +1 (51)

Besides, for any possible value of the residual budget .Bi,t , we can get P{ai,t = k|T } ≤ Bi,t P{ai,t = k|T , Bi,t }P{Bi,t |T } % &   ci,j∗ 2 1 ≤ Bi,t P{jˆi,t = k|T , Bi,t } + P{Bi,t |T } T −t +1 ci,j0

.

≤ P{jˆi,t = k|T } + Therefore, the lemma holds.



ci,j∗ ci,j0

2

1 T −t +1

(52)  

Lemma 6 Given that the DPU algorithm terminates at time slot T , .∀0 < , ρ < 1, we have

86

M. Xiao et al.

  ci,j∗ 2π 2 4 lnT + max{αk lnT , βk ln(4T )(logT + 1)} + .Ezk,T [T ] ≤ 1 + ci,j0 3 where .αk =

8 ρ 2 σk2 ci,k 2

, βk =

√ 2 8 (1−ρ)σk ci,j∗ , σi,k

ri,j∗ zi,j∗ ci,j∗

=

−

(53)

ri,k ci,k .

Proof According to Lemma 5, for a given T and .∀l ≥ 1, we have T N ˆ E[zi,j,T ] = 1 + t=N +1 P {ai,t = k} ≤ 1 + t=N +1 P{ji,t = k}   ci,j∗ 1 T + t=N +1 ci,j0 T − t + 1

.

(54)

where, for the second item, we have

 ≤

T t=N +1 P

min1≤zi,j∗ ,t≤t



Ii,j∗ ,t Ii,k,t ×P ≤ ci,k ci,j∗ rˆi,j,t zi,j,t

Note that .Ii,j,t =

+



Ii,j∗ ,t ci,j∗



 Ii,j∗ ,t Ii,k,t ≤ , zi,k,t ≥ l ci,j∗ ci,k  Ii,k,t T ≤ t=1 ≤ maxl≤zi,k,t ≤t z t t i,j∗ ,t=1 zi,k,t ci,k

T T ˆ t=N +1 P{ji,t = k, zi,k,t ≥ l} = t=N +1 P

.

(55)

2lnt zi,j,t

+

vi,t zi,j,t .

. t,n= 2lnt n

Let .b

We can observe that the

following must hold: .

2bt,zi,k,t ri,j∗ ri,j∗ Ii,j∗ ,t Ii,k,t ri,k ri,k 2vi,t ≤ ; ≥ ; < + + ci,j∗ zi,j∗ ci,j∗ ci,k zi,k ci,k zi,j∗ ci,j∗ zi,k ci,k ci,k zi,j∗ ,t ci,j∗ (56)

if .

Ii,j∗ ,t ci,j∗

 P

.

≤

Ii,k,t ci,k

holds. For the first inequality in Eq. (56)), we can get

Ii,j∗ ,t ri,j∗ ,t ci,j∗ ci,j∗

 = P{ˆri,j∗ ,t ≤ ri,j∗ ,t − vi,t or ri,j∗ ,t } ≤ P{ˆri,j∗ ,t ≤ ri,j∗ ,t − vi,t }   ri,j + P ri,j∗ ,t ≤ zi,j∗ ,t ∗ − zi,j∗ ,t bi,zi,j∗ ,t zi,j∗

(57)

Here, according to Lemmas 2 and 1, we have .P{ˆri,j∗ ,t ≤ ri,j∗ ,t − vi,t } ≤ γ , and r

∗ P{ri,j∗ ,t ≤ zi,j∗ ,t zi,j − zi,j∗ ,t by,zi,j∗ ,t } ≤ e i,j

.

get .P{

Ii,j∗ ,t ci,j∗

≤

∗ ri,j∗ zi,j∗ ci,j∗

2 −2bt,z

z i,j∗ ,t i,j∗ ,t

= t −4 . Therefore, we can

} ≤ t −4 + γ

Similarly, for the second inequality in Eq. (56)), we have .P{ t −4

+ γ . Let .γ =

t −4 ,

we can conclude

Ii,k,t ci,k

≥

ri,k zi,k ci,k }

≤

Unknown Worker Recruitment in Mobile Crowdsourcing



ri,j∗ Ii,j∗ ,t .P ≤ ci,j∗ zi,j∗ ci,j∗

 ≤ 2t

−4

87



Ii,k,t ri,k , and P ≥ ci,k zi,k ci,k



≤ 2t −4

(58) 2bt,zi,k,t ci,k

For the third inequality in Eq. (56)), if it is false, then we can get .σk ≥ 2vi,t zi,j∗ ci,j∗ .

+

For any .0 < ρ < 1, it is true if the following conditions hold: ρσi,k ≥

.

2bt,zi,k,t 2vi,t , (1 − ρ)σi,k ≥ zi,j∗ ci,j∗ ci,k

(59)

From the first and second conditions in Eq. (59)), we can obtain zi,k,t

.

Let .αi,k =

√ 8lnt 2 8ln(2t 4 ) ≥ and zi,j∗ ≥ (log t + 1) 2 c2 (1 − ρ)σi,k ci,j∗ ρ 2 ci,k i,k 8

2 c2 ρ 2 ci,k i,k

and .βi,k =

√ 2 8 (1−ρ)σi,k ci,j∗ .

(60)

We can get

T ˆ l + t=N +1 P{ji,t = k, zi,k,t ≥ l}

.

≤ max{αi,k lnT , βi,k ln(4T 4 )(logT + 1)} +

t T  

t 

2t −4

t=1 zi,j∗ ,t =1 zi,k,t=l

≤ 1 + max{αi,k lnT , βi,k ln(4T 4 )(logT + 1)} + ci,j∗ 2 1 i0 ) T −t+1 ≤ max{αi,k lnT , βi,k ln(4T 4 )(logT

T Since .t=N +1 (

holds.

2π 2 3

(61)

ci,j∗ 2 i0 ) lnT , we can conclude .E[zk,T |T ] ≤ 1 + 2 c ∗ 2 + 1)} + 2π3 + ( i,j i0 ) lnT . Therefore, Lemma 6

(

 

Lemma 7 The total time slot of running the DPU algorithm T is bounded by Bi − ci,j0 .E[T ] > − ci,j∗

 k:ci,k >ci,j∗

ci,k − ci,j∗ ci,j∗

% 1+

2π 2 + 3



ci,j∗ ci,j0

2

      B Bi Bi + max αi,k ln( log +1 , βi,k ln ci,j0 ci,j0 ci,j0

 ln

Bi ci,j0



(62)

T c Proof If the residual budget is no more than the minimal cost, i.e., .Bi − t=1 i,at < ci,j0 , the DPU algorithm terminates. According to Lemma 7, we can get T Bi − ci,j0 < E[t=1 ci,at |T ]

.

T ≤ E[{t=1 ci,j∗ |k:ci,k >ci,j∗ (ci,k − ci,j∗ )P{at = k|T }}|T ]

88

M. Xiao et al.

≤ E[T ]ci,j∗ + k:ci,k >ci,j∗ (ci,k − ci,j∗ )E[E[zi,k,T |T ]] '%     ci,j∗ 2 Bi 2π 2 + ≤ E[T ]ci,j∗ + k:ci,k >ci,j∗ (ci,k − ci,j∗ )E 1 + ln ci,j0 ci,j0 3       Bi Bi Bi + max αi,k ln( log +1 (63) , βi,k ln ci,j0 ci,j0 ci,j0 Substituting .T ≤

Bi ci,j0 ,

we can prove that Lemma 7 holds.

 

6 Conclusion In this chapter, we introduce the UWR issues into MCS systems where the platform needs to recruit workers with high qualities without knowing any information on workers’ qualities in advance. To address such a problem, we model it as a multiarmed bandit game and propose a UWR scheme based on an extended UCB bandit strategy, which is also extended to support the budget constraint. Furthermore, we propose a privacy-preserving UWR scheme to protect the privacy from being disclosed by using the differential privacy technique. In addition, we derive the regret bounds of the proposed three UWR schemes through the theoretical analysis, all of which achieve the sub-linear regret performances.

References 1. Y. Xu, M. Xiao, J. Wu, S. Zhang, G. Gao, Incentive mechanism for spatial crowdsourcing with unknown social-aware workers: a three-stage Stackelberg game approach. IEEE Trans. Mob. Comput. (2022). https://doi.org/10.1109/TMC.2022.3157687 2. J. Wang, F. Wang, Y. Wang, D. Zhang, L. Wang, Z. Qiu, Social-network-assisted worker recruitment in mobile crowd sensing. IEEE Trans. Mob. Comput. 18(7), 1661–1673 (2019) 3. Y. Xu, M. Xiao, A. Liu, J. Wu, Edge resource prediction and auction for distributed spatial crowdsourcing with differential privacy. IEEE Internet Things J. 9(17), 15554–15569 (2022) 4. Y. Wu, F. Li, L. Ma, Y. Xie, T. Li, Y. Wang, A context-aware multiarmed bandit incentive mechanism for mobile crowd sensing systems. IEEE Internet Things J. 6(5), 7648–7658 (2019) 5. M. Xiao, B. An, J. Wang, G. Gao, S. Zhang, J. Wu, CMAB-based reverse auction for unknown worker recruitment in mobile crowdsensing. IEEE Trans. Mob. Comput. 21(10), 3502–3518 (2022) 6. S. Yang, F. Wu, S. Tang, T. Luo, X. Gao, L. Kong, G. Chen, Selecting most informative contributors with unknown costs for budgeted crowdsensing, in 2016 IEEE/ACM 24th International Symposium on Quality of Service (IWQoS) (2016), pp. 1–6 7. C. Ho, J.W. Vaughan, Online task assignment in crowdsourcing markets, in Proceedings of the AAAI Conference on Artificial Intelligence (2012) 8. M. Xiao, J. Wu, H. Huang, L. Huang, C. Hu, Deadline-sensitive user recruitment for mobile crowdsensing with probabilistic collaboration, in 2016 IEEE 24th International Conference on Network Protocols (ICNP) (2016), pp. 1–10

Unknown Worker Recruitment in Mobile Crowdsourcing

89

9. Z. Wang, J. Zhao, J. Hu, T. Zhu, Q. Wang, J. Ren, C. Li, Towards personalized task-oriented worker recruitment in mobile crowdsensing. IEEE Trans. Mob. Comput. 20(5), 2080–2093 (2021) 10. E. Wang, Y. Yang, J. Wu, D. Luan, H. Wang, Worker recruitment strategy for self-organized mobile social crowdsensing, in 2018 27th International Conference on Computer Communication and Networks (ICCCN) (2018), pp. 1–9 11. T. Liu, Y. Zhu, L. Huang, TGBA: a two-phase group buying based auction mechanism for recruiting workers in mobile crowd sensing. Comput. Netw. 149, 56–75 (2019) 12. Y. Hu, J. Wang, B. Wu, S. Helal, Rl-recruiter+: Mobility-predictability-aware participant selection learning for from-scratch mobile crowdsensing. IEEE Trans. Mob. Comput. 21(12), 4555–4568 (2022) 13. Y. Hu, J. Wang, B. Wu, S. Helal, Rl-recruiter+: Mobility-predictability-aware participant selection learning for from-scratch mobile crowdsensing. IEEE Trans. Mob. Comput. 21(12), 4555–4568 (2021) 14. H. Wang, Y. Yang, E. Wang, W. Liu, Y. Xu, J. Wu, Combinatorial multi-armed bandit based user recruitment in mobile crowdsensing, in 2020 17th Annual IEEE International Conference on Sensing, Communication, and Networking (SECON) (2020), pp. 1–9 15. L. Pu, X. Chen, J. Xu, X. Fu, Crowdlet: optimal worker recruitment for self-organized mobile crowdsourcing, in IEEE INFOCOM 2016-The 35th Annual IEEE International Conference on Computer Communications (2016), pp. 1–9 16. M. Abououf, S. Singh, H. Otrok, R. Mizouni, E. Damiani, Machine learning in mobile crowd sourcing: a behavior-based recruitment model. ACM Trans. Internet Technol. (TOIT) 22(1), 16:1–16:28 (2022) 17. L. Wang, D. Yang, Z. Yu, Q. Han, E. Wang, K. Zhou, B. Guo, Acceptance-aware mobile crowdsourcing worker recruitment in social networks. IEEE Trans. Mob. Comput. 22(2), 634– 646 (2021) 18. G. Gao, J. Wu, M. Xiao, G. Chen, Combinatorial multi-armed bandit based unknown worker recruitment in heterogeneous crowdsensing, in IEEE INFOCOM 2020-IEEE Conference on Computer Communications (2020), pp. 179–188 19. M. Xiao, B. An, J. Wang, G. Gao, S. Zhang, J. Wu, Cmab-based reverse auction for unknown worker recruitment in mobile crowdsensing. IEEE Trans. Mob. Comput. 21(10), 3502–3518 (2022) 20. X. Chen, Q. Lin, D. Zhou, Optimistic knowledge gradient policy for optimal budget allocation in crowdsourcing, in International Conference on Machine Learning, ser. JMLR Workshop and Conference Proceedings, vol. 28 (2013), pp. 64–72 21. X. Gao, S. Chen, G. Chen, Mab-based reinforced worker selection framework for budgeted spatial crowdsensing. IEEE Trans. Knowl. Data Eng. 34(3), 1303–1316 (2022) 22. G. Gao, H. Huang, M. Xiao, J. Wu, Y.-E. Sun, Y. Du, Budgeted unknown worker recruitment for heterogeneous crowdsensing using CMAB. IEEE Trans. Mob. Comput. 21(11), 3895–3911 (2021) 23. N. Wang, J. Wu, Cost-efficient heterogeneous worker recruitment under coverage requirement in spatial crowdsourcing. IEEE Trans. Knowl. Data Eng. 7(2), 407–420 (2021) 24. X. Zhang, Z. Yang, Y. Liu, S. Tang, On reliable task assignment for spatial crowdsourcing. IEEE Trans. Emerg. Top. Comput. 7(1), 174–186 (2019) 25. M. Xiao, J. Wu, S. Zhang, J. Yu, Secret-sharing-based secure user recruitment protocol for mobile crowdsensing, in IEEE INFOCOM 2017-IEEE Conference on Computer Communications (2017), pp. 1–9 26. M. Xiao, G. Gao, J. Wu, S. Zhang, L. Huang, Privacy-preserving user recruitment protocol for mobile crowdsensing. IEEE/ACM Trans. Networking 28(2), 519–532 (2020) 27. L. Li, D. Shi, X. Zhang, R. Hou, H. Yue, H. Li, M. Pan, Privacy preserving participant recruitment for coverage maximization in location aware mobile crowdsensing. IEEE Trans. Mob. Comput. 21(9), 3250–3262 (2022) 28. M. Xiao, W. Jin, M. Li, L. Yang, A. Thapa, P. Li, Collusion-resistant worker recruitment in crowdsourcing systems. IEEE Trans. Mob. Comput. 22(1), 129–144 (2021)

90

M. Xiao et al.

29. Y. Han, H. Wu, Privacy-aware participant recruitment in opportunistic device to device networks. IEEE/ACM Trans. Networking 30(3), 1340–1351 (2022) 30. D. Wang, C. Huang, X. Shen, N. Xiong, A general location-authentication based secure participant recruitment scheme for vehicular crowdsensing. Comput. Netw. 171, 107152 (2020) 31. Z. Wang, J. Hu, R. Lv, J. Wei, Q. Wang, D. Yang, H. Qi, Personalized privacy-preserving task allocation for mobile crowdsensing. IEEE Trans. Mob. Comput. 18(6), 1330–1341 (2019) 32. T. Li, T. Jung, Z. Qiu, H. Li, L. Cao, Y. Wang, Scalable privacy-preserving participant selection for mobile crowdsensing systems: participant grouping and secure group bidding. IEEE Trans. Network Sci. Eng. 7(2), 855–868 (2018) 33. K. Yan, G. Lu, G. Luo, X. Zheng, L. Tian, A.M.V.V. Sai, Location privacy-aware task bidding and assignment for mobile crowd-sensing. IEEE Access 7, 131929–131943 (2019) 34. J. Lin, D. Yang, M. Li, J. Xu, G. Xue, Frameworks for privacy-preserving mobile crowdsensing incentive mechanisms. IEEE Trans. Mob. Comput. 17(8), 1851–1864 (2018) 35. J. Xu, G. Chen, Y. Zhou, Z. Rao, D. Yang, C. Xie, Incentive mechanisms for large-scale crowdsourcing task diffusion based on social influence. IEEE Trans. Veh. Technol. 70(4), 3731–3745 (2021) 36. P. Auer, N. Cesa-Bianchi, P. Fischer, Finite-time analysis of the multiarmed bandit problem. Mach. Learn. 47(2–3), 235–256 (2002) 37. S. Bubeck, N. Cesa-Bianchi, Regret analysis of stochastic and nonstochastic multi-armed bandit problems. Found. Trends Mach. Learn. 5(1), 1–122 (2012) 38. F. Li, J. Liu, B. Ji, Combinatorial sleeping bandits with fairness constraints. IEEE Trans. Network Sci. Eng. 7(3), 1799–1813 (2020) 39. E.L. Lawler, Fast approximation algorithms for knapsack problems. Math. Oper. Res. 4(4), 339–356 (1979) 40. W. Chen, Y. Wang, Y. Yuan, Combinatorial multi-armed bandit: general framework and applications, in International Conference on Machine Learning (ICML), ser. JMLR Workshop and Conference Proceedings, vol. 28 (2013), pp. 151–159 41. H. Zhao, M. Xiao, J. Wu, Y. Xu, H. Huang, S. Zhang, Differentially private unknown worker recruitment for mobile crowdsensing using multi-armed bandits. IEEE Trans. Mob. Comput. 20(9), 2779–2794 (2021) 42. A.C.Y. Tossou, C. Dimitrakakis, Algorithms for differentially private multi-armed bandits, in Proceedings of the AAAI Conference on Artificial Intelligence, ed. by D. Schuurmans, M.P. Wellman (2016), pp. 2087–2093 43. T.H. Chan, E. Shi, D. Song, Private and continual release of statistics. ACM Trans. Inf. Syst. Secur. (TISSEC) 14(3), 26:1–26:24 (2011)

Quality-Aware Incentive Mechanism for Mobile Crowdsourcing Haiming Jin and Lu Su

1 Introduction The ubiquity of human-carried mobile devices (e.g., smartphones, tablets, etc.) with a plethora of on-board and portable sensors (e.g., accelerometer, compass, camera, etc.) has given rise to the emergence of various people-centric mobile crowdsourcing (MCS) systems [1–3]. In a typical MCS system, a cloud-based platform aggregates and analyzes the sensory data provided by the public crowd instead of professionals and dedicatedly deployed sensors. The mobile devices of participating users collect and may process in certain level the data before submitting them to the platform. Such MCS systems hold a wide spectrum of applications including healthcare, ambient environment monitoring, smart transportation, indoor localization, etc. For example, MedWatcher [2] is a US FDA advocated MCS system for post-market medical device surveillance. Participating users upload photos of their medical devices to a cloud-based platform using the MedWatcher mobile application, which helps identify visible problems with the devices. The platform aggregates and analyzes user-provided information, sends reports to the FDA, and alerts users about medical device problems. Such a crowdsourcing paradigm enables easier detection of device safety issues and faster propagation of alerts to device users compared to traditional reporting methods such as mail or telephone. Moreover, air quality monitoring [3] is another area where MCS systems obtain their recent popularity. In such systems, crowdsourced air quality data are aggregated from a large number

H. Jin () Shanghai Jiao Tong University, Shanghai, China e-mail: [email protected] L. Su Purdue University, West Lafayette, IN, USA e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 J. Wu, E. Wang (eds.), Mobile Crowdsourcing, Wireless Networks, https://doi.org/10.1007/978-3-031-32397-3_4

91

92

H. Jin and L. Su

Fig. 1 A MedWatcher MCS system example (3 users try to upload the photos of the error message “Er3” on the screens of their blood glucose meters to the MedWatcher platform. The prices that the 3 users ask for cost compensation are 100$, 10$, and 1$, respectively)

of people using air quality sensors ported to their smartphones, which help estimate the city- or district-level air quality. Participating in such crowdsourcing tasks is usually a costly procedure for individual users. On the one hand, it consumes users’ resources, such as computing power, battery, and so forth. On the other hand, a considerable portion of sensing tasks requires the submission of some types of users’ sensitive private information, which causes privacy leakage for participating users. For example, by uploading photos of their medical devices, users reveal the types of their illnesses. By submitting air quality estimation samples, users usually reveal information about their locations. Therefore, without satisfactory rewards that compensate for participating costs and, without privacy-preserving methods that protect peoples personal information, users will be reluctant to carry out the sensing tasks. Aware of the paramount importance of stimulating user participation, the research community has recently developed some game-theoretic incentive mechanisms for MCS systems [4–18]. However, most of the existing mechanisms fail to incorporate one important aspect, that is users’ quality of information (QoI), into their designs. The meaning of QoI varies for different applications. For example, in the aforementioned MedWatcher system [2], QoI refers to the quality (e.g., resolution, contrast, sharpness, etc.) of uploaded photos. Higher quality ones will help the platform better identify visible device problems. As shown in the example in Fig. 1, QoI is also a major factor that should be considered together with the bidding price. Although user 1 has the highest quality photo, her high price prohibits the platform from requesting her data. Furthermore, despite user 3’s low price, the platform will not be interested in her data either because her low-quality photo could hardly contribute to identifying the error message “Er3.” By jointly considering price and QoI, the platform will select user 2 with medium price and acceptable photo quality as the data provider. In air quality monitoring MCS systems [3], QoI means a user’s estimation accuracy of air quality. The QoI of every user could be affected by various factors, including poor sensor quality, noise, lack of sensor calibration, and so forth. And, in this chapter, we will introduce a quality-aware incentive mechanism that considering QoI in the MCS system to obtain high-quality data with little cost.

Quality-Aware Incentive Mechanism for Mobile Crowdsourcing

93

Besides, among the various existing incentive mechanisms developed by the research community to stimulate user participation, one important category adopts the framework of reverse auction [19–34]. In these auction-based mechanisms, a user submits a bid to the platform containing one or multiple tasks he/she is interested in and his/her bidding price for executing these tasks. Based on users’ bids, the platform acting as the auctioneer determines the winners who are assigned to execute the tasks they bid on and the payments paid to the selected winners. Furthermore, designing a truthful auction where every user bids to the platform his/her true interested tasks and the corresponding true task execution cost is a common objective. However, all the aforementioned incentive mechanisms [19–34] fail to consider the preservation of users’ bid privacy. Although the platform is usually considered to be trusted, there exist some honest-but-curious users who strictly follow the protocol of the MCS system but try to infer information about other users’ bids. A user’s bid usually contains his/her private and sensitive information. For example, a user’s bidding task set could imply his/her personal interests, knowledge base, etc. In geotagging MCS systems that provide accurate localization of physical objects (e.g., automated external defibrillator [35], pothole [36, 37]), bidding task sets contain the places a user has visited or will visit, the disclosure of which breaches his/her location privacy. Similar to bidding task set, a user’s bidding price could also be utilized to infer his/her sensitive information. For example, bidding price could imply the type of mobile devices a user uses for an MCS task because usually users tend to bid more if their mobile devices are more expensive. Typically, the change in one user’s bid has the potential to shift the overall payment profile (i.e., payments to all users) significantly. It is possible that a curious user could infer information about other users’ bids from the different payments he/she receives in two rounds of the auction. To address this issue, we incorporate the notion of differential privacy [38, 39], which ensures that the change in any user’s bid will not bring a significant change to the resulting payment profile. Therefore, different from all existing incentive mechanisms for MCS systems, we also design a differentially private incentive mechanism that protects users’ bid privacy against honest-but-curious users. So, we will also introduce a quality-aware incentive mechanism that protects users’ bid privacy against honest-but-curious users in this chapter.

2 Related Work Game theory has been widely utilized to tackle networking problems such as spectrum sharing [40–44], cooperative communication [45, 46], channel and bandwidth allocation [47, 48], and so forth. Similar to many other problems, when it comes to incentive mechanism design in MCS systems, game-theoretic models are also frequently adopted by researchers due to their ability to capture and tackle users’ strategic behaviors.

94

H. Jin and L. Su

Yang et al. [4] designed incentive mechanisms using auction and Stackelberg game for both user-centric and platform-centric models. Their auction-based mechanism, however, did not consider users’ untruthful behaviors about bidding task sets. Duan et al. [5] propose a Stackelberg game-based incentive mechanism which is similar to [4] but able to deal with the asymmetrically incomplete information between users and the platform. Faltings et al. [6] designed an incentive mechanism to ensure the truthfulness of reported data without considering users’ strategic behaviors about bidding prices and task sets. Zhang et al. [49] designed an incentive mechanism specifically for binary crowd labelling systems. Furthermore, [7, 8] designed social cost minimizing incentive mechanisms. Zhang and van der Schaar [9], Xie and Lui [10], and Xie et al. [11] studied crowdsourcing systems with multiple task requesters and users. Zhang et al. [9] proposed a reputation-based incentive protocol for crowdsourcing applications and used repeated gift-giving game to model the interaction between task requesters and users. In [10, 11], the authors integrated the effort and data quality of users into their mechanisms. Another series of work [12–14] utilized online auctions to design incentive mechanisms in crowdsourcing systems where users arrive sequentially. Zhang et al. [50] studied MCS systems based on smartphones and designed an incentive mechanisms with privacy techniques. Zhang et al. [51] studied the privacy compensation for continuous data sensing in MCS systems. Zhao et al. [52] proposed a privacy-preserving and data quality-aware incentive scheme in MCS systems. Liu et al. [53] proposed an incentive mechanism for participants, aiming to protect them from privacy leakage, to ensure the availability of sensing data.

3 Quality-Aware Incentive Mechanisms for MCS Systems In this section, we first introduce a quality ware mechanisms incorporating a crucial metric, called QoI for MCS Systems, and present an overview of MCS systems, our auction model, and design objectives.

3.1 System Overview The MCS system model considered in this chapter consists of a platform residing in the cloud and a set of N users, denoted as .N = {1, · · · , N}. The users execute a set of M sensing tasks, denoted as .T = {τ1 , · · · , τM }, and send their sensory data to the platform. The workflow of the system is described as follows: 1. Firstly, the platform announces the set of sensing tasks, .T, to users. 2. Then, the platform and users enter the auctioning stage in which the platform acts as the auctioneer that purchases the sensory data collected by individual users. Every user .i ∈ N submits his/her bid, which is a tuple .(i , bi ) consisting of

Quality-Aware Incentive Mechanism for Mobile Crowdsourcing

95

the set of tasks .i ⊆ T he/she wants to execute and his/her bidding price .bi for executing these tasks. 3. Based on users’ bids, the platform determines the set of winners, denoted as .S ⊆ → p = {p1 , · · · , pN }. Specifically, a N, and the payment to all users, denoted as .− loser does not execute any task and receives zero payment. 4. After the platform receives winners’ sensory data, it gives the payment to the corresponding winners. One major difference between our work and most of the previous work is that we integrate the quality of information (QoI) corresponding to every user, → q = {q1 , · · · , qN }, into our incentive mechanisms. Generally speaking, denoted as .− QoI indicates the quality of users’ sensory data. The definition of QoI varies for different applications. For example, in the MedWatcher system [2], QoI refers to the quality (e.g., resolution, contrast, sharpness, etc.) of uploaded photos. Photos with higher quality will help the platform better identify visible problems with medical devices. In air quality monitoring MCS systems [3], QoI refers to a user’s estimation accuracy of air quality. We assume that the platform maintains a historical record → of users’ QoI profile .− q used as inputs for winner and payment determination. There are many methods for the platform to calculate users’ QoIs. Intuitively, in the cases where the platform has an adequate amount of ground truth data, QoIs can be obtained by directly calculating the deviation of users’ data from the ground truths. However, even without ground truths, QoIs can still be effectively inferred from users’ data by utilizing algorithms such as those proposed in [54–57]. Alternatively in many applications, QoIs can be inferred from other factors (e.g., the price of a user’s sensors, his/her experience and reputation of executing specific sensing tasks, etc.) using methods proposed in previous studies such as [58]. The problem of which method the platform adopts to calculate users’ QoIs is application dependent and out of the scope of this chapter. Typically, users might know some of the factors that affect their QoIs. However, users usually do not know exactly how QoIs are calculated by the platform. Hence, they do not know the exact values of their QoIs.

3.2 Auction Model In this chapter, we consider strategic and selfish users that aim to maximize their own utilities. The fact that users bid on subsets of tasks motivates us to use reverse combinatorial auction to model the problem. In the rest of the chapter, we use bundle to refer to any subset of tasks of .T. Different from traditional forward combinatorial auction [59, 60], we formally define the concept of reverse combinatorial auction for our problem setting in Definition 1. Definition 1 (RC Auction) In a reverse combinatorial auction (RC auction), each user .i ∈ N is interested in a set of .Ki ≥ 1 bundles, denoted as .Ti = {i1 , · · · , iKi }. For any bundle . ⊆ T, the user has a cost function defined in Eq. (1).

96

H. Jin and L. Su

 Ci () =

j

j

ci ,

if ∃i ∈ Ti s.t.  ⊆ i

+ ∞,

otherwise.

.

(1)

Both .Ti and the cost function .Ci (·) are user i’s private information. If .Ki = 1 for every user, then the auction is defined as a single-minded reverse combinatorial auction (SRC auction). And it is defined as a multi-minded reverse combinatorial auction (MRC auction), if .Ki > 1 for at least one user. In an SRC auction, .Ti contains only user i’s maximum executable task set . i . That is, . i consists of all the sensing tasks that user i is able to execute. Since he/she is not capable to carry out tasks beyond . i , his/her cost for any bundle . ⊆  i can be equivalently viewed as .+∞. Similarly in an MRC auction, the union of  i j all the bundles in .Ti is . i . That is, . K j =1 i =  i . If user i is a winner of the RC auction, he/she will be paid .pi for executing the corresponding set of sensing tasks. In contrast, he/she will not be allocated any sensing task and will receive zero payment if he/she is a loser. We present the definitions of the utility of a user and the profit of the platform formally in Definitions 2 and 3. Definition 2 (A User’s Utility) The utility of any user .i ∈ N is  ui =

pi − ci , if i ∈ S

(2)

.

0,

otherwise.

Definition 3 (Platform’s Profit) The profit of the platform given users’ QoI profile − → q is

.

.

→ u0 = V− q (S) −



pi ,

(3)

i∈S N + → where the value function .V− q (·) : 2 → R maps the winner set .S to the value − → → that the winners bring to the platform. Furthermore, .V− q (·) is monotonic in . q . That → → } such that .q ≥ q holds q = {q1 , · · · , qN } and .− q = {q1 , · · · , qN is, for any .− i i → − → .∀i ∈ N, we have .V− (S) ≥ V (S). q q

Similar to the traditional VCG mechanism design [61, 62], we aim to design mechanisms that maximize the social welfare, which is formally defined in Definition 4. Definition 4 (Social Welfare) The social welfare of the whole MCS system is .

usocial = u0 +

 i∈N

→ ui = V− q (S) −

 i∈S

ci .

(4)

Quality-Aware Incentive Mechanism for Mobile Crowdsourcing

97

3.3 Design Objective In this chapter, we aim to design dominant-strategy mechanisms in which for every user there exists a dominant strategy [63] defined in Definition 5. Definition 5 (Dominant Strategy) A strategy .sti is the dominant strategy for user i if and only if for any other strategy .st i and any strategy profile of the other users, denoted as .st−i , the property .ui (sti , st−i ) ≥ ui (st i , st−i ) holds. In our SRC auction, each user submits to the platform a bid .(i , bi ) consisting of his/her declared interested bundle .i and the bidding price .bi . Since users are strategic, it is possible that he/she declares a bid that deviates from the true value .( i , ci ). However, one of our goals for the SRC auction is to design a truthful mechanism defined in Definition 6. Definition 6 (Truthfulness) An SRC auction is truthful if and only if it is the dominant strategy for every user .i ∈ N to bid his/her true value .( i , ci ). Noticed from Definition 6 that we aim to ensure the truthfulness of both the cost .ci and bundle . i . Besides truthfulness, another design objective for the SRC auction is to ensure that every user receives non-negative utility from participating. Such property is critical in incentive mechanisms because it ensures that users will not be disincentivized to participate for receiving negative utilities. This property is defined as individual rationality in Definition 7. Definition 7 (Individual Rationality) A mechanism is individual rational (IR) if and only if .ui ≥ 0 is satisfied for every user .i ∈ N. As mentioned in Sect. 3.2, our mechanism aims to maximize the social welfare. However, as will be proved in Sect. 3.4, the problem of maximizing the social welfare in the SRC auction is NP-hard. Hence, we aim to design a polynomial-time mechanism that gives us approximately optimal social welfare with a guaranteed approximation ratio. In the domain of multi-minded combinatorial auction, requiring truthfulness limits the family of mechanisms that can be used, as pointed out in [64]. Hence, in our MRC auction, we aim to design a dominant-strategy mechanism that can still yield a guaranteed approximation ratio to the optimal social welfare without ensuring truthfulness. In fact, as mentioned in [60], the requirement of truthfulness is only to obtain close-to-optimal social welfare with strategic user behaviors, but not the real essence. Therefore, as long as the approximation ratio is guaranteed when users play their dominant strategies, it is justifiable for us to relax the truthfulness requirement. Additionally, we also require our mechanism to be individual rational and have a polynomial computational complexity. The authors in [60, 65] address the issue of mechanism design for multi-minded forward combinatorial auctions. Their mechanisms cannot ensure that the users have dominant strategies and cannot be applied to reverse combinatorial auctions. However, in contrast, we are able to design a dominant-strategy incentive mechanism for

98

H. Jin and L. Su

Table 1 Summary of design objectives Model SRC MRC

Dominant strategy √ √

Truthful √

IR √ √

.

.

.

.

×

.

Approx. ratio Guaranteed Guaranteed

Complexity Polynomial Polynomial

the MRC auction in this chapter. We summarize our design objectives for both the SRC and MRC auctions in Table 1.

3.4 SRC Auction In this section, we introduce the mathematical formulation, mechanism design, an intuitive walk-through example and the corresponding analysis for the SRC auction.

3.4.1

Mathematical Formulation

In our SRC auction, each user’s bid .(i , bi ) consists of his/her declared interested bundle .i and the bidding price .bi . Although our model is valid for any general value → function .V− q (·) that satisfies Definition 3, to simplify our analysis we assume that → .V− (·) is the sum of the value, .vi , contributed by every winner .i ∈ S. Furthermore, q we assume that .vi is proportional to the total QoI provided by this user. Given users’ − → bidding bundle profile .  = {1 , · · · , N } and the winner set .S, the platform’s → value function .V− q (·) can be represented by → V− q (S) = .



vi =

i∈S



αqi |i |,

(5)

i∈S

where .α is a coefficient that transforms QoI to monetary reward. Another aspect that distinguishes our work from previous work is that we consider QoI coverage in the SRC auction. For the task that none of the users capable to execute it has adequately high QoI, collective efforts of multiple users → are necessary to ensure sensing quality. We use .Qτj ,− q (S) to denote the total QoI → that all winners have on task .τj ∈ T. Furthermore, we approximate .Qτj ,− q (S) as the sum of the QoIs of the winners that execute this task. Therefore, QoI coverage is equivalent to guaranteeing that every task is executed by users with sufficient → amount of QoI in total. Based on this additive assumption of QoI, .Qτj ,− q (S) can be represented by .

→ Qτj ,− q (S) =

 i:τj ∈i ,i∈S

qi .

(6)

Quality-Aware Incentive Mechanism for Mobile Crowdsourcing

99

Since we aim to maximize the social welfare given in Definition 4, the winner determination and pricing can be decoupled into two separate problems. We formulate the SRC auction winner determination (SRC-WD) problem as the following integer linear program. SRC-WD Problem:  . max (αqi |i | − bi )xi. (7) i∈N

s.t.



qi xi ≥ Qj ,

∀τj ∈ T.

(8)

∀i ∈ N.

(9)

i:τj ∈i ,i∈N

xi ∈ {0, 1},

Constants The SRC-WD input constants .α, users’ bid profile   problem takes as → (1 , b1 ), · · · , (N , bN ) , users’ QoI profile .− q , and tasks’ QoI requirement profile − → . Q = {Q1 , · · · , QM }. .

Variables In the SRC-WD problem, we have a set of binary variables .{x1 , · · · , xN } for every user .i ∈ N. If user i is in the winner set .S, then .xi = 1. Otherwise, .xi = 0. Objective Function Since the platform does not know the true values of users’  interested bundles and the corresponding costs, . ( 1 , c1 ), · · · , ( N , cN ) , the objective function that it directly tries to maximize is the social welfare based on  → w = {w1 , · · · , wN }, in which users’ bid profile . (1 , b1 ), · · · , (N , bN ) . We use .− .wi = αqi |i |−bi , to denote the marginal social welfare profile of all users based on   users’ bids. Then, we have the objective function . i∈S wi = i∈S (αqi |i |−bi ) =  i∈N (αqi |i | − bi )xi . Constraints Constraint 8 represents the QoI coverage for every task .τj ∈ T, which → ensures that the total QoI of all the winners for this task, calculated as .Qτj ,− q (S) =   q = q x , is no less than the QoI requirement . Q j. i:τj ∈i ,i∈S i i:τj ∈i ,i∈N i i Next, we prove the NP-hardness of the SRC-WD problem. Theorem 1 The SRC-WD problem is NP-hard. Proof In this proof, we demonstrate that the NP-complete minimum weight set cover (MWSC) problem is polynomial-time reducible to the SRC-WD problem. The reduction starts with an instance of the MWSC problem consisting of a universe of elements .U = {τ1 , · · · , τM } and a set of N sets .O = {1 , · · · , N } whose union equals .U. Every set .i ∈ O is associated with a non-negative weight .wi . The MWSC problem is to find the subset of .O with the minimum total weight whose union contains all the elements in .U. Based on the instance of the MWSC problem, we construct an instance of the SRC-WD problem. Firstly, we transform .i into .i such that for every element in .i there exist .li ∈ Z+ copies of the same element in .i . We require that every element + times. After the reduction, we obtain an .τj ∈ U is covered for at least .Lj ∈ Z

100

H. Jin and L. Su

→ instance of the SRC-WD problem in which users’ QoI profile is .− q = {l1 , · · · , lN }, − → users’ bidding bundle profile is .  = {1 , · · · , N }, users’ marginal social welfare − → → profile is .− w = {−w1 , · · · , −wN }, and tasks’ QoI requirement profile is . Q = {L1 , · · · , LM }. Noticed that the SRC-WR problem represents a richer family of problems in which any user i’s QoI, .qi , and any task j ’s QoI requirement, .Qj , could take any value in .R+ . Furthermore, the marginal social welfare can take any value in .R. Hence, every instance of the MWSC problem is polynomial-time reducible to an instance of the SRC-WD problem. The SRC-WD problem is NP-hard.



3.4.2

Mechanism Design

Because of the NP-hardness of the SRC-WD problem, it is impossible to compute the set of winners that maximize the social welfare in polynomial time unless .P = NP. As a result, we cannot use the off-the-shelf VCG mechanism [61, 62] since the truthfulness of VCG mechanism requires that the social welfare is exactly maximized. Therefore, as mentioned in Sect. 3.3, we aim to design a mechanism that approximately maximizes the social welfare while guaranteeing truthfulness. Myerson’s characterizations of truthfulness for single-parameter auctions [66] are not directly applicable in our scenario because our SRC auction is a doubleparameter auction that considers both bundle and cost truthfulness. Moreover, different from the characterizations of truthfulness for single-minded forward combinatorial auctions proposed in [59], we describe and prove the necessary and sufficient conditions for a truthful SRC auction in Lemma 1. Lemma 1 An SRC auction is truthful if and only if the following two properties hold: • Monotonicity. Any user i who wins by bidding .(i , bi ) still wins by bidding any .b < bi and any . ⊃ i given that other users’ bids are fixed. i i • Critical payment. Any winner i with bid .(i , bi ) is paid the supremum of all bidding prices .bi such that bidding .(i , bi ) still wins, which is defined as user i’s critical payment. Proof It is easily verifiable that a truthful bidder will never receive negative utility. If user i’s any untruthful bid .(i , bi ) is losing or .i ⊆  i , his/her utility from bidding .(i , bi ) will be non-positive. Therefore, we only need to consider the case in which .(i , bi ) is winning and .i ⊆  i . • Because of the property of monotonicity, .( i , bi ) is also a winning bid. Suppose the payment for bid .(i , bi ) is p and that for bid .( i , bi ) is .p. Every bid .( i , bi ) with .bi > p is losing because .p is the user i’s critical payment given bundle . i . From monotonicity, bidding .(i , b ) is also losing. Therefore, the critical i payment for .(i , bi ) is at most that for .( i , bi ), which means .p ≤ p. Hence, the user will not increase his/her utility by bidding .(i , bi ) instead of .( i , bi ).

Quality-Aware Incentive Mechanism for Mobile Crowdsourcing

101

Algorithm 1: QoI-SRC Auction Winner Determination − → − → → → Input: T, N, − w,− q , Q,  ; Output: S; // Initialization 1 N− ← ∅, S ← ∅; // Select non-negative marginal social welfare users 2 foreach i s.t. wi ≥ 0 do 3 S ← S ∪ {i};

4 N− ← N \ S; // Calculate residual QoI requirement

5 foreach j s.t. τj ∈ T do  6 Q j ← Qj − min{Qj , i:τj ∈i ,i∈S qi }; // Main  loop j :τj ∈T Qj = 0 do // Find the user with the minimum marginal social welfare effectiveness |wi | 8 l = arg mini∈N−  ; min{Q ,q }

7 while

j :τj ∈i

9 10 11 12

j

i

S ← S ∪ {l}; N− ← N− \ {l}; // Update residual requirement foreach j s.t. τj ∈ T do Q j ← Q j − min{Q j , ql };

13 return S;

• Then, we consider the case in which bidding truthfully .( i , ci ) wins. This bid earns the same payment .p as .( i , bi ). Then, his/her utilities from these two bids will be the same. If bidding .( i , ci ) loses, then we have .ci > p ≥ bi . Hence, bidding .( i , bi ) will receive negative utility. Therefore, .( i , bi ) will also not increase his/her utility compared to .( i , ci ). Thus, we conclude that an SRC auction is truthful if and only if the monotonicity and critical payment properties hold.

We utilize the rationale provided in Lemma 1 to design a quality of information aware SRC (QoI-SRC) auction. Specifically, we present the winner determination and pricing mechanisms of the QoI-SRC auction, respectively, in Algorithms 1 and 2. → w using users’ The marginal social welfare profile .−  platform calculates users’  − → bids . (1 , b1 ), · · · , (N , bN ) and utilizes . w as input to the winner determination algorithm shown in Algorithm 1. Firstly, the platform includes all users with nonnegative marginal social welfare into the winner set .S (lines 2 and 3). By removing the current winners from .N, the platform gets the set of users .N− with negative marginal social welfare (line 4). Then, the platform calculates tasks’ residual QoI − → − → requirement profile . Q by subtracting from . Q the QoI provided by the currently

102

H. Jin and L. Su

Algorithm 2: QoI-SRC Auction Pricing − → → → Input: S, α, − q ,− w, ; − → Output: p ; // Initialization → 1 N+ ← ∅, − p ← {0, · · · , 0}; // Find non-negative marginal welfare users 2 foreach i s.t. wi ≥ 0 do 3 N+ ← N+ ∪ {i}; // Main loop

4 foreach i ∈ S do  5 run Algorithm 1 on N \ {i} until j :τj ∈i Q j = 0; 6 S ← the winner set when step 5 stops; 7 8 9 10 11 12

// Calculate payment if |S | < |N+ | then pi ← αqi |i |; else foreach k ∈ S \ N+ do − → Q ← tasks’ residual QoI requirement profile when winner k is selected;   j :τ ∈ min{Qj ,qi } pi ← max pi , αqi |i | − wk  j i min{Q ,q } ; j :τj ∈k

j

k

→ 13 return − p;

selected winners (lines 5 and 6). The main loop (lines 7–12) is executed until every task’s QoI requirement is satisfied. In the main loop, winner selection is based on marginal social welfare effectiveness (MSWE), defined as the ratio between the absolute value of user i’s marginal social welfare .|wi | and his/her effective QoI contribution . j :τj ∈i min{Q j , qi }. In every iteration, the user with the minimum MSWE among the remaining users in .N− is included into .S (lines 8–9). After that, − → the platform updates .N− and tasks’ residual QoI requirement profile . Q (lines 10– 12). Algorithm 2 describes the corresponding pricing mechanism. It takes the winner → → p . Firstly, .− p is initialized as set .S as input and outputs the payment profile .− a zero vector (line 1). Then, the platform includes all users with non-negative marginal social welfare into .N+ (line 2–3). The main loop (lines 4–12) calculates the platform’s payment to every winner. For every winner .i ∈ S, the winner determination mechanism in Algorithm 1 is executed with all users except user i until the QoI requirement of every task in .i has been fully satisfied (line 5). We reach the point such that it is impossible for user i to be selected as a winner in future iterations of Algorithm 1. Then, the platform gets the current winner set .S (line 6) and calculates .pi differently in the following two cases: • Case 1 (lines 7 and 8). Any winner i belonging to case 1 has .wi ≥ 0. Hence, this user’s critical payment is the bidding price .bi that satisfies .wi = αqi |i |−bi = 0. That is, .pi = αqi |i |.

Quality-Aware Incentive Mechanism for Mobile Crowdsourcing

103

• Case 2 (lines 10 and 11). For any winner i belonging to case 2, we go through every user .k ∈ S \ N+ . We calculate user i’s maximum bidding price .bi to be able to substitute user k as the winner. That is, .bi satisfies Eq. (10). .

bi − αqi |i | |wk |  . ,q } = min{Q j :τj ∈i j :τj ∈k min{Qj , qk } j i



(10)

This means  . bi

= αqi |i | − wk 

j :τj ∈i

min{Q j , qi }

j :τj ∈k

min{Q j , qk }

.

(11)

Finally, the maximum value among all .bi ’s is used as the payment to user i.

4 Quality-Aware Incentive Mechanism Considering the Bid Privacy for MCS Systems In this section, we introduce a quality-aware mechanism considering the bid privacy and present an overview of MCS systems, the aggregation method, our auction model, and design objectives.

4.1 System Overview The MCS system considered in this chapter consists of a cloud-based platform and a set of N participating users denoted as .N = {w1 , · · · , wN }. In this chapter, we are particularly interested in MCS systems that host a set of K classification tasks, denoted as .T = {τ1 , · · · , τK }, namely ones that require users to locally decide the classes of the objects or events he/she has observed, and report his/her local decisions (i.e., labels of the observed objects or events) to the platform. Here, we assume that all tasks in .T are binary classification tasks, which constitute a significant portion of the tasks posted on MCS platforms. Examples of such tasks include tagging whether or not a segment of road surface has potholes or bumps [36, 37], labeling whether or not traffic congestion happens at a specific road segment [67], etc. Each binary classification task .τj ∈ T has a true class label .lj , unknown to the platform, which is either .+1 or .−1. If user .wi is selected to execute task .τj , he/she will provide a label .li,j to the platform. Currently, a major challenge in designing reliable MCS systems lies in the fact that the sensory data provided by individual users are usually unreliable due to various reasons including carelessness, background noise, lack of sensor calibration, poor sensor quality, etc. To overcome this issue, the platform has to aggregate

104

H. Jin and L. Su

the labels provided by multiple users, as this will likely cancel out the errors of individual users and infer the true label. We describe the workflow1 of the MCS system as follows: • The platform firstly announces the set of binary classification tasks, .T, to the users. • Then, the users and the platform start the auctioning stage, where the platform acts as the auctioneer purchasing the labels provided by the users. Every user .wi submits his/her bid .bi = (i , ρi ), which is a tuple consisting of the set of tasks .i he/she wants to execute and his/her bidding price .ρi for providing labels about these tasks. We use .b = (b1 , · · · , bN ) to denote users’ bid profile. • Based on users’ bids, the platform determines the set of winners (denoted as .S ⊆ N) and the payment .pi paid to each user .wi . We use .p = (p1 , · · · , pN ) to denote users’ payment profile. • After the platform aggregates users’ labels to infer the true label of every task, it gives the payment to the corresponding winners. Every user .wi has a reliability level .θi,j ∈ [0, 1] for task .τj , which is the probability that the label .li,j provided by user .wi about task .τj equals the true label .lj , i.e., .Pr[li,j = lj ] = θi,j . We use the matrix .θ = [θi,j ] ∈ [0, 1]N ×K to denote the reliability level matrix of all users. We assume that the platform maintains a historical record of the reliability level matrix .θ utilized as one of the inputs for winner and payment determination. There are many methods that the platform could use to estimate .θ. In the cases where the platform has access to the true labels of some tasks a priori, it can assign these tasks to users in order to estimate .θ as in [68]. When ground truth labels are not available, .θ can still be effectively estimated from users’ previously submitted data using algorithms such as those in [54–57, 69, 70]. Alternatively, in many applications, .θ can be inferred from some explicit characteristics of the users (e.g., a user’s reputation and experience of executing certain types of sensing tasks, the type and price of a user’s sensors) using the methods proposed in [58]. The issue of exactly which method is used by the platform to calculate .θ is application dependent and out of the scope of this chapter.

4.2 Aggregation Method In this chapter, we reasonably assume that the platform utilizes a weighted .lj for each task .τj based on aggregation method to calculate the aggregated label

the labels collected from users. That is,

1 Note that we are specifically interested in the scenario where all users and tasks arrive at the same time. We leave the investigation of the online scenario where users and tasks arrive sequentially in an online manner in our future work.

Quality-Aware Incentive Mechanism for Mobile Crowdsourcing

. lj = sign



105





αi,j li,j ,

(12)

i:wi ∈S,τj ∈i

where .αi,j is the weight corresponding to the label .li,j . In fact, many sophisticated state-of-the-art data aggregation mechanisms, such as those proposed in [54–56, 69, 70], also adopt the weighted aggregation method to calculate the aggregation results. Given the aggregation method, the platform selects winners so that the aggregation error of each task .τj ’s label is upper bounded by a predefined threshold .δj . That is,

j = lj ] ≤ δj holds for every task .τj ∈ T, where the platform aims to ensure that .Pr[L

.Lj denotes the random variable corresponding to

.lj . We directly apply with minor adaptation in this chapter the results derived in [33] (Theorem 1 and Corollary 1), formally summarized in Lemma 2, regarding the relationship between the selected winners’ reliability levels and the upper bounds of tasks’ aggregation error. Lemma 2 If the platform utilizes a weighted aggregation method that calculates each task .τj ’s aggregated label

.lj according to Eq. (12) with .αi,j = 2θi,j − 1, and  .

i:wi ∈S,τj ∈i

1 , − 1) ≥ 2 ln δj 

(2θi,j

2

(13)

j = lj ] ≤ δj . where .δj ∈ (0, 1), then we have that .Pr[L We refer to Eq. (13) as the error bound constraint in the rest of this chapter.

j = Essentially, Lemma 2 presents a necessary and sufficient condition for .Pr[L algorithm. That lj ] ≤ δj to hold for each task .τj ∈ T for a weighted aggregation

 is, the aggregated label

.lj should be calculated as

.lj = sign i:wi ∈S,τj ∈i (2θi,j −  2 1)li,j and the sum of the values .(2θi,j − 1) ’s for all winner .wi ’s that execute task

1 .τj should not be smaller than the threshold .2 ln δj . Intuitively, the larger the value (2θi,j − 1)2 is, the more informative the label .li,j will be to the platform. When the value .(2θi,j − 1)2 approaches 0, or equivalently .θi,j approaches 0.5, the label .li,j will be closer to a random noise.

.

4.3 Auction Model In the rest of the chapter, we will refer to any subset of tasks of .T as a bundle. Since in the MCS system considered in this chapter every user bids on one bundle of tasks, we use single-minded reverse combinatorial auction with heterogeneous cost (hSRC auction), formally defined in Definition 8, to model the problem. Definition 8 (hSRC Auction) We define the single-minded reverse combinatorial auction with heterogeneous cost, namely hSRC auction, as follows. In the hSRC auction, any user .wi has a set of .Ki possible bidding bundles denoted as .Ti =

106

H. Jin and L. Su

{i,1 , · · · , i,Ki }. For providing labels about all the tasks in each bundle .i,k ∈ Ti , the user has a cost .ci,k . Furthermore, every user .wi is only interested in one of the bundles in .Ti , denoted as .i∗ with cost .ci∗ . Note that the hSRC auction defined in Definition 8 is a generalization of traditional single-minded combinatorial auctions, such as those in [59, 60]. Typically, in traditional single-minded combinatorial auctions, all the possible bidding bundles of a user have the same cost. However, in our hSRC auction, the cost .ci,k ’s for every bundle .i,k ∈ Ti do not necessarily have to be the same. In MCS systems, users usually have different costs for executing different bundles, which makes our definition of hSRC auction more suitable to the problem studied in this chapter. In Definitions 9, we define a user’s truthful bid. Definition 9 (Truthful Bid) We define bid .bi∗ = (i∗ , ci∗ ) which contains user .wi ’s true interested bundle .i∗ and the corresponding cost .ci∗ as her truthful bid. In Definitions 10 and 11, we present the formal definitions of a user’s utility and the platform’s total payment. Definition 10 (User’s Utility) Suppose a user .wi bids .i,k ∈ Ti in the hSRC auction. If he/she is a winner, he/she will be paid .pi by the platform. Otherwise, he/she will not be allocated any task and receives zero payment. Therefore, the utility of the user .wi is  ui =

pi − ci,k ,

if wi ∈ S

0,

otherwise.

.

(14)

Definition 11 (Platform’s Payment) The platform’s total payment to all users given the payment profile .p and the winner set .S is R(p, S) = .

 i:wi ∈S

pi .

(15)

4.4 Design Objective Since users are strategic in our hSRC auction, it is possible that a user could submit a bid different from the truthful bid defined in Definition 9 in order to obtain more utility. To address this problem, one of our goals is to design a truthful mechanism, where every user maximizes his/her utility by bidding his/her truthful bid regardless of other users’ bids. In practice, ensuring exact truthfulness for the hSRC auction is too restrictive. Therefore, we turn to a weaker but more practical notion of .γ truthfulness in expectation [39, 71], formally defined in Definition 12.

Quality-Aware Incentive Mechanism for Mobile Crowdsourcing

107

Definition 12 (.γ -truthfulness) An hSRC auction is .γ -truthful in expectation, or γ -truthful for short, if and only if for any bid .bi = bi∗ and any bid profile of other users .b−i , there is

.

.

    E ui (bi∗ , b−i ) ≥ E ui (bi , b−i ) − γ ,

(16)

where .γ is a small positive constant. .γ -truthfulness ensures that no user is able to make more than a slight .γ gain in his/her expected utility by bidding untruthfully. Therefore, we reasonably assume that each user .wi would bid his/her truthful bid .bi∗ , if our hSRC auction satisfies .γ -truthfulness. Apart from .γ -truthfulness, another desirable property of our hSRC auction is individual rationality, which implies that no user has negative utility. This property is crucial in that it prevents users from being disincentivized by receiving negative utilities. We formally define this property in Definition 13.

Definition 13 (Individual Rationality) An hSRC auction is individual rational if and only if .ui ≥ 0 holds for every user .wi ∈ N. Simply paying users according to the output payment profile of the auction poses threats to the privacy of users’ bids. Because the change in one user’s bid has the potential to shift the payment profile significantly, it is possible for a curious user to infer other users’ bids from the different payments he/she receives in two rounds of auction. Therefore, we aim to design a differentially private mechanism [38, 39], formally defined in Definition 14. Definition 14 (Differential Privacy) We denote the proposed hSRC auction as a function .M(·) that maps an input bid profile .b to a payment profile .p. Then, .M(·) is . -differentially private if and only if for any possible set of payment profiles .A and any two bid profiles .b and .b that differ in only one bid, we have .

    Pr M(b) ∈ A ≤ exp( )Pr M(b ) ∈ A ,

(17)

where . is a small positive constant usually referred to as privacy budget. Differential privacy ensures that the change in any user’s bid will not bring a significant change to the resulting payment profile. Hence, it is difficult for the curious users to infer information about other users’ bids from the outcome (i.e., payment profile) of the mechanism. In this chapter, to achieve differential privacy, we introduce randomization to the outcome of our mechanism, similar to [39, 72, 73]. In short, we aim to design a .γ -truthful, individual rational, and . -differentially private incentive mechanism in this chapter.

108

H. Jin and L. Su

4.5 Mathematical Formulation In this section, we present our formal mathematical problem formulation. In this chapter, we adopt the natural and commonly used optimal single-price payment, as in [72, 74, 75], as our optimal payment benchmark because it is within a constant factor of the payment of any mechanism with price differentiation, as proved in [75]. In this chapter, therefore, we aim to design a single-price mechanism that pays every winner in .S according to the same price p. To simplify our analysis, we assume that the possible values of the cost .ci,k for a user .wi to execute a bundle of tasks .i,k ∈ Ti forms a finite set .C. The smallest and largest element in .C is .cmin and .cmax , respectively. Given the winner set .S, for an individual rational single-price mechanism, the platform’s total payment is minimized if and only if the price p equals the largest cost of the users in .S, that is, .p = maxwi ∈S ci,k . This is because otherwise the platform can always let .p = maxwi ∈S ci,k and obtain a smaller total payment while maintaining individual rationality. Therefore, the set .P containing all possible prices should satisfy that .P ⊆ C. Furthermore, we define that a price p is feasible if and only if it is possible to select a set of winners .S among the users with bidding prices .ρi ≤ p such that the error bound constraint defined in Eq. (13) is satisfied for every task. Then, we define the price set .P as the set containing all values in the set .C that are feasible. Thus, obviously we have .cmax ∈ P ⊆ C. Next, we formulate the total payment minimization (TPM) problem as the following optimization program. TPM Problem:  . min pxi . (18) i:wi ∈N

s.t.



 (2θi,j − 1)2 xi ≥ 2 ln

i:wi ∈N,τj ∈i

p − ρi xi ≥ 0, xi ∈ {0, 1}, ∀wi ∈ N, p ∈ P.

1 , δj

∀τj ∈ T.

(19)

∀wi ∈ N.

(20) (21)

Constants The TPM problem takes as inputs the price set .P, users’ bid profile .b, the reliability level matrix .θ, the vector .δ = (δ1 , · · · , δK ), as well as the task and user set .T and .N. Variables In the TPM problem, we have a vector of N binary variables .x = (x1 , · · · , xN ). For every user .wi ∈ N, there is a binary variable .xi indicating whether this user is selected as a winner. That is,  1, if wi ∈ S .xi = 0, otherwise.

Quality-Aware Incentive Mechanism for Mobile Crowdsourcing

109

Furthermore, another variable in the TPM problem is the price p, which could take any value in the price set .P.  Objective Function Based on the definition of variables .x and p, . i:wi ∈N pxi represents the platform’s total payment to all the winners. Hence, the TPM problem aims to find the vector .x and price p that minimize the platform’s total payment. Constraints Constraint (19) is exactly the error bound constraint represented by Inequality (13) in Lemma 2, which ensures that the aggregation error of every task .τj ∈ T is no larger than the predefined threshold .δj . To simplify presentation, we

 introduce the following notations: .qi,j = (2θi,j − 1)2 , .Qj = 2 ln δ1j , .q = [qi,j ] ∈ [0, 1]N ×K , and .Q = (Q1 , · · · , QK ), and thus, Constraint (19) can be simplified as the following Inequality (22):  .

qi,j xi ≥ Qj ,

∀τj ∈ T.

i:wi ∈N,τj ∈i

(22)

Furthermore, Constraint (20) ensures that for each user .wi , we have .p ≥ ρi , if the user is a winner. This means that any feasible solution to the TPM problem satisfies individual rationality, if users submit truthful bids. Apart from Constraints (19) and (20), we also consider two other inherent constraints, namely approximate truthfulness and differential privacy for users’ bids, which means that the hSRC auction that corresponds to any feasible solution is approximately truthful and differentially private. Due to the difficulty in mathematically formulating the two constraints, we take them into consideration without explicitly formulating them, in the TPM problem. Next, in Theorem 2, we prove the NP-hardness of the TPM problem. Theorem 2 The TPM problem is NP-hard. Proof Firstly, we transform the TPM problem into a modified TPM problem by fixing the price .p = 1 and relaxing Constraint (20), as well as the inherent approximate truthfulness and differential privacy constraints. Clearly, the modified TPM problem is a special case of the TPM problem. Thus, we turn to proving the NP-hardness of the modified TPM problem, instead. We start our proof by introducing an instance of the minimum set cover (MSC) problem with a universe of K elements .U = {τ1 , · · · , τK } and a set of N sets .H = {1 , · · · , N }. The objective of the MSC problem is to find the minimumcardinality subset of .H whose union contains all the elements in .U. We construct an instance of the modified TPM problem based on this instance of the MSC problem. Firstly, we construct .i from .i where every .τj ∈ i has .hi,j ∈ Z+ copies in .i . Furthermore, we require that the selected sets cover every .τj ∈ U for at least .Hj times. Therefore, we get an instance of the modified TPM problem where .q = [hi,j ] ∈ (Z+ )N ×K , .Q = (H1 , · · · , HK ), and the bidding bundle profile . = ( , · · · ,  ). In fact, the modified TPM problem represents a richer family N 1 of problems where elements in .q and .Q can be positive real values. Therefore, every

110

H. Jin and L. Su

instance of the NP-complete MSC problem is polynomial-time reducible to the modified TPM problem. The modified TPM problem, and thus, the original TPM problem, is NP-hard.



4.6 Mechanism Design Because of the NP-hardness of the TPM problem shown in Theorem 2, even given the price p, it is impossible to calculate in polynomial time the set of winners that minimize the platform’s total payment unless .P = NP. Let alone we eventually need to select an optimal price from the price set .P. Therefore, we aim to design a polynomial-time mechanism that gives us an approximately optimal total payment with a guaranteed approximation ratio to the optimal total payment .ROPT . In addition, we also take into consideration the bid privacypreserving objective when designing the mechanism. We present our mechanism in Algorithm 3, namely differentially private hSRC (DP-hSRC) auction, that satisfies all our design objectives. Algorithm 3 takes as inputs the privacy budget . , the cost upper bound .cmax , the user set .N, the task set .T, the price set .P, users’ bid profile .b, the .q matrix, and the .Q vector. It outputs the winner set .S and the payment p paid to each winner. Firstly, it sorts users according to the ascending order of their bidding prices such that .ρ1 ≤ ρ2 ≤ · · · ≤ ρN (line 1). Then, it initializes several parameters (lines 2– 5). It finds the minimum price .pmin in .P (line 2) and the index .imin of the largest bidding price that does not exceed .pmin (line 3). The algorithm constructs an index set .I containing all the integers from .imin to N (line 4). Set .I contains every user index i such that a winner set .Si is calculated among the users with bidding prices that are not larger than .ρi . In the last step of the initialization, the algorithm creates an extra bidding price .ρN +1 by adding a small positive constant .δ to .cmax (line 5) to ensure that .ρN +1 is greater than .∀p ∈ P. The purpose of creating .ρN +1 is to make sure that every price .p ∈ P is considered by lines 14 and 15 in the main loop (lines 6–15) for exactly once. After the initialization phase, Algorithm 3 calculates the winner set for every possible price .p ∈ P (lines 6–15). Intuitively, we need to calculate the winner set for every given price .p ∈ P. However, for all possible prices between two consecutive bidding prices, that is, .∀p ∈ P ∩ [ρi , ρi+1 ), the winner sets are the same. Therefore, to reduce the computational complexity and remove its dependency on the number of possible prices (i.e., .|P|), we only need to calculate the winner set for every price .p ∈ {ρimin , ρimin +1 , · · · , ρN }. At the beginning of every iteration of the main loop (line 6–15), Algorithm 3 initializes the winner set .Si as .∅, the residual .Q vector as .Q, and the candidate winner set .N as every user .wk with bidding price .ρk that is not larger than .ρi (line 7). The inner loop (lines 8–13) is executed until the error bound constraints for all tasks are satisfied or equivalently until .Q = 0K×1 . In every iteration of the inner loop (lines 8–13), the user .wimax that provides the most improvement to the feasibility of Constraint 19 is selected as the new winner (line 9).

Quality-Aware Incentive Mechanism for Mobile Crowdsourcing

111

Algorithm 3: DP-hSRC Auction

12 13

Input: , cmax , b, q, Q, N, T, P; Output: S, p; sort users according to the ascending order of bidding prices such that ρ1 ≤ ρ2 ≤ · · · ≤ ρN ; // Initialization pmin ← minp∈P p; imin ← arg maxi:ρi ≤pmin ρi ; I ← {imin , imin + 1, · · · , N }; // Add a small constant δ > 0 to cmax ρN +1 ← cmax + δ; // Calculates the winner sets foreach i ∈ I do Si ←  ∅, Q ← Q, N ← {wk |ρk ≤ ρi }; while j :τj ∈T Q j = 0 do  imax = arg maxi:wi ∈N j :τj ∈i min{Q j , qi,j }; Si ← Si ∪ {wimax }; N ← N \ {wimax }; // Update the residual Q vector foreach j s.t. τj ∈ T do Q j ← Q j − min{Q j , qimax ,j };

14 15

// Assign the same winner set Si to every possible price in [ρi , ρi+1 ) foreach p ∈ P ∩ [ρi , ρi+1 ) do S(p) ← Si ;

1 2 3 4 5 6 7 8 9 10 11

16 randomly pick a price p according to the distribution   exp − x|S(x)| 2N cmax  , ∀x ∈ P;  Pr[p = x] =  y|S(y)| y∈P exp − 2N cmax // Obtain the corresponding winner set 17 S ← S(p); 18 return {S, p};

Hence, .wimax is included in .Si (line 10) and excluded from .N (line 11). After .wimax is selected, the algorithm updates the residual .Q vector (lines 12 and 13). To ensure differential privacy, we introduce randomization to the output price. We extend the exponential mechanism proposed in [39] and set the probability that the output price p ofAlgorithm 3 equals a price .x ∈ P to be proportional to the value .exp − x|S(x)| 2N cmax . That is, 

x|S(x)| . Pr[p = x] ∝ exp − , ∀x ∈ P. 2Ncmax

(23)

112

H. Jin and L. Su

One important rationale of setting the probability of every possible price as the form in Eq. (23) is that the price resulting in a smaller total payment will have a larger probability to be sampled. In fact, the probability increases exponentially with the decrease of the total payment and the distribution is substantially biased toward low total payment prices. Therefore, we can both achieve differential privacy anda guaranteed  approximation to the optimal payment. Algorithm 3 normalizes x|S(x)| .exp − 2N cmax and randomly picks a price p according to the following distribution (line 16) defined in Eq. (24):   exp − x|S(x)| 2N cmax  , ∀x ∈ P.  . Pr[p = x] =  y|S(y)| exp − y∈P 2N cmax

(24)

After a price p is sampled, the winner set .S is set to be the one corresponding to p, namely .S(p) (line 17). Finally, it returns the winner set .S and the price p (line 18).

5 Conclusion In this chapter, we design two quality-aware incentive mechanisms for MCS systems. We design the first quality-aware incentive mechanism incorporating a crucial metric QoI. For the SRC auction, we design a truthful, individual rational, and computationally efficient mechanism that approximately maximizes the social welfare with a guaranteed approximation ratio. For the MRC auction, we design an iterative descending mechanism that achieves close-to-optimal social welfare with a guaranteed approximation ratio while satisfying individual rationality and computational efficiency. We design the second quality-aware incentive mechanism which incentivizes user participation without disclosing their sensitive bid information. The proposed mechanism is based on a novel design of single-minded reverse combinatorial auction with heterogeneous cost and thus bears several advantageous properties including approximate truthfulness, individual rationality, and computational efficiency.

References 1. X. Sheng, J. Tang, X. Xiao, G. Xue, Sensing as a service: challenges, solutions and future directions. IEEE Sens. J. 13(10), 3733–3741 (2013) 2. MedWatcher App. http://www.fda.gov/MedicalDevices/Safety/ReportaProblem/ucm385880. htm

Quality-Aware Incentive Mechanism for Mobile Crowdsourcing

113

3. Y. Cheng, X. Li, Z. Li, S. Jiang, Y. Li, J. Jia, X. Jiang, AirCloud: a cloud-based air-quality monitoring system for everyone, in Proceedings of the 12th ACM Conference on Embedded Network Sensor Systems (2014), pp. 251–265 4. D. Yang, G. Xue, X. Fang, J. Tang, Crowdsourcing to Smartphones: Incentive Mechanism Design for Mobile Phone Sensing, in Proceedings of the 18th Annual International Conference on Mobile Computing and Networking (2012), pp. 173–184 5. L. Duan, T. Kubo, K. Sugiyama, J. Huang, T. Hasegawa, J. Walrand, Reputation-based incentive protocols in crowdsourcing applications, in Proceedings of the IEEE INFOCOM (2012), pp. 2140–2148 6. B. Faltings, J. Li, R. Jurca, Incentive mechanisms for community sensing. IEEE Trans. Comput. 66(1), 115–128 (2014) 7. Z. Feng, Y. Zhu, Q. Zhang, L. Ni, A. Vasilakos, TRAC: Truthful auction for location-aware collaborative sensing in mobile crowdsourcing, in IEEE INFOCOM 2014-IEEE Conference on Computer Communications (2014), pp. 1231–1239 8. J.-S. Lee, B. Hoh. Sell your experiences: a market mechanism based incentive for participatory sensing, in Proceedings of the 2010 IEEE International Conference on Pervasive Computing and Communications (PerCom) (2010), pp. 60–68 9. Y. Zhang, M. van der Schaar, Reputation-based incentive protocols in crowdsourcing applications, in Proceedings of the IEEE INFOCOM (2012), pp. 2140–2148 10. H. Xie, J.C. Lui, Modeling crowdsourcing systems: design and analysis of incentive mechanism and rating system. SIGMETRICS Perform. Eval. Rev. 42(2), 52–54 (2014) 11. H. Xie, J.C. Lui, J.W. Jiang, W. Chen, Incentive mechanism and protocol design for crowdsourcing systems, in Proceedings of the 2014 52nd Annual Allerton Conference on Communication, Control, and Computing (Allerton) (2014), pp. 140–147 12. Z. Feng, Y. Zhu, Q. Zhang, H. Zhu, J. Yu, J. Cao, L. Ni, Towards truthful mechanisms for mobile crowdsourcing with dynamic smartphones, in 2014 IEEE 34th International Conference on Distributed Computing Systems (2014), pp. 11–20 13. X. Zhang, Z. Yang, Z. Zhou, H. Cai, L. Chen, X. Li, Free market of crowdsourcing: incentive mechanism design for mobile sensing. IEEE Trans. Parallel Distrib. Syst. 25(12), 3190–3200 (2014) 14. D. Zhao, X.-Y. Li, H. Ma, How to crowdsource tasks truthfully without sacrificing utility: online incentive mechanisms with budget constraint, in IEEE INFOCOM 2014-IEEE Conference on Computer Communications (INFOCOM) (2014), pp. 1213–1221 15. S. He, D.-H. Shin, J. Zhang, J. Chen, Toward optimal allocation of location dependent tasks in crowdsensing, in IEEE INFOCOM 2014-IEEE Conference on Computer Communications (INFOCOM) (2014), pp. 745–753 16. I. Koutsopoulos, Optimal incentive-driven design of participatory sensing systems, in IEEE INFOCOM 2014-IEEE Conference on Computer Communications (INFOCOM) (2013), pp. 1402–1410 17. T. Luo, H.-P. Tan, L. Xia, Optimal incentive-driven design of participatory sensing systems, in IEEE INFOCOM 2014-IEEE Conference on Computer Communications (INFOCOM) (2014), pp. 1402–1410 18. Z. Song, C. Liu, J. Wu, J. Ma, W. Wang, QoI-Aware multitask-oriented dynamic participant selection with budget constraints. IEEE Trans. Veh. Technol. 63(9), 4618–4632 (2014) 19. J. Sun, H. Ma, Privacy-preserving verifiable incentive mechanism for online crowdsourcing markets, in Proceedings of the 2014 23rd International Conference on Computer Communication and Networks (ICCCN) (2014), pp. 1–8 20. L. Gao, F. Hou, J. Huang, Optimal incentive-driven design of participatory sensing systems, in IEEE INFOCOM 2014-IEEE Conference on Computer Communications (INFOCOM) (2015), pp. 2803–2811 21. Y. Wen, J. Shi, Q. Zhang, X. Tian, Z. Huang, H. Yu, Y. Cheng, X. Shen, Quality-driven auctionbased incentive mechanism for mobile crowd sensing. IEEE Trans. Veh. Technol. 64(9), 4203– 4214 (2015)

114

H. Jin and L. Su

22. X. Zhang, G. Xue, R. Yu, D. Yang, J. Tang, Truthful incentive mechanisms for crowdsourcing, in IEEE INFOCOM 2014-IEEE Conference on Computer Communications (INFOCOM) (2015), pp. 2830–2838 23. Y. Wei, Y. Zhu, H. Zhu, Q. Zhang, G. Xue, Truthful online double auctions for dynamic mobile crowdsourcing, in IEEE INFOCOM 2014-IEEE Conference on Computer Communications (INFOCOM) (2015), pp. 2074–2082 24. H. Zhang, B. Liu, H. Susanto, G. Xue, T. Sun, Incentive mechanism for proximity-based Mobile Crowd Service systems, in IEEE INFOCOM 2014-IEEE Conference on Computer Communications (INFOCOM) (2016), pp. 1–9 25. J. Wang, J. Tang, D. Yang, E. Wang, G. Xue, Quality-aware and fine-grained incentive mechanisms for mobile crowdsensing, in Proceedings of the2016 IEEE 36th International Conference on Distributed Computing Systems (ICDCS) (2016), pp. 354–363 26. D. Yang, G. Xue, X. Fang, J. Tang, Incentive mechanisms for crowdsensing: crowdsourcing with smartphones. IEEE ACM Trans. Netw. 24(3), 1732–1744 (2016) 27. D. Zhao, X.-Y. Li, H. Ma, Budget-feasible online incentive mechanisms for crowdsourcing tasks truthfully. IEEE ACM Trans. Netw. 24(2), 647–661 (2016) 28. H. Jin, L. Su, H. Xiao, K. Nahrstedt, INCEPTION: incentivizing privacy-preserving data aggregation for mobile crowd sensing systems, in Proceedings of the Twelfth ACM International Symposium on Mobile Ad Hoc Networking and Computing (2016), pp. 341–350 29. J. Lin, D. Yang, M. Li, J. Xu, G. Xue, BidGuard: a framework for privacy-preserving crowdsensing incentive mechanisms. In: Proceedings of the 2016 IEEE conference on communications and network security (CNS) (2016), pp. 145–153 30. Z. Zheng, F. Wu, X. Gao, H. Zhu, G. Chen, S. Tang, A budget feasible incentive mechanism for weighted coverage maximization in mobile crowdsensing. IEEE Trans. Mob. Comput. 16(9), 2392–2407 (2016) 31. D. Zhao, H. Ma, L. Liu, Frugal online incentive mechanisms for mobile crowd sensing. IEEE Trans. Veh. Technol. 66(4), 3319–3330 (2016) 32. T. Luo, S.K. Das, H.P. Tan, L. Xia, Incentive mechanism design for crowdsourcing: an all-pay auction approach. ACM Trans. Intell. Syst. Technol. 7(3), 1–26 (2016) 33. H. Jin, L. Su, H. Xiao, K. Nahrstedt, CENTURION: incentivizing multi-requester mobile crowd sensing, in IEEE INFOCOM 2014-IEEE Conference on Computer Communications (INFOCOM) (2017), pp. 1–9 34. X. Gan, X. Wang, W. Niu, G. Hang, X. Tian, X. Wang, J. Xu, Incentivize multi-class crowd labeling under budget constraint. IEEE J. Sel. Areas Commun. 35(4), 893–905 (2017) 35. MyHeartMap. http://www.med.upenn.edu/myheartmap/ 36. J. Eriksson, L. Girod, B. Hull, R. Newton, S. Madden, H. Balakrishnan, The pothole patrol: using a mobile sensor network for road surface monitoring, in Proceedings of the 6th international conference on Mobile systems, applications, and services (2008), pp. 29–39 37. P. Mohan, V.N. Padmanabhan, R. Ramjee, Nericell: rich monitoring of road and traffic conditions using mobile smartphones, in Proceedings of the 12th ACM Conference on Embedded Network Sensor Systems (2008), pp. 323–336 38. C. Dwork, F. McSherry, K. Nissim, A. Smith, Calibrating noise to sensitivity in private data analysis, in Proceedings of the Theory of Cryptography: Third Theory of Cryptography Conference, TCC 2006, New York, NY, USA, March 4–7, 2006 (Springer, Berlin, 2006), pp. 265–284 39. F. McSherry, K. Talwar, Mechanism design via differential privacy, in Proceeding of the 48th Annual IEEE Symposium on Foundations of Computer Science (FOCS’07) (2007), pp. 94–103 40. J. Huang, R.A. Berry, M.L. Honig, Auction-based spectrum sharing. Mob. Netw. Appl. 11(3), 405–418 (2006) 41. X. Chen, J. Huang, Spatial spectrum access game: nash equilibria and distributed learning, in Proceedings of the Twelfth ACM International Symposium on Mobile Ad Hoc Networking and Computing (2012), pp. 205–214 42. J. Jia, Q. Zhang, Q. Zhang, M. Liu, Revenue generation for truthful spectrum auction in dynamic spectrum access, in Proceedings of the Twelfth ACM International Symposium on Mobile Ad Hoc Networking and Computing (2009), pp. 3–12

Quality-Aware Incentive Mechanism for Mobile Crowdsourcing

115

43. H. Jin, G. Sun, X. Wang, Q. Zhang, Spectrum trading with insurance in cognitive radio networks, in IEEE INFOCOM 2014-IEEE Conference on Computer Communications (INFOCOM) (2012), pp. 2041–2049 44. M. Dong, G. Sun, X. Wang, Q. Zhang, Combinatorial auction with time-frequency flexibility in cognitive radio networks, in IEEE INFOCOM 2014-IEEE Conference on Computer Communications (INFOCOM) (2012), pp. 2282–2290 45. D. Yang, X. Fang, G. Xue, Truthful auction for cooperative communications, in Proceedings of the Twelfth ACM International Symposium on Mobile Ad Hoc Networking and Computing (2011) 46. X. Chen, B. Proulx, X. Gong, J. Zhang, Social trust and social reciprocity based cooperative D2D communications, in Proceedings of the Twelfth ACM International Symposium on Mobile Ad Hoc Networking and Computing (2013), pp. 187–196 47. Q. Yu, J. Chen, Y. Fan, X. Shen, Y. Sun, Multi-channel assignment in wireless sensor networks: a game theoretic approach, in IEEE INFOCOM 2014-IEEE Conference on Computer Communications (INFOCOM) (2010), pp. 1–9 48. M. Dong, H. Jin, G. Sun, X. Wang, W. Liu, X. Wang, Non-cooperative game based social welfare maximizing bandwidth allocation in WSN, in 2011 IEEE Global Telecommunications Conference-GLOBECOM 2011 (2011), pp. 1–5 49. Q. Zhang, Y. Wen, X. Tian, X. Gan, X. Wang, Incentivize crowd labeling under budget constraint, in Proceedings of the 2015 IEEE Conference on Computer Communications (INFOCOM) (2015), pp. 2812–2820 50. X. Zhang, L. Liang, C. Luo, L. Cheng, Privacy-preserving incentive mechanisms for mobile crowdsensing. IEEE Pervasive Comput. 17(3), 47–57 (2018) 51. Z. Zhang, S. He, J. Chen, J. Zhang, REAP: an efficient incentive mechanism for reconciling aggregation accuracy and individual privacy in crowdsensing. IEEE Trans. Inf. Forensics Secur. 13(12), 2995–3007 (2018) 52. B. Zhao, S. Tang, X. Liu, X. Zhang, PACE: privacy-preserving and quality-aware incentive mechanism for mobile crowdsensing. IEEE Trans. Mob. Comput. 20(5), 1924–1939 (2021) 53. Y. Liu, H. Wang, M. Peng, J. Guan, Y. Wang, An incentive mechanism for privacy-preserving crowdsensing via deep reinforcement learning. IEEE Internet Things J. 8(10), 8616–8631 (2021) 54. Q. Li, Y. Li, J. Gao, B. Zhao, W. Fan, J. Han, Resolving conflicts in heterogeneous data by truth discovery and source reliability estimation, in Proceedings of the 2014 ACM SIGMOD international conference on Management of data (2014), pp. 1187–1198 55. L. Su, Q. Li, S. Hu, S. Wang, J. Gao, H. Liu, T. Abdelzaher, J. Han, X. Liu, Y. Gao, L. Kaplan, Generalized decision aggregation in distributed sensing systems, in Proceedings of the 2014 IEEE Real-Time Systems Symposium (2014), pp. 1–10 56. Q. Li, Y. Li, J. Gao, L. Su, B. Zhao, M. Demirbas, W. Fan, J. Han, A confidence-aware approach for truth discovery on long-tail data. Proc. VLDB Endow. 8(4), 425–436 (2014) 57. S. Wang, D. Wang, L. Su, L. Kaplan, T.F. Abdelzaher, Towards cyber physical systems in social spaces: the data reliability challenge, in Proceedings of the2014 IEEE Real-Time Systems Symposium (2014), pp. 74–85 58. H. Li, B. Zhao, A. Fuxman, The wisdom of minority: discovering and targeting the right group of workers for crowdsourcing, in Proceedings of the 23rd International Conference on World Wide Web (2014), pp. 165–176 59. L. Blumrosen, N. Nisan, Combinatorial auctions, in Algorithmic Game Theory (Cambridge University Press, London, 2007) 60. M. Babaioff, R. Lavi, E. Pavlov, Single-value combinatorial auctions and algorithmic implementation in undominated strategies. J. ACM 56(1), 1–32 (2009) 61. E.H. Clarke, Multipart pricing of public goods, in Public Choice (Springer, Netherlands, 1971) 62. T. Groves, Incentives in teams. Econometrica: Journal of the Econometric Society (JSTOR) 41(4), 617–631 (1973) 63. M.J. Osborne, A. Rubinstein, A course in game theory, Electronic edn. (The MIT Press, Cambridge, 1994)

116

H. Jin and L. Su

64. M. Babaioff, R. Lavi, E. Pavlov, Mechanism design for single-value domains, in Proceedings of the ACM AAAI (2005), pp. 241–247 65. Z. Zheng, F. Wu, S. Tang, G. Chen, Unknown combinatorial auction mechanisms for heterogeneous spectrum redistribution, in Proceedings of the 15th ACM International Symposium on Mobile ad HOC Networking and Computing (2014), pp.3–12 66. R.B. Myerson, Optimal auction design, in Mathematics of Operations Research (INFORMS) (1981) 67. A. Thiagarajan, L. Ravindranath, K. LaCurts, S. Madden, H. Balakrishnan, S. Toledo, J. Eriksson, VTrack: accurate, energy-aware road traffic delay estimation using mobile phones, in Proceedings of the 7th ACM Conference on Embedded Networked Sensor Systems (2009), pp. 85–98 68. D. Oleson, A. Sorokin, G. Laughlin, V. Hester, J. Le, L. Biewald, Programmatic gold: targeted and scalable quality assurance in crowdsourcing, in AAAI Press HCOMP (2011), pp. 43–48 69. C. Meng, W. Jiang, Y. Li, J. Gao, L. Su, H. Ding, Y. Cheng, Truth discovery on crowd sensing of correlated entities, in Proceedings of the 12th ACM Conference on Embedded Network Sensor Systems (2015), pp. 169–182 70. Y. Li, J. Gao, C. Meng, Q. Li, L. Su, B. Zhao, W. Fan, J. Han, A survey on truth discovery. SIGKDD Explor. Newsl. 17(2), 1–16 (2016) 71. A. Gupta, K. Ligett, F. McSherry, A. Roth, K. Talwar, Differentially private combinatorial optimization, in Proceedings of the twenty-first annual ACM-SIAM symposium on Discrete Algorithms. Society for Industrial and Applied Mathematics (2010), pp. 1106–1125 72. R. Zhu, Z. Li, F. Wu, K. Shin, G. Chen, Differentially private spectrum auction with approximate revenue maximization, in Proceedings of the Twelfth ACM International Symposium on Mobile Ad Hoc Networking and Computing (2014), pp. 185–194 73. R. Zhu, K.G. Shin, Differentially private and strategy-proof spectrum auction with approximate revenue maximization, in IEEE INFOCOM 2014-IEEE Conference on Computer Communications (INFOCOM) (2015), pp. 918–926 74. A. Gopinathan, Z. Li, A prior-free revenue maximizing auction for secondary spectrum access, in IEEE INFOCOM 2014-IEEE Conference on Computer Communications (INFOCOM) (2011), pp. 86–90 75. N. Immorlica, A.R. Karlin, M. Mahdian, K. Talwar, Balloon popping with applications to ascending auctions, in Proceedings of the 48th Annual IEEE Symposium on Foundations of Computer Science (FOCS’07) (2007), pp. 104–112

Incentive Mechanism Design for Mobile Crowdsourcing Without Verification Chao Huang, Haoran Yu, Jianwei Huang, and Randall Berry

1 Introduction 1.1 Motivations The rapid growth of the Internet has redefined the wisdom of the crowd by resorting to distributed crowdsourced workers for solving various online tasks. A mobile crowdsourcing platform can solve complex problems via recruiting nonexpert workers to complete decomposed simple tasks and appropriately aggregating the workers’ solutions. Mobile crowdsourcing has become prevalent in practical applications, including but not limited to: • Voting Systems: A voting task requires a crowdsourced worker to select an answer from multiple choices. For example, in Amazon Mechanical Turk, workers assigned with image labeling tasks are asked to determine whether a photo contains a certain type of animal. Then, the annotated results are utilized to train machine learning models [1]. In the blockchain-enabled platform Steem, people worldwide use their mobile phones to express their individual perspectives

C. Huang University of California, Davis, CA, USA H. Yu Beijing Institute of Technology, Beijing, China J. Huang The Chinese University of Hong Kong, Shenzhen, China e-mail: [email protected] R. Berry () Northwestern University, Evanston, IL, USA e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 J. Wu, E. Wang (eds.), Mobile Crowdsourcing, Wireless Networks, https://doi.org/10.1007/978-3-031-32397-3_5

117

118

C. Huang et al.

regarding online content through voting. This helps boost the efficiency of content moderation as well as promote the creation of high-quality content [2]. • Information Sharing Systems: Many mobile crowdsourcing platforms aim to share various types of information among the crowd to increase user experience. For example, in the mobile Crowdsensing App CUPUS, users utilize builtin sensors in their phones (e.g., microphones and thermometers) to sense environmental information (e.g., weather conditions and temperature levels) and share the sensed information with the platform and other users [3]. This greatly reduces the platform’s costs on the direct deployment of centralized managed sensors as well as enhancing user experience via real-time data sharing [4]. • Creative Systems: Even highly advanced technologies cannot replace humans in terms of creativity. Only humans can accomplish creative tasks such as coding and brainstorming. In IdeaConnection, a popular crowdsourcing website, users are encouraged to submit innovative solutions to technical challenges such as accurate measurement of oceanographic data [5]. The creation of innovative approaches has a huge potential in advancing technologies and reshaping the state-of-the-art solutions, which could facilitate improving a person’s quality of life. As the crowdsourced workers make decentralized decisions and are likely to be strategic, it is important for the mobile crowdsourcing platform to design effective incentive mechanisms to motivate high-quality solutions from the workers. A highquality solution requires a worker to exert effort, and the mobile crowdsourcing platform needs to properly motivate effort exertion and also truthful reporting of the solution. Based on the possibility of verifying workers’ reported solutions, the existing studies on the design of incentive mechanisms for mobile crowdsourcing can be mainly divided into the following two types: • Verifiable Information: The platform can access the ground truth later to verify the quality of the workers’ reported solutions. For example, for a weather prediction task, the platform can calculate the incentive amount to the workers after the weather information becomes available [6]. • Unverifiable Information: It is not possible for the mobile crowdsourcing platform to access the ground truth in order to verify the quality of the workers’ solutions. This can be because the ground truth is very costly to obtain. For example, in environmental sensing, it is prohibitive to judge the accuracy of each piece of environmental information provided by distributed mobile sensors, as it requires the deployment of centralized managed sensors to validate all the reported information. This chapter focuses on the much more challenging scenario where the ground truth information is difficult (or even impossible) to obtain. In particular, we survey efficient incentive mechanisms to eliciting high-quality and truthful solutions from mobile crowdsourced workers when the ground truth information is unverifiable.

Incentive Mechanism Design for Mobile Crowdsourcing Without Verification

119

1.2 Key Challenges Despite the existing research works on mobile crowdsourcing without verification, there are a few practically important and technically challenging problems that are under-explored. We list some of them as follows: • Worker Heterogeneity: In practical mobile crowdsourcing applications, workers are likely to differ in various aspects, such as innate work skill levels, computational resources, energy constraints, sensing capabilities, mobility patterns, etc. That is, different workers are prone to generate data with different accuracy. While much existing work considers homogeneous workers, it is important to formulate a general analytical framework to study the impact of worker heterogeneity and incorporate such an impact into the incentive mechanism design. • Worker Collusion: Most existing incentive mechanisms (e.g., the rich strand of peer prediction mechanisms) calculate the rewards to workers based on the comparison among workers’ reported solutions. This can be vulnerable to worker collusion, i.e., the workers can coordinate to misreport their solutions. For example, in a mechanism where workers are rewarded based on the consistency of their reported solutions, if the workers can manage to communicate with each other, the best strategy for workers is to collude and report the same solutions. This would be disastrous for the mobile crowdsourcing platform due to paying high costs for useless solutions. In fact, without access to the ground truth, it is practically significant and technically challenging to design incentive mechanisms that help mitigate worker collusion, prevent spam solutions, and hence improve the crowdsourcing system performance. • Information Incompleteness: In some cases, it is difficult for a platform to access and make use of workers’ personal data due to an increasing tendency of privacy protection. Without the personal data distribution, the platform can hardly estimate the workers’ private information (e.g., sensing capabilities), making it hard to design an effective incentive mechanism. Hence, it is of paramount importance to understand how to design an effective mechanism using online learning techniques to overcome the information incompleteness and learn the workers’ sensing capabilities (and privacy costs) accurately and promptly over time. • Information Asymmetry: Usually, the mobile crowdsourcing platform has more information than workers regarding worker characteristics through market research and past experiences. For example, for Waze and Amazon Mechanical Turk, each worker’s historical performance is known by the platform but not by the other workers [1]. The information asymmetry between the platform and the workers is practically important and under-explored. It calls for research to study the impact of information asymmetry on the incentive mechanism design and further investigate how the crowdsourcing platform can take advantage of such asymmetry via strategically revealing such information to the workers.

120

C. Huang et al.

1.3 Chapter Outline The remainder of the chapter is organized as follows. In Sect. 2, we present a generic model for mobile crowdsourcing systems without ground truth verification. In Sects. 3 and 4, we discuss promising approaches and techniques to addressing worker heterogeneity and worker collusion. In Sects. 5 and 6, we discuss potential solutions to information incompleteness and information asymmetry. We present a few open problems and conclude in Sect. 7.

2 Model In Sect. 2.1, we introduce the workers’ possible decisions and payoffs. In Sect. 2.2, we introduce the platform’s potential decisions and payoff. In Sect. 2.3, we discuss the interactions between the workers and the platform.

2.1 Workers’ Decisions and Payoffs In this subsection, we first introduce the task and workers and then define each worker’s strategies and payoff function.

2.1.1

Task and Workers

Consider a mobile crowdsourcing platform that aims to obtain solutions to a task via a set .N = {1, 2, · · · , N} of workers. For the ease of presentation, we consider binary-solution tasks throughout the entire chapter. In practice, many mobile crowdsourcing applications focus on binary tasks, e.g., judging whether a certain spot has a traffic jam, which draw extensive attention in the literature [7, 8]. Let .X = {1, 0} denote the task’s solution space, where 1 means Yes (e.g., the traffic jam occurs) and 0 means No (e.g., the traffic jam does not occur). We further define: • .x ∈ X: the true solution (i.e., ground truth), which neither the platform nor the workers know. • .xiest ∈ X: worker i’s estimated solution after completing the task. The estimated solution .xiest may be inaccurate and hence different from the ground truth x. rep • .xi ∈ X: worker i’s reported solution to the platform. Due to strategic rep considerations, the reported solution .xi may or may not be the same as the est worker’s estimated solution .xi .

Incentive Mechanism Design for Mobile Crowdsourcing Without Verification

2.1.2

121

Worker Effort Exertion Strategy

Each worker can decide whether to exert effort completing the task, and the accuracy (i.e., quality) of his solution stochastically depends on his chosen effort level. Specifically, a worker can choose to either exert effort or not exert effort, and we use .ei ∈ {0, 1} to denote worker i’s effort level [7, 9]. If worker i does not exert effort, i.e., .ei = 0, he will generate the correct solution (which is the task’s true solution) with a probability .0.5 at zero cost. We assume that without exerting effort a worker has no information about the true solution, so the estimated solution is equally likely to be correct or wrong [7, 9]. We can extend our analysis to the scenario where even without any effort a worker still has some information about the true solution. In this scenario, the solution would always be more accurate than random guessing [8]. Exerting effort (i.e., .ei = 1) improves a worker’s solution accuracy at a cost .ci ≥ 0, and he can generate the correct solution with probability .pi ∈ (0.5, 1]. For example, in mobile traffic sensing, workers who travel to the designated spot with an effort exertion cost will generate more accurate sensing information than those who stay at home. More specifically,   est  0.5, . P rob xi =x = pi ,

2.1.3

if ei = 0 (with zero cost), if ei = 1 (with a cost ci ≥ 0).

(1)

Worker Solution Reporting Strategy

Each worker also needs to decide whether to truthfully report his solution to the platform. For a worker i who does not exert effort, he can only apply the random reporting strategy denoted by .ri = rd. The .rd strategy is adopted for ease of exposition. In fact, if a worker exerts no effort, his solution is equally likely to be correct or wrong. Hence, one could equivalently view that the worker is either truthfully reporting its solution or untruthfully reporting it. For those workers who exert effort, they can either truthfully or untruthfully report their solutions. We use .ri ∈ {1, −1} to denote the reporting strategy, with .ri = 1 indicating truthful reporting and .ri = −1 indicating untruthful reporting. More specifically,

rep .x i

=

⎧ est ⎪ ⎪ ⎨xi ,

1 − xiest , ⎪

if ri = 1,

if ri = −1, ⎪ ⎩1 or 0 with an equal probability, if r = rd. i

(2)

For ease of exposition, we use .si  (ei , ri ) to denote worker i’s effort exertion and reporting strategy, with .si ∈ Si  {(0, rd), (1, 1), (1, −1)}. In our model, if a worker exerts effort but untruthfully reports (i.e., uses .(1, −1)), the accuracy of his reported solution will be smaller than .0.5. We can further analyze the case where a

122

C. Huang et al.

worker has an accuracy smaller than .0.5 a priori. The worker can simply choose to reverse his solution and achieve an accuracy larger than .0.5.

2.1.4

Incentive Mechanism

After workers report their solutions, the platform will distribute rewards to workers according to some predefined incentive mechanisms. The reward can be money or some type of credits and points that workers can use within the platform. Let the mapping .F : x rep → R ⊂ R+ N denote a general incentive mechanism where .R = {Ri }i∈N . Each worker i may obtain a reward .Ri ∈ R+ for reporting the solution to the platform.

2.1.5

Worker Payoff

We define worker i’s expected payoff as .

ui (e, r) = Ri · Pi (e, r) − ci · ei .

(3)

In (3), .Pi (e, r) is the probability that worker i obtains a reward, and hence .Ri · Pi (e, r) captures the expected reward. The term .ci · ei captures the effort exertion cost.

2.2 Platform’s Decisions and Payoff 2.2.1

Platform Decisions

The platform needs to decide a proper incentive mechanism .F : x rep → R ⊂ R+ N to incentivize high-quality and truthful information from crowdsourced workers. As will be mentioned in Sect. 6, the platform may further design an information revelation strategy to affect workers’ beliefs and hence maximize the platform payoff.

2.2.2

Platform Payoff

The mobile crowdsourcing platform seeks to achieve a good tradeoff between the accuracy (i.e., quality) of the aggregated solution and the cost of incentivizing the workers. Specifically, we define the platform’s expected payoff as follows: .

U (R) = βP agr (R) − (1 − β)C tot (R),

(4)

Incentive Mechanism Design for Mobile Crowdsourcing Without Verification

Stage I Decide Incentive Mechanism

Platform

123

Stage II Decide Effort & Reporting Strategies

Workers

Fig. 1 Two-stage platform–worker interaction

where .P agr (R) captures the accuracy (i.e., quality) of the aggregated solution, tot (R) captures the total incentives, and .β ∈ [0, 1] is a coefficient balancing the .C two terms. A large .β implies that the platform attaches more importance to the quality of workers’ reported solution than the incentive cost.

2.3 Platform–Worker Interaction Canonically, the interactions between the platform and workers can be formulated as a two-stage Stackelberg game (illustrated in Fig. 1). The platform moves first and decides the incentive mechanism in Stage I. The workers are followers and behave according to the decided incentive mechanism. That is, the workers decide their effort and reporting strategies in Stage II. It is interesting to note the workers in Stage II also interact in a game-theoretical fashion. The intuition is that due to the lack of ground truth verification, the platform needs to design an incentive mechanism that rewards workers based on the correlation of their reported solutions. This gives rise to a greater flexibility of workers’ strategic manipulations and makes the mechanism design more challenging.

3 Approaches to Worker Heterogeneity In Sect. 3.1, we give motivating examples of worker heterogeneity and ask several key questions. In Sect. 3.2, we present the mechanism design. In Sect. 3.3, we present the key results and insights associated with the solution.

124

C. Huang et al.

3.1 Motivating Examples and Key Questions Extensive research works on mobile crowdsourcing without verification have emerged over the past decade. Much existing literature is restricted to the case of homogeneous workers who have the same solution accuracy (i.e., for each candidate solution, the probabilities that the workers generate the correct solution are the same). This is not realistic in many domains. For example, in mobile crowdsensing, different sensors may generate sensing data with different accuracy. In our work [7, 10], we consider a general model of workers with heterogeneous solution accuracy (i.e., different workers may generate a good solution with different probabilities). More specifically, we consider workers with heterogeneous solution accuracy (i.e., sensing capabilities). Specifically, there are k workers with a high solution accuracy .ph and the remaining workers have a low accuracy .pl , where .k ∈ {1, 2, · · · , N − 1} and .0.5 < pl < ph ≤ 1. Let .Nh and .Nl denote the set containing high-accuracy workers and low-accuracy workers, respectively. We further assume that the workers have the same cost for effort exertion, i.e., .ci = c ≥ 0, ∀i ∈ N. It is challenging to design a mechanism considering heterogeneous solution accuracy, as it requires the mechanism to account for the worker diversity, which makes the elicitation of truthfulness more difficult. Furthermore, the study of incentive mechanism design accounting for worker heterogeneity leads to several key research questions:

•

? Questions

• Question 3a: given an incentive mechanism, how will the heterogeneous workers behave? • Question 3b: how will the worker heterogeneity affect the platform’s tradeoff?

In the remainder of this section, we focus on addressing the above key questions.

3.2 Solution: Majority Voting Mechanism We adopt the majority voting mechanism, a popular practical method for mobile crowdsourcing without verification, to study the workers’ behaviors and the platform’s tradeoff problem. In this mechanism, a worker receives a reward if his solution is consistent with the majority solution from the other workers. More majority specifically, let .x−i denote the majority solution from worker i’s perspective, and it has the following form:

Incentive Mechanism Design for Mobile Crowdsourcing Without Verification

majority .x −i

⎧ ⎪ ⎪ ⎨1, = 0, ⎪ ⎪ ⎩tie,

if if if



report

>

report


Important

• Under the majority voting mechanism, the workers with higher solution accuracy are more likely to spend effort and truthfully provide high-quality solutions.

Incentive Mechanism Design for Mobile Crowdsourcing Without Verification

127

• The platform payoff may decrease in the number of high-accuracy workers. More high-accuracy workers bring marginally decreasing benefit to the platform, but the rewards used as incentives may significantly increase.

4 Approaches to Worker Collusion In Sect. 4.1, we give motivating examples of worker collusion and propose several key questions. In Sect. 4.2, we detail the mechanism design. In Sect. 4.3, we present the key results and insights.

4.1 Motivating Examples and Key Questions Due to the lack of ground truth, a key challenge is that workers may collude to misreport their solutions, since the quality and truthfulness of their reports are hard to verify [11]. Existing mechanisms in the literature restricted the gathered information to voluntary reports from the workers, which did not address this potential collusion issue. For example, in the majority voting mechanism (as in Sect. 3.2) where workers are rewarded based on the consistency of their solutions, workers may coordinate to misreport their solutions, even if all the solutions are incorrect [4]. The worker collusion challenge leads to the following two questions:

•

? Questions

• Question 4a: how to design an incentive mechanism that can mitigate the worker collusion? • Question 4b: how will the workers behave according to the incentive mechanism that mitigates collusion?

We focus on addressing the above questions in the rest of the section.

4.2 Solution: Truth Detection Mechanism A potential approach to mitigate collusion is to utilize a truth detection technology [12]. A truth detection technology is some type of interface which can give a

128

C. Huang et al.

signal indicating if a worker is being truthful or lying. In practice, the inference can be based on a number of factors including (but not limited to) physiological measures (e.g., pupil dilation and facial electromyography [12, 13]). Some recent works empirically validated the credibility of even using recorded non-real-time information (e.g., recorded via phone cameras), such as facial expressions and verbal cues, to infer truthfulness [14]. In 2018, several European countries initiated the IBORDERCTRL project, which utilizes smart truth detection systems to enhance border security [15]. Truth detection can be applied in various mobile crowdsourcing scenarios. For example, in crowdsensing, after the workers upload their sensing information on road conditions, the platform can require the workers to answer predefined, individualized questions via the cameras in their phones. The platform can also utilize geographical measures (e.g., GPS service) to help identify whether a worker has been to the task location. Doing this allows the platform to better judge the effort exertion behavior of the worker. In Amazon Mechanical Turk, workers are also required to pass a bot test via several simple image labeling questions before they start to finish online tasks. Such a process helps verify the correctness of workers’ responses to questions and hence helps reveal the quality of workers’ reported solutions. Since truth detection can help test the truthfulness of workers’ reporting, it has a great potential to assist the incentive design for mobile crowdsourcing without verification. Next, we provide a detailed incentive design based on truth detection. Each worker may obtain a truthful reward .Ri from the platform if selected for truth detection. Before reporting their estimated solutions, all workers are informed that they will be tested by the truth detector with a probability .prob ∈ [0, 1] (which is the platform’s decision). If worker i is chosen to interact with the truth detector (after he reports the estimated solution), he will be asked the following question: “Have you invested effort finishing the task?”1 Based on his answer and the truth detection output, he will receive a truthful reward .Ri as follows (see Fig. 3): • If worker i’s answer to the question is “Yes” (i.e., 1), – He will obtain a bonus .Ri = bon > 0 if the truth detector judges him being truthful, i.e., the truth detected result is 1. – He will obtain a penalty .Ri = −pen < 0 (i.e., pay the platform a positive amount pen) if the truth detector judges him lying, i.e., the truth detected result is 0.

1 Our question focuses on finding out whether a worker has exerted effort, which will affect the accuracy of his estimation. In a real-world implementation, a platform can utilize supplemented “more objective” questions to better judge the effort exertion behavior. For example, an article review platform can present multiple sentences in a list and require the workers to choose one that fits the gist of the article most. A correct sentence choice will likely suggest that the workers exert efforts reading the article.

Incentive Mechanism Design for Mobile Crowdsourcing Without Verification

Truth Detected Result (of Effort Exertion) Answer to the Question

Fig. 3 Reward matrix for truth detection. Here, we have .0 ≤ bon, pen ≤ upp, where .upp > 0 is a finite constant. The platform cannot provide an arbitrarily large bonus or penalty due to practical constraints

129

0

1

0

0

0

1

- pen

bon

• If worker i’s answer to the question is “No” (i.e., 0), he will obtain no bonus or penalty, i.e., .Ri = 0. Detection Accuracy Let .q ∈ [0.5, 1] denote the accuracy of the truth detection, i.e., it can correctly judge whether a worker is telling the truth with a probability q. Therefore, if worker i indeed invests effort and says “Yes” to the question, he will obtain the bonus bon with probability q and obtain the penalty .−pen with probability .1 − q. If worker i invests effort and says “No”, he will obtain zero reward that does not depend on the accuracy q.

4.3 Results and Insights To derive the incentive mechanism for the truth detection, we first need to ensure that the workers will truthfully answer the truth detection question. The following theorem characterizes the conditions of bonus and penalty parameters under which this can be achieved. Theorem 2 (Incentive Compatibility) A worker will truthfully answer the truth detection question if q(bon + pen) ≥ max{bon, pen}.

.

(6)

Proof Let .Ri (ei , wi ) denote worker i’s truthful reward, where .wi ∈ {0, 1} is his answer to the truth detection question, i.e., .wi = 1 corresponds to answering “Yes” to the question and .wi = 0 corresponds to “No.” Recall that the truth detection question is “have you invested effort finishing the task?” We discuss two cases when worker i exerts effort (i.e., .ei = 1) and does not exert effort (i.e., .ei = 0). • Case A: .ei = 1, i.e., worker i exerts effort. Answering “Yes” to the truth detection question (i.e., truthfully reporting since .ei = 1) yields

130

C. Huang et al.

E[Ri (1, 1)] = q · bon + (1 − q)(−pen) .

= q(bon + pen) − pen,

(7)

while answering “No” to the question (i.e., lying ) yields E[Ri (1, 0)] = 0.

.

(8)

To ensure truthfulness in this case, we need .E[Ri (1, 1)] ≥ E[Ri (1, 0)], which is equivalent to q(bon + pen) ≥ pen.

.

(9)

• Case B: .ei = 0, i.e., worker i does not exert effort. Similarly, truthfully answering the truth detection question yields E[Ri (0, 0)] = 0,

.

(10)

while lying yields E[Ri (0, 1)] = q(−pen) + (1 − q) · bon .

= bon − q(bon + pen).

(11)

To ensure truthfulness in this case, we need .E[Ri (0, 0)] ≥ E[Ri (0, 1)], which is equivalent to q(bon + pen) ≥ bon.

.

(12)

Combining the two conditions in Case A and Case B that ensure truthfulness, we obtain q(bon + pen) ≥ max{bon, pen}.

.

(13)

Next, we discuss how the workers would behave under the truth detection mechanism. Given the reward policy .pd  (bon, pen, prob), each worker decides the effort exertion .ei and the reporting strategy .ri . Theorem 3 characterizes each worker’s optimal strategy. Theorem 3 Given any .pd, there exists a cost threshold c∗ (pd)  prob · [q · bon + q · pen − pen],

.

so that worker i’s optimal strategy .si∗ (pd) is characterized by

(14)

Incentive Mechanism Design for Mobile Crowdsourcing Without Verification

si∗ (pd) =

 (0, rd),

if ci > c∗ (pd),

(1, 1),

if ci ≤ c∗ (pd).

.

131

(15)

A formal proof of Theorem 3 is given in Appendix B in [16]. In (14), the cost threshold .c∗ (pd) represents the reward threshold given by the platform to worker i, when he chooses .(1, 1). Hence, Theorem 3 shows that if the platform’s truthful reward exceeds the worker’s cost, the worker should exert effort and truthfully report; otherwise, the worker exerts no effort and random reports. Our truth detection mechanism helps address the collusion issue. Under truth detection, a worker has no incentive to collude because his payoff does not depend on the other workers’ reports. More specifically, recall that the truth detection is in response to the question “Have you exerted effort finishing the task?” When workers collude to misreport their solutions without investing any effort, there is a good chance (i.e., the detection accuracy .q > 0.5) that they will be detected if they lie about exerting effort. When considering this, a worker will deviate to exerting effort and truthfully report if the provided truthful reward is sufficiently enough. Hence, under a proper reward design, truth detection can prevent worker collusion. We end this section by providing key insights associated with Questions 4a and 4b as follows.

•

> Important

• One can design an incentive mechanism based on a truth detection technology, which relies on the independent verification of the correctness of each worker’s response to some question. • The truth detection can prevent worker collusion, i.e., workers will not benefit by using any colluding reporting strategies (e.g., always report the same solution).

5 Approaches to Information Incompleteness In Sect. 5.1, we motivate information incompleteness and ask some key questions. In Sect. 5.2, we propose an online mechanism design. In Sect. 5.3, we present the key results and insights.

5.1 Motivating Examples and Key Questions Most prior work on mobile crowdsourcing without verification relies on the strong assumption that workers’ solution accuracy levels are public information, i.e., such

132

C. Huang et al.

information is known by both the platform and the workers [4, 7, 9, 17, 18]. This assumption may not hold in many practical scenarios. It is becoming more difficult for a platform to access and make use of personal data due to an increasing desire for privacy protection [19]. Without personal data, the platform cannot easily estimate the workers’ solution accuracy levels. We aim to design an incentive mechanism without requiring the platform to know even the workers’ accuracy distribution. This leads to the following two questions:

•

? Questions

• Question 5a: how should the platform design the incentive mechanism without knowing the workers’ accuracy distribution? • Question 5b: how accurate and how fast can the incentive mechanism learn the workers’ true accuracy distribution?

5.2 Solution: Randomized Learning Mechanism We build our mechanism design based on majority voting. However, the online learning mechanism design is very challenging in our context, as there is no accessible ground truth to verify either the workers’ task solutions or their accuracy reports. We address such a challenge by proposing an online learning framework, where the workers estimate the population’s accuracy distribution based on their accuracy reports as tasks are assigned and finished over time. In practice, the platform may require the workers to report their accuracy levels. For example, in mobile crowdsensing, the platform may require the workers to report the types of their devices, which reflect the sensing accuracy. In scholarly peer review, the platform may ask the reviewers to report their estimated expertise level before reviewing the submissions. Through the online mechanism, the platform gradually learns the proper reward level to incentivize workers to truthfully report their accuracy. The online nature of the proposed mechanism enables real-time implementation. This is a desirable property, as the ability to finish rapidly incoming tasks on the fly can help reduce the delay and save cost for the platform. The mathematical details behind the online mechanism design can be highly involved and tedious. For additional details, we refer the interested readers to our previous work in [20, 21]. We only include the main ideas and intuitions in the following discussions. In an online learning mechanism, the platform and the workers repeatedly interact with each other. We use .t ∈ {1, 2, · · · , T } to denote the time slot index. The interaction between the platform and workers in each slot t can be summarized in three steps (see also Fig. 4):

Incentive Mechanism Design for Mobile Crowdsourcing Without Verification

133

Platform estimates accuracy distribution and sets reward bundle

Worker submit accuracy reports

Workers submit

Workers choose effort exertion and reporting strategies Platform sets reward

Workers choose strategies

Time slot

Fig. 4 Randomized online learning mechanism

In Step 1, each worker submits its own accuracy level, which can be different from the worker’s actual accuracy level. In Step 2, the platform estimates the workers’ accuracy distribution using all historical accuracy reports. Then, it uses the estimated distribution information to calculate the reward amounts to users. In Step 3, workers choose the effort exertion and reporting strategies. For the online learning mechanism to be effective, it should incentivize the workers to truthfully report their accuracy levels. The key is to use randomized reward design in Step 2, in which the randomization can offset the benefits of workers from misreporting their accuracy levels. For example, if a worker i underreports his accuracy, he will have a larger chance of matching the majority solution. This is because the platform will use his accuracy report to calculate other workers’ rewards, and his under-reporting will lead other workers’ rewards to be larger. With larger rewards, other workers are more likely to exert effort (and truthfully report solutions). As a result, worker i will have a larger chance of matching the majority solution. However, when the platform randomly chooses a reward, under-reporting increases the probability of getting a small (or even zero) reward. This reduces the benefit a worker obtains from under-reporting his accuracy level. Similar logic applies when workers over-report accuracy levels.

5.3 Results and Insights We analyze the performance of the online mechanism in terms of learning the workers’ accuracy distribution. Theorem 4 summarizes this result. Theorem 4 The randomized online mechanism achieves a sub-linear regret .o(T ) in learning the workers’ true accuracy distributions. Furthermore, the mechanism is asymptotically optimal. The detailed proof is given in Appendix D in [21]. Theorem 4 implies that if the platform uses a properly designed randomized mechanism, it can induce the workers to truthfully report their accuracy levels in the long term, thus effectively learning the true accuracy distribution of the workers.

134

C. Huang et al.

We summarize the key results and insights related to Questions 5a and 5b as follows.

•

> Important

• Without knowledge of workers’ accuracy distributions, the platform can design an online incentive mechanism via repeatedly asking workers to report their accuracy levels in addition to the task solutions. • To incentivize workers to truthfully report their accuracy levels, the mechanism should use randomization techniques to offset the benefits of workers from misreporting accuracy levels. • Our proposed mechanism learns the workers’ true accuracy distribution in .o(T ) and it is asymptotically optimal.

6 Approaches to Information Asymmetry In Sect. 6.1, we motivate information asymmetry and ask some key questions. In Sect. 6.2, we propose the mechanism design. In Sect. 6.3, we present the key results and insights.

6.1 Motivating Examples and Key Questions Most prior literature (e.g., [4, 22, 23]) has studied the problem as a game with symmetric information. The common assumption is that both the workers and the platform have the same information regarding the environment, e.g., worker capabilities. In many mobile crowdsourcing platforms, however, information regarding the worker characteristics is asymmetric between the platform and workers. Usually, the platform has more information regarding worker characteristics through market research and past experiences. For example, for Waze and Amazon Mechanical Turk, each worker’s historical performance is known by the platform but not by the other workers [1]. In this chapter, we consider information asymmetry between the platform and the workers, which is practically important, under-explored, and more difficult to deal with than the models in existing studies. As will be seen, such information asymmetry significantly complicates the analysis of both the workers’ behaviors and the platform’s incentive design.

Incentive Mechanism Design for Mobile Crowdsourcing Without Verification

135

The study of information asymmetry raises a few interesting research questions:

•

? Questions

• Question 6a: does the platform have an incentive to reveal the asymmetric information to workers? • Question 6b: if yes, does the platform have an incentive to manipulate the revealed information (i.e., not telling the truth)?

We focus on answering the above key questions in the remainder of this section.

6.2 Solution: Bayesian Persuasion Mechanism Similar to Sect. 3.1, we consider a mix of high- and low-accuracy workers. The platform knows the number of each type (i.e., the number of high-accuracy workers k), but the workers do not. Due to information asymmetry, besides the incentive design, the platform also needs to decide how to reveal information. In particular, the platform can strategically announce the worker information to induce desired worker behavior and maximize its payoff. We model the information asymmetry and revelation between the platform and the workers using a Bayesian persuasion framework (e.g., [24–26]) as follows (see also Fig. 5): • Step 1: neither the platform nor the workers know k, and the platform must commit to a long-term information revelation strategy. • Step 2: the value of k is realized and observed by the platform, but not the workers. • Step 3: the platform announces a value that may be different from the real value of k to the workers, according to its previously committed strategy. The detailed mathematical formulation can be found in [27, 28]

is realized and observed by platform

Before is realized, platform commits to information policy

Step 1: Platform commits

Platform announces a value according to information policy Step 2: Platform observes

Step 3: Platform announced

Bayesian persuasion framework

Fig. 5 A Bayesian persuasion framework

136

C. Huang et al.

6.3 Results and Insights We study two types of workers: (1) naive workers who fully believe in the platform’s announcement and (2) strategic workers who update their prior belief based on the announced information. The naive worker case serves as a benchmark, and it can model the scenario where workers are confident in the platform’s announced information (especially those platforms with good reputations). It can also model the scenario where workers have limited reasoning capabilities to deduce the authenticity of the announcement. The strategic worker case, however, leads to more intriguing results. Such a case fits the scenario where workers do not trust the platform and where workers are strategic and have high reasoning capabilities (e.g., [10, 20]). As we will show below, the platform’s optimal information revelation strategies are very different when facing these two types of workers. Theorem 5 gives platform’s equilibrium information revelation strategy. Theorem 5 (i) For naive workers, the platform should always announce a high value, regardless of the realization of k. (ii) For strategic workers, it is not always optimal to announce a high value. For a detailed proof of this see Appendices E and J in [28]. We discuss the intuition behind Theorem 5 below. • Case (i): the naive workers fully trust the platform’s announcement. When hearing a higher value, the naive workers believe that there is a larger proportion of high-accuracy workers. The workers would think that they have a larger probability of getting the reward and hence the platform can use a smaller reward amount to incentivize workers to exert effort. This leads to a higher platform payoff. • Case (ii): this is counter-intuitive, as one may think that the platform should lure the workers into believing in a high overall worker capability (by always announcing a higher value) to maximize its payoff. Intuitively speaking, announcing a high value with proper frequencies can improve the platform payoff, as this will lead the strategic workers to believe the announcement and act accordingly. However, when the platform announces a high value too frequently, the strategic workers may doubt such announcement (based on the Bayesian update) and act as if the realized value is low. Hence, larger rewards are required to incentivize them, which will hurt the platform payoff. We end this section by providing the key insights related to Questions 6a and 6b as follows.

Incentive Mechanism Design for Mobile Crowdsourcing Without Verification

137

•

> Important

• The platform can manipulate the naive workers by always announcing a high value. • It is not always optimal to announce a high value to strategic workers. In certain cases, the platform may even announce a lower value than the actual one. • The platform needs to carefully design information policy, as truthful announcement is not optimal for either naive or strategic workers.

7 Conclusion and Open Problem 7.1 Future Challenges and Open Issues The research on incentive mechanism design for mobile crowdsourcing without verification is far from complete, and there are many interesting open issues to address. We discuss some of them as follows.

7.1.1

Joint Optimization of Information Elicitation and Aggregation

Existing studies mainly focus on either deriving mechanisms to elicit truthful and high-quality information or proposing effective schemes to properly aggregate workers’ reported information. The interaction between information elicitation and information aggregation is little understood, especially when the ground truth information is unavailable. It is important to establish a unified framework to understand the intertwining between elicitation and aggregation and further propose a joint optimization framework to simultaneously improve the efficiency and effectiveness of information elicitation and information aggregation.

7.1.2

Competitive Market

In practical applications, a worker may face more than one mobile crowdsourcing platform competing for market share with different tasks and reward options. Besides eliciting workers’ truthful solutions to the designated tasks, the crowdsourcing platforms (e.g., oligopolies) also need to effectively compete with other platforms by careful design of incentive and pricing mechanisms. Therefore, we posit that a novel and thorough analytical framework is needed to understand the competition in mobile crowdsourcing markets where there is no verifiable ground truth information.

138

7.1.3

C. Huang et al.

Worker Bounded Rationality

Existing research studies typically assume that the workers are fully rational, expected payoff maximizers. This canonical assumption in economics may fail to hold in many practical cases. More specifically, crowdsourced workers may not be able to make fully rational decisions and hence may not behave according to the optimal solution as specified by the mechanism. Empirical studies have also demonstrated that the workers are looking for a satisfactory solution instead of the optimal one due to their cognitive limitations and the time available to make decisions. While the bounded rationality in mobile crowdsourcing without verification has received very little attention, we posit that the consideration of such a bounded rationality will have a major impact on the practical incentive mechanism design, and more research endeavors should be made along this line.

7.1.4

Worker Privacy and Moral Issues

Prior work has been focusing on deriving efficient mechanisms to elicit highquality and truthful solutions from workers. However, some sophisticated ethical issues may arise. First, the workers might be sensitive to the private information leaked during the crowdsourcing process. For example, a worker may worry that the collected information is not being properly used. Also, the worker may have the right to protect his “cognitive liberty,” which limits the crowdsourcing platform’s right to peer into an individual’s “thought” without his/her consent. Such ethical considerations are another important direction of future work.

7.2 Conclusion In this chapter, we first focus on the workers’ heterogeneous capabilities and investigate how this heterogeneity affects the worker behaviors and the mobile crowdsourcing platform’s incentive mechanism design without verification. We then propose a novel truth detection mechanism that addresses worker collusion that is likely to happen in many mobile crowdsourcing applications. Next, we study two information cases (i.e., information incompleteness and information asymmetry) and investigate their impact on the incentive mechanism design. Our analysis and results reveal the impact of various important factors (i.e., worker heterogeneity/collusion, information incompleteness/asymmetry) on the incentive mechanism design, as well as initialize the exploration of mechanism design in multiple dimensions (e.g., a joint design of incentives and information revelation). We also outline some challenges and open issues that deserve further study in the future.

Incentive Mechanism Design for Mobile Crowdsourcing Without Verification

139

References 1. http://www.mturk.com 2. https://steem.com/ 3. http://www.iot.fer.hr/index.php/cupus-crowdsensing-app-en/ 4. N. Miller, P. Resnick, R. Zeckhauser, Eliciting informative feedback: The peer-prediction method. Manag. Sci. 51(9), 1359–1373 (2005) 5. https://www.ideaconnection.com 6. Y. Luo, N.B. Shah, J. Huang, J. Walrand, Parametric prediction from parametric agents. Oper. Res. 66(2), 313–326 (2018) 7. C. Huang, H. Yu, J. Huang, R.A. Berry, Crowdsourcing with heterogeneous workers in social networks, in Proceedings of the 2019 IEEE Global Communications Conference (GLOBECOM) (2019), pp. 1–6 8. Y. Liu, Y. Chen, Learning to incentivize: eliciting effort via output agreement, in Proceedings of the IJCAI (2016), pp. 3782–3788 9. C. Huang, H. Yu, J. Huang, R.A. Berry, Incentivizing crowdsourced workers via truth detection, in Proceedings of the 2019 IEEE Global Conference on Signal and Information Processing (GlobalSIP) (2019) 10. C. Huang, H. Yu, J. Huang, R. Berry, Eliciting information from heterogeneous mobile crowdsourced workers without verification. IEEE Trans. Mob. Comput. 21(10), 3551–3564 (2021) 11. G. Radanovic, B. Faltings, R. Jurca, Incentives for effort in crowdsourcing using the peer truth serum. ACM Trans. Intell. Syst. Technol. 7(4), 1–28 (2016) 12. I. Krajbich, C. Camerer, J. Ledyard, A. Rangel, Using neural measures of economic value to solve the public goods free-rider problem. Science 326(5952), 596–599 (2009) 13. R.H. Nugroho, M. Nasrun, C. Setianingsih, Lie detector with pupil dilation and eye blinks using Hough transform and frame difference method with fuzzy logic, in Proceedings of the 2017 International Conference on Control, Electronics, Renewable Energy and Communications (ICCREC) (2017), pp. 40–45 14. S.M. Ho, X. Liu, C. Booth, A. Hariharan, Saint or sinner? language-action cues for modeling deception using support vector machines, in Proceedings of the Social, Cultural, and Behavioral Modeling: 9th International Conference, SBP-BRiMS 2016, Washington, DC, USA, June 28-July 1, 2016 (Springer, Berlin, 2016), pp. 325–334 15. https://www.iborderctrl.eu/ 16. C. Huang, H. Yu, R. Berry, J. Huang, Using truth detection to incentivize workers in mobile crowdsourcing. IEEE Trans. Mob. Comput. 21(6), 2257–2270 (2020) 17. J. Witkowski, B. Nebel, D.C. Parkes, Robust peer prediction mechanisms, Ph.D. dissertation (University of Freiburg, Freiburg im Breisgau, Germany, 2015) 18. Y. Kong, G. Schoenebeck, Equilibrium selection in information elicitation without verification via information monotonicity, in ITCS, vol. 94 (2018), pp. 1–20 19. G. Liao, X. Chen, J. Huang, Optimal privacy-preserving data collection: a prospect theory perspective, in Proceedings of the GLOBECOM 2017-2017 IEEE Global Communications Conference (2017), pp. 1–6 20. C. Huang, H. Yu, J. Huang, R.A. Berry, Online crowd learning with heterogeneous workers via majority voting, in Proceedings of the 2020 18th International Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks (WiOPT) (2020), pp. 1–8 21. C. Huang, H. Yu, J. Huang, R. Berry, Online crowd learning through strategic worker reports. IEEE Trans. Mob. Comput. (2022) https://doi.org.10.1109/TMC.2022.3172965. 22. B. Waggoner, Y. Chen, Output agreement mechanisms and common knowledge, in Proceedings of the AAAI Conference on Human Computation and Crowdsourcing, vol. 2(1) (2014) 23. V. Shnayder, A. Agarwal, R. Frongillo, D.C. Parkes, Informed truthfulness in multi-task peer prediction, in Proceedings of the 2016 ACM Conference on Economics and Computation (2016), pp. 179–196

140

C. Huang et al.

24. V.P. Crawford, J. Sobel, Strategic information transmission, in Econometrica: Journal of the Econometric Society (1982), pp. 1431–1451 25. E. Kamenica, M. Gentzkow, Bayesian persuasion. Am. Econ. Rev. 101(6), 2590–2615 (2011) 26. M. Kurschilgen, I. Marcin, Communication is more than information sharing: the role of statusrelevant knowledge. Games Econom. Behav. 113, 651–672 (2019) 27. C. Huang, H. Yu, J. Huang, R.A. Berry, Strategic information revelation in crowdsourcing systems without verification, in Proceedings of the IEEE INFOCOM 2021-IEEE Conference on Computer Communications (2021), pp. 1–10 28. C. Huang, H. Yu, J. Huang, R. Berry, Strategic information revelation mechanism in crowdsourcing applications without verification. IEEE Trans. Mob. Comput. 22(5), 2989–3003 (2021)

Part III

Key Technical Components: Task Allocation

Stable Worker–Task Assignment in Mobile Crowdsensing Applications Fatih Yucel, Murat Yuksel, and Eyuphan Bulut

1 Introduction Mobile crowdsensing (MCS) aims to complete spatio-temporal sensing tasks, which usually require massive expenses and execution times when performed individually, using the help from mobile participants (workers). This happens through recruitment of mobile users and leveraging the sensing capabilities (e.g., microphone, camera, and GPS) on their mobile devices. In an MCS system, there are mainly four entities; namely, the platform, requesters, tasks, and workers. Requesters define the tasks and post them to the platform with the requirements of their tasks, e.g., deadline, reward, and budget. Workers are the mobile users that register to the system with a set of their capabilities and limitations, e.g., a regional service area. The platform, knowing the tasks requested and the works eligible for each task in the system, makes the assignment of tasks to the requesters. This assignment can be made through a predefined logic with some goal, e.g., maximum tasks matched with minimum cost to requesters. Moreover, the platform, instead of performing the matching itself, can let the workers and requesters communicate and agree on an assignment in a distributed way. In this case, the platform acts as a mediator between the workers and task requesters. One of the key and most studied problems in MCS systems is the assignment of sensing tasks to workers. The challenging part of this problem is there are many parameters that can be considered and objectives can vary depending on

F. Yucel · E. Bulut () Department of Computer Science, Virginia Commonwealth University, Richmond, VA, USA e-mail: [email protected]; [email protected] M. Yuksel Department of Electrical and Computer Engineering, University of Central Florida, Orlando, FL, USA e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 J. Wu, E. Wang (eds.), Mobile Crowdsourcing, Wireless Networks, https://doi.org/10.1007/978-3-031-32397-3_6

143

144

F. Yucel et al.

the system design. On the one side, there are task requesters who want their tasks to be completed in the best way (e.g., with the minimum cost to them and with the minimum delay), and, on the other side, there are workers who would like to make the best profit from the rewards they obtain once the tasks are completed and after their costs are taken out. There is also the platform that might be getting some registration cost from each user (i.e., task requester and worker) or some fee from each task completed and thus may aim to match as many tasks as possible to the eligible workers or maximize the total quality of service (QoS) received by the task requesters [1, 2]. In most of the existing studies in the literature, however, the objective during the task assignment process is defined in the favor of either one side (workers or task requesters) or for the system/platform itself. However, such assignments that do not take into account the individual needs and preferences of different entities may result in dissatisfied users and impair their future participation. This is because users in practice may not want to sacrifice their individual convenience for the sake of system utility or the other side’s benefit. In this chapter, we study the task assignment problem in MCS systems considering the preferences of entities involved in an MCS scenario. Preference-aware or stable matching has indeed been extensively studied in general bipartite matching problems especially in the economics literature [3]. However, these studies do not consider the features that are specific to MCS systems such as budget of task requesters, uncertainty in matching opportunities due to unknown worker trajectories, and time constraints of tasks. The stable matching problem for task assignment in MCS systems can indeed be defined in many different ways because of the varying settings of MCS scenarios, and, in each, the solution can be based on different approaches. Thus, we overview the different stable matching definitions studied recently in the MCS domain for the task assignment problem and provide a summary of proposed solutions. We also refer the readers to the actual studies for the details of the solutions. We hope this will highlight the spectrum of different stability definitions considered for MCS systems and summarize the differences. The remainder of the chapter is organized as follows: in Sect. 2, we first start with a background on worker–task assignment problem and proposed solutions in the MCS literature, as well as with a background on stable matching theory and its applications in several domains. We then provide a motivation for using stable matching in MCS in Sect. 3. In Sect. 4, we provide with the classification of the MCS scenarios studied while considering preference awareness in the task assignments. We provide the blocking or unhappy pair definitions considered in different MCS settings and discuss how the stability is defined in each. We also summarize the algorithms proposed to find the stable matchings in such settings. Finally, we discuss the open problems that need to be studied in the MCS systems while considering the preference awareness and conclude the chapter in Sect. 5.

Stable Worker–Task Assignment in Mobile Crowdsensing Applications

145

2 Background In this section, we first provide an overview of worker–task assignment solutions in different MCS settings. Then, we look at the matching problems and solutions studied considering the user preferences in different domains. During these overviews as well as through the rest of this chapter, we base our discussion considering the three main categories of MCS scenarios illustrated in Fig. 1. In participatory sensing, workers can interrupt their daily schedule to carry out the assigned tasks (e.g., measuring air quality at a specific location) at the expense of additional cost, e.g., traveling distance. In opportunistic sensing, workers do not alter their schedules and perform the assigned tasks (e.g., traffic monitoring) only when they happen to be in the task regions, thus without an additional cost but with less likelihood of visiting task regions. Finally, in hybrid sensing, workers provide some flexibility through a set of alternative paths they can follow and let the platform decide which one to use to increase the utility from the matching.

2.1 Worker–Task Assignment in Mobile Crowdsensing The overall performance of an MCS system and the satisfaction of its users are highly dependent on the efficiency of the assignments; thus, there have been many task assignment solutions proposed in different studies. These studies have considered various objectives in the task assignment process such as maximizing the number of completed tasks [5], minimizing the completion times of tasks [6], minimizing the incentives provided to the users [7], assuring the task or sensing quality [2] under some constraints on traveling distance [8], energy consumption [1], and expenses of task requesters [9]. Beyond these works, the issues of security [10], privacy [11], and truthfulness [12] have also been considered in the worker recruitment process.

Opportunistic Semi-Opportunistic

Participatory Task Fig. 1 Three different MCS scenarios considered [4]. Solid line is the regular path that the user follows. Dashed blue lines are alternative similar paths to the user’s regular path. Red line is the path the user is forced to follow to complete the tasks

146

F. Yucel et al.

In participatory MCS, since workers need to travel between the task regions to perform the assigned tasks, a key factor that shapes the task assignment process is the travel costs of the workers. In [13], the authors investigate the problem of minimizing the total travel costs of the workers while maximizing the number of completed tasks and keeping the rewards to be paid to the workers as low as possible. In [8], the authors study the task assignment problem in an online setting and aim to maximize the total task quality while ensuring that the travel costs of the workers do not exceed their individual travel budgets. In [14], the authors adopt a system model in which each worker has a maximum traveling distance that needs to be considered in the assignment process, and the objective is to maximize the profit of the platform. The authors propose a deep reinforcement learning-based scheme that significantly outperforms the other heuristic algorithms. In [15], the goal is also to minimize the travel distance of workers, and however, differently from the aforementioned studies, the authors consider the issue of user privacy and present a mechanism that finds the task assignments without exposing any private information about workers or task requesters. Lastly, in [16], the authors study the destinationaware task assignment problem in participatory crowdsourcing systems. On the other hand, in opportunistic MCS, the main objectives are to maximize the coverage and to minimize the completion times of the tasks due to the uncontrolled mobility, i.e., a task can only be performed if its region resides on the trajectory of a worker. In [17], the authors study the maximum coverage task assignment problem in opportunistic MCS with worker trajectories that are known beforehand. It is assumed that each task needs to be performed at a certain point of interest and has a weight that indicates how important its completion is to the platform, which has a fixed budget and can hence recruit only so many workers. The objective of the platform is to select a set of workers within the budget constraints, which maximizes the weighted coverage over the set of tasks according to the given trajectories of workers. The authors develop a .(1 − 1/e)-approximate algorithm with a time complexity of .O(n5 ), where n is the number of workers in the system. [18] studies the same problem and proposes a greedy algorithm that, despite not having a theoretical guarantee, outperforms the algorithm proposed in [17] in terms of achieved coverage in certain settings and runs in .O(n2 ) time. Adding the delivery probability of the sensed data to the goals notably changes the problem being studied as shown in [19]. In this design, after carrying out a task, a worker should either deliver the sensed data to the server through one of the collection points (i.e., Wi-Fi APs) on his trajectory or transmit it to another user who will deliver it for him. Thus, here, not only does the platform need to estimate whether and when workers would visit task regions and collection points, but it is also crucial to obtain and utilize the encounter frequencies of workers to improve the delivery probability of the sensed data. The authors present different approximation algorithms for the systems with deterministic and uncertain worker trajectories and evaluate their performance on real datasets. The data delivery aspect of the problem in [19] has also been studied in [20] and [21]. They both utilize Nash Bargaining Theory to decide on whether or not selfish data collectors and mobile (relay) users who only take part in delivery of sensed data would like to

Stable Worker–Task Assignment in Mobile Crowdsensing Applications

147

cooperate with each other according to their utility in either scenario. However, in [21], the authors consider a more complete mobile social network model and present an enhanced data collection mechanism. Another aspect to the MCS system design is the uncertainty of workers’ trajectories. In [22], the problem of maximizing spatio-temporal coverage in vehicular MCS with uncertain but predictable vehicle (i.e., worker) trajectories is investigated. The authors first prove that the problem is NP-hard when there is a budget constraint and then propose a greedy approximation algorithm and a genetic algorithm. In [23], the authors also assume predictable worker trajectories. However, they focus on the task assignment problem in a mobile social network where task assignments and delivery of sensed data are realized in an online manner when task requesters and workers encounter with each other. They aim to minimize the task completion times and propose different approximation algorithms to optimize both worst-case and average-case performance. For predictions of worker trajectories, [22] uses spatiotemporal trajectory matrices, while [23] assumes that user inter-meeting times follow an exponential distribution, which is used widely in mobile social networks [24–26] literature. Recently, there are studies [27, 28] that look at the task assignment problem in a hybrid system model to simultaneously leverage the advantages of participatory and opportunistic MCS. In [27], the authors propose a two-phased task allocation process, where opportunistic task assignment is followed by participatory task assignment. The objective behind this design is to maximize the number of tasks that are performed in an opportunistic manner, which is much less costly compared to participatory MCS, and then to ensure that the tasks that cannot be completed by opportunistic workers are assigned to workers that are willing to perform tasks in a participatory manner to alleviate the coverage problem in opportunistic MCS. On the other hand, in [28], the workers carry out the sensing tasks only in opportunistic mode, but they provide the matching platform with multiple paths that they would take if requested, instead of a single path as in classic opportunistic MCS. This enables the platform to find a matching with a high task coverage. However, none of the studies considers the stability in the assignment based on the preferences of the workers and task requesters.

2.2 Matching Under Preferences Stable matching problem is introduced in the seminal paper of Gale and Shapley [29] and can simply be defined as the problem of finding a matching between two groups of objects such that no pair of objects favor each other over their partners in the matching. Gale and Shapley have also introduced what is called the deferred acceptance procedure that finds stable matchings in both one-to-one matching scenarios (e.g., stable marriages) and many-to-one matching scenarios with capacity constraints (e.g., stable college admissions) in .O(mn) time, where m and n are the size of the sets being matched. Since its introduction in [29], the concept

148

F. Yucel et al.

of stability has been utilized in different problems including hospital resident assignment [30], resource allocation in device-to-device communications [31], SDN controller assignment in data center networks [32], supplier and demander matching in electric vehicle charging [33–35], and peer-to-peer energy sharing among mobile devices [36–38]. Some matching problems allow or require nodes in one or both sides to be matched with multiple nodes, i.e., many-to-one and many-to-many matching problems. A few studies investigate the issue of stability in such matching problems. For instance, [39] and [40] study the many-to-one stable matching of students– colleges and doctors–hospitals, respectively. In [39], all colleges define a utility and a wage value for students and aim to hire the best set of students (i.e., with the highest total utility) within their budget constraints. Each student also forms a preference list over colleges. The authors prove that there may not exist a stable matching in this setting and even checking the existence is NP-hard. However, they provide a polynomial time algorithm that finds pairwise stable matchings in the so-called typed weighted model where students are categorized into groups (e.g., Master’s and PhD students) and colleges are restricted to define a set of possible wages for each group, i.e., they cannot define a particular wage for each student. [40] studies the same problem and proposes two different fully polynomial time approximation algorithms with some performance guarantee in terms of coalitional stability for general and proportional (i.e., the wage of doctors are proportional to their utility for hospitals) settings. However, the study does not provide an experimental analysis of the algorithms or discuss their actual/expected performance in these settings. Moreover, the proposed solutions can only be applied to a limited set of scenarios. There are some studies that look at the stable matching problem in settings with incomplete information on user preferences or dynamic user arrivals/departures. [41] and [42] both study the dynamic stable taxi dispatching problem considering passenger and taxi preferences. However, the objective adopted in [42] is to find locally optimal stable assignments for a given time-point, whereas that in [41] is to minimize the number of unhappy taxi–passenger pairs globally. The authors in [43] investigate the stable matching problem in the presence of uncertainty in user preferences. [44] looks at the problem of minimizing the number of partner changes that need to be made in a stable matching to maintain stability when preference profiles of some users change in time. Lastly, [45] studies an interesting combination of famous stable marriage and secretary (hiring) problems. The concept of stability is studied in multi-dimensional matching problems as well. In [46], the authors introduce the three-dimensional stable matching problem. In this problem, there are three sets of different types, each individual from a set has a preference list over all pairs from the other two sets, and the goal is to form stable/satisfactory families of three, where each individual in a family is a member of a different set. Wu [47] investigates a different version of this problem, where each individual has a one-dimensional preference list over the individuals from the other two sets instead of over all pairs of individuals as in [46]. In [48], the authors extend the stable roommates problem [49] to a three-dimensional setting, where a

Stable Worker–Task Assignment in Mobile Crowdsensing Applications

149

set of individuals are assigned into groups of three instead of two based on their preferences. Lastly, in [50], the authors study the problem of matching data sources, servers, and users in a stable manner in video streaming services under restricted preference settings. In a typical MCS system, the objectives of workers and task requesters can be defined as to maximize their profits and to have their tasks completed with the highest quality possible, respectively. Thus, they are likely to have preferences over possible assignments they can get, and the task assignment in MCS can be consequently characterized as a matching problem under preferences. Apart from the studies that we will present in the next section, there are only a few studies that consider user preferences in mobile crowdsensing (or in mobile crowdsourcing). In [51], the authors study the budget-constrained many-to-many stable task assignment problem, which they prove to be NP-hard, and propose efficient approximation algorithms. Similarly, in [52], stability in many-to-many assignments has also been studied considering a competition congestion metric. In [53], the authors study the same problem, but in a system model with capacity constraints. On the other hand, [54] considers a many-to-one matching setting and introduce additional constraints (e.g., minimum task quality requirements) that are taken into account in the matching process, along with user preferences. Lastly, in [55], the authors consider a budget-constrained MCS system where the quality of a worker is identical for all tasks and present an exponential-time algorithm to find weakly stable many-to-one matchings. Note that there are also studies (e.g., [56]) that use auctions for fairness in crowdsensing systems, but these are out of the scope of this work.

3 Why Should We Care About Stability in MCS? The answer of this question is rather obvious in MCS campaigns with no central authority, where the task assignments are made in a distributed manner or by a piece of software that runs on the cloud and is managed jointly by the users. This is because aside from malicious intent, there is no reason for the users of such systems to adopt a task assignment mechanism that would favor certain individuals, and thus the long-term functioning of such systems can only be made possible by considering user preferences in a fair way and producing stable assignments where no user has an incentive to deviate from their assigned partners. However, in MCS campaigns with a central service provider (SP) which aims to maximize its profits, the question set forth above is more of a business question than an engineering question. The main motivation for an SP to consider the user preferences and provide a stable worker–task assignment would be to ensure the continuous participation of the existing users and to promote their willingness to perform the assigned tasks. However, this may not align perfectly with the SP’s goal of maximizing its own utility. So, here the SP faces a critical business decision: it should either choose to maximize its own utility by disregarding user preferences,

150

F. Yucel et al.

but potentially suffer from the consequences of doing so (e.g., unhappy users abandoning the platform), or should prioritize user preferences to keep the users actively participating in the MCS campaign at the cost of its own utility. In order to demonstrate this trade-off between the utility of SP and user happiness, we have performed a series of experiments and shown that [57] a task assignment solely maximizing the utility of the SP without considering user preferences may make the majority of users unhappy with their assignments. In such cases, if the users that get such dissatisfying assignments do not obey the task assignment results and not perform the task, as they can be selfish and can consider their own benefit, the SP will face a significant utility loss. Furthermore, if this dissatisfaction causes even a certain part of the users to abandon the system in each assignment cycle, over time, the numbers of workers and task requesters participating in the campaign will decrease and this will result in an exponential utility loss for the SP. Even though these results show the value of considering user preferences, they do not necessarily indicate that the SP should always only care about user preferences and ignore its own utility. In fact, in some scenarios, the SP may benefit from producing task assignments [57], which maximize its own utility while keeping the conflicts with user preferences as minimal as possible (i.e., the system utility and user preferences as primary and secondary objectives, respectively).

4 Stable Task Assignments in Different MCS Applications In this section, we present the stability definitions considered in three different MCS scenarios (i.e., participatory, opportunistic, and hybrid). Throughout the section, we also use the terms uniform and proportional MCS systems, where the former refers to the MCS scenarios when the QoS provided by each worker is the same for all tasks, and the latter refers to the MCS scenarios where the rewards that are offered to the workers are proportional to the QoS they provide. Note that these are exclusive to the three aforementioned categories (i.e., participatory, opportunistic, and hybrid). Throughout the chapter, we assume a system model with a set of workers .W = {w1 , w2 , . . . , wn } and a set of sensing tasks .T = {t1 , t2 , . . . , tm }. We also define .ct (w) as the cost of performing task t for worker w and use .rt (w) for the reward that worker w is offered to carry out task t.

4.1

Participatory MCS

An MCS system is called participatory if the available workers considered in the task assignment process are actively waiting for new tasks to be assigned and are willing to go to the locations of the assigned tasks immediately or whenever requested by the task requesters. In other words, the task assignment mechanisms in these systems

Stable Worker–Task Assignment in Mobile Crowdsensing Applications

151

can be assumed to have the ability to control the mobility of workers, generally within some constraints specified by workers. This, of course, creates additional cost to the task requesters, as workers need to travel to task locations and ask for more rewards to cover this cost. In MCS systems where each task can recruit multiple workers within their budget constraints, the stability can be defined in two different ways: pairwise and coalitional. Due to the many-to-one nature of task assignments and budget constraints, the conditions of both pairwise and coalitional stability differ from the classic stability conditions specified in [29], and thus existing stable matching solutions cannot be used to find pairwise or coalitional stable matchings in such systems. Moreover, depending on the relation (i.e., proportional or not) between the QoS provided by workers and the reward they gain, the hardness of the problem and the corresponding solution approach completely change. Since a rational worker will aim to maximize their profit and will not accept to perform the tasks that cost higher than the corresponding rewards that will be paid, we can define the preference list of worker w as Pw = ti1 , ti2 , . . . , tik ,

.

(1)

where .Pw ⊆ T, .∀t ∈ Pw , .rt (w) > ct (w), and .∀t  = tij , t  = tij +1 , .rt  (w)−ct  (w) > rt  (w) − ct  (w). We denote the j th task (.tij ) in .Pw by .Pw (j ) and utilize .t  w t  notation to express that .t  precedes .t  in .Pw . On the other hand, a rational task requester will try to maximize the total quality of service (QoS) that can be obtained from the recruited workers considering his/her budget constraint. Let .qt (w) denote the QoS that worker w can provide for task t and .bt denote the budget of task t. Then, we can define the preference list of task t as Pt = S1 , S2 , . . . , Sk ,

.

(2)

  where .∀S ∈ Pt , .S ⊆ W and . w∈S rt (w) ≤ bt , and .∀Si , Si+1 ∈ Pt , . w∈Si qt (w) ≥  w∈Si+1 qt (w). Here, given a matching .M and .w ∈ W, t ∈ T, the partner1 of worker w is denoted by .M(w) and the partner set of task t is denoted by .M(t). If .M(u) = ∅ for user .u ∈ W ∪ T, it means user u is unmatched in .M. Note that the last set in the preference list of each task t is .∅, so we have .S t ∅, .∀S ∈ (Pt \ ∅). Also, even though the preference lists of workers do not include .∅, since we assume that the workers in our system are rational, we have .t w ∅, for all .w ∈ W and .t ∈ Pw . We denote the remaining budget of task t in .M by .btM = bt − rt (M(t)).

1 The partner of a worker refers to the task that the worker is assigned to perform, while the partner set of a task refers to the set of all workers assigned with the task.

152

F. Yucel et al.

Table 1 Mobile crowdsensing scenarios (i.e., uniform (U) and proportional (P)) and corresponding applicable algorithms (.∗ indicates the algorithm is applicable but has a very poor performance since it is not specifically designed for that scenario) MCS Type P. & U. P. & N.U. N.P. & U. N.P. & N.U.

UTA [58]  

PSTA [58]    

Heuristic [58]    

SJA [55] 

.φ-STA

 

[59]

.θ -STA

[59]

* *  

Definition 1 (Unhappy Pair) Given a matching .M, a worker w, and a task t form an unhappy (blocking) pair . w, t if .t w M(w), and there is a subset .S ⊆ M(t) such that .{w} t S and .rt (w) ≤ btM + rt (S). Then, a matching .M is said to be pairwise stable if it does not admit any unhappy pairs. Definition 2 (Unhappy Coalition) Given a matching .M, a subset of workers .S ⊆ W and a task t form an unhappy (blocking) coalition . S, t if .∀w ∈ S, .t w M(w) and there is a subset .S  ⊆ M(t) such that .S t S  and .rt (S) ≤ btM + rt (S  ). Similarly, a matching .M is said to be coalitionally stable if it does not admit any unhappy coalition. Following these different unhappy pair and stability definitions and using a classification of MCS systems based on the variability in the QoS provided by the workers for different tasks (uniform/non-uniform) and the relationship between the QoS provided by the workers and the rewards they are offered (proportional/nonproportional), three different stable task assignment algorithms, namely UTA, PSTA, and Heuristic, have been provided for different MCS classes and scenarios in [58]. These algorithms are summarized in Table 1. In [58], we prove that UTA and PSTA algorithms always produce pairwise stable task assignments in uniform and proportional MCS systems, respectively. With simulation results, we also show that our algorithms significantly outperform the state-of-the-art stable task assignment algorithms in most scenarios. Specifically, PSTA and Heuristic algorithms usually achieve the highest outward and overall user happiness, respectively. In participatory MCS systems where the number of workers is scarce, it is also possible that some workers are assigned to multiple tasks to complete all the tasks, while still assigning each task to a single worker. While this provides workers with an opportunity to earn more rewards, the consideration of preferences and stability can be different. Note that in such systems, the task assignments can be performed either instantly or in a predetermined way. In the former, workers are assigned one task at a time, and they are assigned a new one only after completing their currently assigned task. However, this creates an uncertainty for the worker and also the assignments made become not optimal. On the other hand, if the task assignments for all workers could be planned in a foreseeable future (e.g., the next hour or day), such issues can be avoided. This is studied in [60] and a task assignment algorithm

Stable Worker–Task Assignment in Mobile Crowdsensing Applications

153

that not only considers the scheduling of tasks for workers but also respects user preferences so that no user will have a desire to deviate from their assignment (i.e., stable) is proposed.

4.2 Opportunistic MCS An MCS system is called opportunistic if the mobility of workers is not controllable and the tasks can be completed by workers only when they happen to visit the task locations in their regular mobility patterns. Thus, the task assignments should be made considering the likelihood that the workers will visit the task regions within an acceptable time frame. As workers are not directed to a certain location, there is also no travel cost associated with task assignments, and however, it may take longer for workers to visit task regions in an opportunistic manner (compared to participatory sensing). A particularly important objective in the opportunistic MCS systems is to maximize the sensing coverage over a set of points of interest (POIs), which has recently been studied in [9, 18, 61, 62]. However, these studies either do not consider budget constraints of task requesters or assume that there is only a single task requester (i.e., a single budget constraint) in the system. This may not be a practical assumption as there can be multiple task requesters with a unique set of goals and an individual budget constraint. Moreover, some task requesters may prefer to allocate a separate budget for different sets of POIs. Thus, in such systems, stable assignment that considers each requester’s preferences would benefit all. When the utility functions are additive (i.e., the total utility of a set of workers for a task is equal to the sum of their individual utilities). However, when the utility functions are additive (...), the stability can be handled easily. On the other hand, the coverage of workers over a set of POIs is usually non-additive because of the commonly covered POIs by different workers. In order to handle such scenarios, in [63], we propose the following definition for unhappy coalitions. Definition 3 (Unhappy Coalition with Non-additive Worker Utilities) Given a matching .M, a task t and a subset S of workers form an unhappy coalition (denoted by . S, t ) if the following conditions hold for a subset .S  of the workers assigned to task t in .M: • Task t would be better off with S than with .S  , i.e., Ut (S ∪ (M(t) \ S  )) > Ut (M(t)),

.

(3)

where .Ut (S) defines the non-additive utility of set S of workers [63], • Task t can replace .S  with S without violating her budget constraint, i.e., rt (S) − rt (S  ) ≤ btM ,

.

(4)

154

F. Yucel et al.

M where .bt is the remaining budget of task t in .M (i.e., .btM = bt −  w∈M(t) rt (w)), • Every worker w in S prefers task t to task .t  to whom he is currently assigned in .M, i.e.,

∀w ∈ S, gt (w) > gt  (w),

.

(5)

where .gt (w) = rt (w) − ct (w) is the net profit of performing task t for worker w and .gt  (w) = 0 if worker w is currently unmatched (i.e., .M(w) = t  = ∅). Given this definition, a matching .M is considered coalitionally stable if it does not contain any unhappy coalitions. However, we show that in some MCS instances it may not be possible to find a stable matching under this definition. Thus, we propose .α-stability if the matching obtained achieves not more than .α dissatisfaction for each worker. This .α-stability is studied in different scenarios (e.g., proportional rewards) and corresponding algorithms that guarantee certain .α values are provided. These algorithms are adapted from the well-known online budgeted maximum coverage (OBMC) problem [64]. In an opportunistic MCS setting, it is also possible that worker trajectories can be uncertain and hence not known in advance. Thus, existing solutions fail to produce an effective task assignment. Moreover, the uncertainty in worker trajectories requires a different stability definition. In [65], we study this problem with the following unhappy pair definition. Definition 4 (Decision–time Unhappy Pair) A worker–task pair .(wi , tj ) is said to be a decision–time unhappy pair if the following conditions hold for any time-step s in .[tj .b, tj .d] (i.e., beginning and deadline of task j ): • • • •

Worker .wi has a positive remaining capacity. Task .tj is unassigned. Worker .wi is in region .tj .r. Either (i) SP matches worker .wi to task .tj , but at least one of them would be better off otherwise, i.e., .

Wi (s) > Wi,j (s) or Tj (s) > Tj,i (s)

(6)

• Or (ii) SP does not match worker .wi to task .tj , but they both would be better off otherwise, i.e., .

Wi,j (s) > Wi (s) and Tj,i (s) > Tj (s).

(7)

Here, .Wi (s) and .Wi,j (s) refer to the expected total reward worker .wi would get in time frame .[s, T ] if he was not assigned to task .tj at time-step s, and otherwise, respectively. Similarly, .Tj (s) and .Tj,i (s) refer to the expected sensing quality to be

Stable Worker–Task Assignment in Mobile Crowdsensing Applications

155

received by task .tj if it is not assigned to worker .wi at time-step s and otherwise, respectively. Thus, a matching .M is called an online stable matching if it does not admit any decision–time unhappy pairs. While it is straightforward to see that the optimal matching strategy toward such stable matching would match a worker–task pair if (7) holds, the challenge comes from the computation of the values of .Wi (s),  (s), and .T (s) because .A (which is defined as the set of all possible worker .W j s i,j visit scenarios for the time frame .[s, T ] given the visit probabilities of the workers for all task regions [65]) grows exponentially with the number of users and the length of the assignment period (T ). In [65], we compute these values efficiently without actually forming the set .As and develop an efficient solution that always finds online stable matchings.

4.3 Hybrid MCS Besides a purely participatory or a purely opportunistic MCS system, it is also possible to have a hybrid MCS system to take advantage of both systems while avoiding the issues in each. The key issue in the participatory mode is that the paths assigned to workers are likely to disturb their daily schedules and introduce significant additional travel costs, whereas the opportunistic mode mainly suffers from the issue of poor coverage, as a task cannot be carried out if its region will not be visited in time by any worker in the system during their self-defined trips. A hybrid (or semi-opportunistic) sensing mode can address these issues and finds a middle ground between the participatory and opportunistic modes [28]. In the hybrid mode, the workers provide the matching platform with alternative paths (e.g., dashed lines in Fig. 1) they would be willing to take within their comfort zones in addition to the path they would normally take. This yields a wider range of task assignment options for both workers and tasks and hence not only improves the task coverage but also expands the set of tasks that workers can carry out, allowing them to increase their profits by performing more tasks. However, existing studies [27, 28] do not consider the stability in the assignment based on the preferences of the workers and task requesters; thus, the resulting assignment may impair their longterm participation in the MCS campaign. The three-dimensional version of stable matching was indeed introduced by Knuth [66] by considering three sets of agents (e.g., woman, man, dogs) and their preferences on the others. Moreover, several variants that consider cyclic preference relations [67] as well as one-dimensional preference lists over all individuals from the other two sets [47] have also been studied. Such three-dimensional stable matching has also been considered in several other application domains such as server-data source–user matching [50] in video streaming services. In some recent studies, three-dimensional stability is also considered in spatialcrowdsourcing context. However, these studies have a limited understanding of user preferences and stability. For example, [68] only considers the preferences of users

156

F. Yucel et al.

on the potential places (that the tasks will be completed) based on their proximity, while workers and task requesters do not have preferences over each other. On the other hand, there are also studies [69, 70] that consider trichromatic matching (i.e., matching of three items such as tasks, workers, and workplaces/PoIs) with some stability definitions. However, these studies mainly focus on task scheduling within a deadline without considering the matching stability based on user preferences and aim to maximize the number of matched items. In this section, we present a totally different scenario that is studied in [4] where only the nodes in two (i.e., workers, tasks) of the three sets have preferences over each other depending on the features of the nodes in the third set (i.e., acceptable paths of workers). Each worker .wi provides the service provider (SP) with a set of paths .Pi = {pi,1 , pi,2 , . . . , pi,ai } that he finds acceptable from his current location to his destination. In each assignment period, it is the responsibility of SP to find a satisfactory assignment between workers and tasks by matching workers to one of their acceptable paths and assigning a subset of tasks on their selected paths. Each path .pi,j has a capacity .ci,j associated with it, which indicates the maximum number of tasks that worker .wi is willing to perform if he is assigned to path .pi,j . Given these constraints, then we define the following to base our stable solution on. Definition 5 (Unhappy Triad) Given a matching .M, worker .wi , path .pi,j , and a set S of tasks form an unhappy triad denoted by . wi , pi,j , S if • S is an acceptable assignment for .wi , i.e., 1 ≤ |S| ≤ ci,j , S ⊆ Lw i , and S ⊆ Ti,j .

.

(8)

• .wi is an acceptable assignment for each .tk ∈ S, i.e., wi ∈ Ltk and tk ∈ Ti,j .

.

(9)

• Each task .tk ∈ S either prefers worker .wi to their current assignment .wh in .M, i.e., qi,k > qh,k where qh,k = 0 if wh = ∅,

.

(10)

or is already assigned to worker .wi , i.e., .Mu (tk ) = wi . • Worker .wi prefers the task set S to his current assignment in .M, i.e.,  .

th ∈S

rh,i >



rk,i .

(11)

tk ∈Mu (wi )

Thus, given an unhappy triad . wi , pi,j , S , we see from the first two conditions that it is possible to assign the tasks in the set S to worker .wi through path .pi,j without violating any feasibility constraints and see from the last two conditions

Stable Worker–Task Assignment in Mobile Crowdsensing Applications

157

that this would make at least one task in S and worker .wi strictly better off without making any task in S worse off. A matching is said to be 3D-stable if it does not contain any unhappy triads. In [4], we provide two different algorithms for different MCS systems. In the first algorithm, we provide a solution that always finds the stable matching in uniform MCS systems. In the second algorithm, we consider the general MCS instances where stable matchings may not exist and propose an approximation algorithm that finds near-optimal matchings in terms of stability.

5 Conclusion and Open Problems In this chapter, we studied the worker–task assignment problem in MCS systems while considering their preferences or aiming to obtain a stable matching solution. Due to the various kinds of MCS scenarios possible and many parameters (e.g., budget of task requesters, capacity of workers) that can be considered in their design, obtaining a stable matching is very challenging, and existing solutions cannot be applied directly. Thus, we provided an overview of the recent studies that considered stability in their design while assigning the tasks in the system to the eligible workers. Considering the three main categories of MCS systems, namely participatory, opportunistic, and hybrid, we have provided the core blocking or unhappy pair definitions considered to define the stability in each scenario and discussed in what conditions the proposed algorithms can find exact stable or approximate stable solutions. Besides the studies considered in this chapter, as the stability has only been studied in a few recent studies in MCS literature, there are many potential interesting problems that stability can be studied during the task assignment process. For example, in MCS systems that are defined within a mobile social network (MSN) using the local communication technologies (e.g., device to device (D2D), Bluetooth, and Wi-Fi) between the users, the decision of online task assignments while considering the stability of the decisions made is a challenging and not studied problem. Note that in such MCS systems, not only the task assignments happen in a distributed fashion and when the workers and task requesters meet each other, but also the delivery of tasks completed happens opportunistically and when the same parties meet each other again. Thus, two problems should be considered together. While the short-distance communication helps reduce the overhead on cellular networks and allows for local user recruitment and sensed data collection even if the cellular network coverage is poor [19, 71], the uncertainty increases in the system, making the stability management much harder (which can get more challenging with correlation among the mobility patterns of users [72]). Having budget constraints of task requesters and capacity constraints of workers make the problem even further challenging. Some recent studies have looked at the task assignment and worker recruitment problem [73–76]) in such MSN-based opportunistic crowdsensing scenarios. While these studies leverage the opportunistic encounters of nodes for

158

F. Yucel et al.

task assignment and communication between nodes, they do not consider stability in the assignment. Thus, this problem is still an open problem. Another key aspect that has been overlooked so far is the benefit of cooperation between workers. In MCS systems with non-trivial tasks, it may be the case that two workers who cannot carry out a certain task individually can do so if they are both assigned to the task and work in a cooperative manner. Therefore, their total utility for the task would be larger than the sum of their individual utilities. Additional costs, however, may need to be incurred to make them work cooperatively, which need to be considered in the assignment process, along with the potential benefits to be reaped. This is similar to the assignment problem with non-additive utility functions studied in [63], but a major difference is that the total (coverage-based) utility of two (or more) workers for a task in the model considered there cannot be larger than the sum of their individual utilities. In this chapter, we assumed a system model with rational and reliable participants. However, there may be, for example, workers who are trying to spread misinformation by submitting fabricated data. When the possibility of having such malicious users are taken into consideration, user preferences may become uncertain. We have also assumed that the sets of workers and tasks were known to the matching platform before the sensing campaign actually starts. Yet for many real-world applications, a more realistic model would allow users to join and leave the system and allow task requesters to publish new tasks and withdraw some of their existing tasks in real time during the campaign. Lastly, it is possible to improve the long-term efficiency of the proposed algorithms by forming the task assignments for an assignment period by modifying the assignments in the previous task assignment period(s) instead of creating a new task assignment from scratch in each assignment period. This has the potential to largely reduce the total running time of the proposed algorithms, especially in MCS applications, where user preferences do not change significantly between consecutive assignment periods. Thus, this is another interesting problem that can be explored with the stability in mind.

References 1. J. Wang, J. Tang, G. Xue, D. Yang, Towards energy-efficient task scheduling on smartphones in mobile crowd sensing systems. Comput. Netw. 115, 100–109 (2017) 2. T. Hu, M. Xiao, C. Hu, G. Gao, B. Wang, A QoS-sensitive task assignment algorithm for mobile crowdsensing. Pervasive Mob. Comput. 41, 333–342 (2017) 3. L. Tesfatsion, in Gale-Shapley Matching in an Evolutionary Trade Network Game (Iowa State University, Ames, 1998) 4. F. Yucel, E. Bulut, Three-dimensional stable task assignment in semi-opportunistic mobile crowdsensing, in IEEE 23rd International Symposium on a World of Wireless, Mobile and Multimedia Networks (WoWMoM) (2022) 5. L. Kazemi, C. Shahabi, Geocrowd: enabling query answering with spatial crowdsourcing, in Proceedings of the 20th International Conference on Advances in Geographic Information Systems (2012), pp. 189–198

Stable Worker–Task Assignment in Mobile Crowdsensing Applications

159

6. M. Abououf, R. Mizouni, S. Singh, H. Otrok, A. Ouali, Multi-worker multi-task selection framework in mobile crowd sourcing. J. Netw. Comput. Appl. 130, 52–62 (2019) 7. T. Liu, Y. Zhu, L. Huang, TGBA: a two-phase group buying based auction mechanism for recruiting workers in mobile crowd sensing. Comput. Netw. 149, 56–75 (2019) 8. W. Gong, B. Zhang, C. Li, Location-based online task assignment and path planning for mobile crowdsensing. IEEE Trans. Veh. Technol. 68, 1772–1783 (2018) 9. Z. Zheng, F.Wu, X. Gao, H. Zhu, S. Tang, G. Chen, A budget feasible incentive mechanism for weighted coverage maximization in mobile crowdsensing. IEEE Trans. Mob. Comput. 16, 2392–2407 (2016) 10. M. Xiao, J. Wu, S. Zhang, J. Yu, Secret-sharing-based secure user recruitment protocol for mobile crowdsensing, in IEEE Conference on Computer Communications (INFOCOM) (2017), pp. 1–9 11. L. Wang, D. Yang, X. Han, T. Wang, D. Zhang, X. Ma, Location privacy-preserving task allocation for mobile crowdsensing with differential geo-obfuscation, in Proceedings of the 26th International Conference on World Wide Web (2017), pp. 627–636 12. J. Li, Z. Cai, J. Wang, M. Han, Y. Li, Truthful incentive mechanisms for geographical position conflicting mobile crowdsensing systems. IEEE Trans. Comput. Soc. Syst. 5, 324–334 (2018) 13. Y. Liu, B. Guo, Y. Wang, W. Wu, Z. Yu, D. Zhang, TaskMe: multi-task allocation in mobile crowd sensing, in Proceedings of the 2016 ACM International Joint Conference on Pervasive and Ubiquitous Computing (2016), pp. 403–414 14. X. Tao, W. Song, Task allocation for mobile crowdsensing with deep reinforcement learning, in 2020 IEEE Wireless Communications and Networking Conference (WCNC) (2020), pp. 1–7 15. B. Zhao, S. Tang, X. Liu, X. Zhang, W. Chen, iTAM: bilateral privacy-preserving task assignment for mobile crowdsensing. IEEE Trans. Mob. Comput. 20 , 3351–3366 (2020) 16. Y. Zhao, K. Zheng, Y. Li, H. Su, J. Liu, X. Zhou, Destination-aware task assignment in spatial crowdsourcing: a worker decomposition approach. IEEE Trans. Knowl. Data Eng. 32, 2336– 2350 (2019) 17. M. Zhang, P. Yang, C. Tian, S. Tang, X. Gao, B. Wang, F. Xiao, Quality-aware sensing coverage in budget-constrained mobile crowdsensing networks. IEEE Trans. Veh. Technol. 65, 7698– 7707 (2015) 18. J. Chen, J. Yang, Maximizing coverage quality with budget constrained in mobile crowdsensing network for environmental monitoring applications. Sensors 19, 2399 (2019a) 19. M. Karaliopoulos, O. Telelis, I. Koutsopoulos, User recruitment for mobile crowdsensing over opportunistic networks, in 2015 IEEE Conference on Computer Communications (INFOCOM) (2015), pp. 2254–2262 20. Y. Zhan, Y. Xia, Y. Liu, F. Li, Y. Wang, Incentive-aware time-sensitive data collection in mobile opportunistic crowdsensing. IEEE Trans. Veh. Technol. 66, 7849–7861 (2017) 21. Y. Zhan, Y. Xia, J. Zhang, Y. Wang, Incentive mechanism design in mobile opportunistic data collection with time sensitivity. IEEE Internet Things J. 5, 246–256 (2017) 22. Z. He, J. Cao, X. Liu, High quality participant recruitment in vehicle-based crowdsourcing using predictable mobility, in IEEE Conference on Computer Communications (INFOCOM) (2015), pp. 2542–2550 23. M. Xiao, J. Wu, L. Huang, R. Cheng, Y. Wang, Online task assignment for crowdsensing in predictable mobile social networks. IEEE Trans. Mob. Comput. 16, 2306–2320 (2016) 24. J. Wu, M. Xiao, L. Huang, Homing spread: community home-based multi-copy routing in mobile social networks, in IEEE Conference on Computer Communications (INFOCOM) (2013), pp. 2319–2327 25. A. Dhungana, E. Bulut, Energy sharing based content delivery in mobile social networks, in 20th IEEE International Symposium on “A World of Wireless, Mobile and Multimedia Networks”(WoWMoM) (2019), pp. 1–9 26. Y. Li, D. Jin, P. Hui, S. Chen, Contact-aware data replication in roadside unit aided vehicular delay tolerant networks. IEEE Trans. Mob. Comput. 15, 306–321 (2015) 27. J. Wang, F. Wang, Y. Wang, L. Wang, Z. Qiu, D. Zhang, B. Guo, Q. Lv, HyTasker: hybrid task allocation in mobile crowd sensing. IEEE Trans. Mob. Comput. 19, 598–611 (2019)

160

F. Yucel et al.

28. W. Gong, B. Zhang, C. Li, Z. Yao, Task allocation in semi-opportunistic mobile crowdsensing: paradigm and algorithms. Mob. Netw. Appl. 25, 772–782 (2020) 29. D. Gale, L.S. Shapley, College admissions and the stability of marriage. Am. Math. Mon. 69(1), 9–15 (1962) 30. National Resident Matching Program (2021). https://www.nrmp.org/matching-algorithm/ 31. Z. Zhou, C. Gao, C. Xu, T. Chen, D. Zhang, S. Mumtaz, Energy-efficient stable matching for resource allocation in energy harvesting-based device-to-device communications. IEEE Access. 5, 15184–15196 (2017) 32. T. Wang, F. Liu, J. Guo, H. Xu, Dynamic SDN controller assignment in data center networks: stable matching with transfers, in The 35th Annual IEEE International Conference on Computer Communications (INFOCOM) (2016), pp. 1–9 33. R. Zhang, X. Cheng, L. Yang, Flexible energy management protocol for cooperative EV-to-EV charging. IEEE Trans. Intell. Transp. Syst. 20, 172–184 (2018) 34. F. Yucel, K. Akkaya, E. Bulut, Efficient and privacy preserving supplier matching for electric vehicle charging. Ad Hoc Netw. 90, 101730 (2019) 35. E. Bulut, M. Kisacikoglu, K. Akkaya, Spatio-temporal non-intrusive direct V2V charge sharing coordination. IEEE Trans. Veh. Technol. 68, 9385–9398 (2019) 36. E. Bulut, B. Szymanski, Mobile energy sharing through power buddies, in IEEE Wireless Communications and Networking Conference (WCNC), San Francisco, CA, USA, March 19– 22, 2017 (2017). pp. 1–6 37. A. Dhungana, T. Arodz, E. Bulut, Exploiting peer-to-peer wireless energy sharing for mobile charging relief. Ad Hoc Netw. 91, 01882 (2019) 38. E. Bulut, S. Hernandez, A. Dhungana, B. Szymanski, Is crowdcharging possible?, inIEEE 27th International conference on Computer Communication and Networks, ICCCN 2018, Hangzhou, China, July 30 - August 2, 2018 (2018), pp. 1–9 39. A. Ismaili, N. Hamada, Y. Zhang, T. Suzuki, M. Yokoo, Weighted matching markets with budget constraints. J. Artif. Intell. Res. 65, 393–421 (2019) 40. Y. Kawase, A. Iwasaki, Approximately stable matchings with budget constraints, in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 32 (2018) 41. B. Zhao, P. Xu, Y. Shi, Y. Tong, Z. Zhou, Y. Zeng, Preference-aware task assignment in ondemand taxi dispatching: an online stable matching approach, in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 33 (2019), pp. 2245–2252 42. H. Zheng, J. Wu, Online to offline business: urban taxi dispatching with passenger-driver matching stability, in 2017 IEEE 37th International Conference on Distributed Computing Systems (ICDCS) (2017), pp. 816–825 43. H. Aziz, P. Biró, S. Gaspers, R. Haan, N. Mattei, B. Rastegari, Stable matching with uncertain linear preferences. Algorithmica. 82, 1410–1433 (2020) 44. R. Bredereck, J. Chen, D. Knop, J. Luo, R. Niedermeier, Adapting stable matchings to evolving preferences, in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34 (2020), 1830–1837 45. K. Eriksson, J. Sjöstrand, P. Strimling, Optimal expected rank in a two-sided secretary problem. Oper. Res. 55, 921–931 (2007) 46. D. Knuth, N. De Bruijn, in Stable Marriage and Its Relation to Other Combinatorial problems: An Introduction to the Mathematical Analysis of Algorithms. (American Mathematical Society, Providence, 1997) 47. J. Wu, Stable matching beyond bipartite graphs, in 2016 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW) (2016), pp. 480–488 48. K. Iwama, S. Miyazaki, K. Okamoto, Stable roommates problem with triple rooms, in Proceeding of the 10th KOREA–JAPAN Joint Workshop on Algorithms and Computation (WAAC) (2007), pp. 105–112 49. R. Irving, An efficient algorithm for the “stable roommates” problem. J. Algorithms. 6, 577– 595 (1985) 50. L. Cui, W. Jia, Cyclic stable matching for three-sided networking services. Comput. Netw. 57, 351–363 (2013)

Stable Worker–Task Assignment in Mobile Crowdsensing Applications

161

51. C. Dai, X. Wang, K. Liu, D. Qi, W. Lin, P. Zhou, Stable task assignment for mobile crowdsensing with budget constraint. IEEE Trans. Mob. Comput. 20, 3439–3452 (2020) 52. G. Yang, B. Wang, X. He, J. Wang, H. Pervaiz, Competition-congestion-aware stable workertask matching in mobile crowd sensing. IEEE Trans. Netw. Serv. Manag. 18, 3719–3732 (2021) 53. M. Abououf, S. Singh, H. Otrok, R. Mizouni, A. Ouali, Gale-shapley matching game selection—A framework for user satisfaction. IEEE Access. 7, 3694–3703 (2018) 54. X. Yin, Y. Chen, C. Xu, S. Yu, B. Li, Matchmaker: stable task assignment with bounded constraints for crowdsourcing platforms. IEEE Internet Things J. 8, 1599–1610 (2020) 55. Y. Chen, X. Yin, Stable job assignment for crowdsourcing, in IEEE Global Communications Conference (GLOBECOM) (2017), pp. 1–6 56. B. Holzbauer, B. Szymanski, E. Bulut, Socially-aware market mechanism for participatory sensing, in Proceedings of ACM International Workshop on Mission-Oriented Wireless Sensor Networking (MiSeNet) at Mobicom 2012, Istanbul, Turkey, August 26, 2012 (2012), pp. 9–14 57. F. Yucel, E. Bulut, User satisfaction aware maximum utility task assignment in mobile crowdsensing. Comput. Netw. 172, 107156 (2020) 58. F. Yucel, M. Yuksel, E. Bulut, QoS-based budget constrained stable task assignment in mobile crowdsensing. IEEE Trans. Mob. Comput. 20, 3194–3210 (2020) 59. Y. Kawase, A. Iwasaki, Approximately stable matchings with budget constraints, in ThirtySecond AAAI Conference on Artificial Intelligence (2018), pp. 1113–1120 60. F. Yucel, E. Bulut, Time-dependent stable task assignment in participatory mobile crowdsensing, in IEEE 45th Conference on Local Computer Networks (LCN) (2020), pp. 433–436 61. M. Zhang, P. Yang, C. Tian, S. Tang, X. Gao, B. Wang, F. Xiao, Quality-aware sensing coverage in budget-constrained mobile crowdsensing networks. IEEE Trans. Veh. Technol. 65, 7698– 7707 (2016) 62. Z. Feng, Y. Zhu, Q. Zhang, L. Ni, A. Vasilakos, TRAC: truthful auction for locationaware collaborative sensing in mobile crowdsourcing, in IEEE Conference on Computer Communications (INFOCOM), Toronto, Canada (2014), pp. 1231–1239 63. F. Yucel, M. Yuksel, E. Bulut, Coverage-aware stable task assignment in opportunistic mobile crowdsensing. IEEE Trans. Veh. Technol. 70, 3831–3845 (2021) 64. D. Rawitz, A. Rosén, Online budgeted maximum coverage, in 24th Annual European Symposium on Algorithms (ESA) (2016) 65. F. Yucel, E. Bulut, Online stable task assignment in opportunistic mobile crowdsensing with uncertain trajectories. IEEE Internet Things J. 9, 9086–9101 (2021) 66. D. Knuth, in Mariages Stables (Les Presses De L’Universit De Montral, Montréal, 1976) 67. C. Ng, D. Hirschberg, Three-dimensional stable matching problems. SIAM J. Discret. Math. 4, 245–252 (1991) 68. B. Li, Y. Cheng, Y. Yuan, G. Wang, L. Chen, Three-dimensional stable matching problem for spatial crowdsourcing platforms, in 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (2019), pp. 1643–1653 69. T. Song, Y. Tong, L. Wang, J. She, B. Yao, L. Chen, K. Xu, Trichromatic online matching in real-time spatial crowdsourcing, in IEEE 33rd International Conference on Data Engineering (ICDE) (2017), pp. 1009–1020 70. B. Zheng, C. Huang, C. Jensen, L. Chen, N. Hung, G. Liu, G. Li, K. Zheng, Online trichromatic pickup and delivery scheduling in spatial crowdsourcing, in IEEE 36th International Conference on Data Engineering (ICDE) (2020), pp. 973–984 71. Y. Wang, H. Li, T. Li, Participant selection for data collection through device-to-device communications in mobile sensing. Pers. Ubiquit. Comput. 21, 31–41 (2017) 72. E. Bulut, S. Geyik, B. Szymanski, Utilizing correlated node mobility for efficient DTN routing. Pervasive Mob. Comput. 13, 150–163 (2014) 73. M. Xiao, J. Wu, L. Huang, Y. Wang, C. Liu, Multi-task assignment for crowdsensing in mobile social networks, in IEEE Conference on Computer Communications (INFOCOM) (2015), pp. 2227–2235 74. M. Xiao, J. Wu, L. Huang, R. Cheng, Y. Wang, Online task assignment for crowdsensing in predictable mobile social networks. IEEE Trans. Mob. Comput. 16, 2306–2320 (2017)

162

F. Yucel et al.

75. Y. Han, H. Wu, Minimum-cost crowdsourcing with coverage guarantee in mobile opportunistic D2D networks. IEEE Trans. Mob. Comput. 16, 2806–2818 (2017) 76. F. Yucel, E. Bulut, Location-dependent task assignment for opportunistic mobile crowdsensing, in IEEE 17th Annual Consumer Communications & Networking Conference (CCNC) (2020), pp. 1–6

Spatiotemporal Task Allocation in Mobile Crowdsensing Honglong Chen, Guoqi Ma, and Yang Huang

1 Introduction With the rapid development of technology, mobile devices equipped with rich embedded sensors (e.g., cameras, microphones, GPS, compass, etc.) are increasingly used to collect information about the surrounding environments and human activities, which makes mobile crowdsensing (MCS) [1] become a new paradigm to collect large sensing information by using crowd power. Nowadays, MCS has received extensive attention and has been widely applied in various fields, such as positioning and navigation [2, 3], environmental monitoring [4], city perception [5– 7], public safety [8], social recommendation [9], etc. Recently, more and more studies have focused on the problem of spatiotemporal task allocation. Wang et al. [10] considered the spatial constraints and the valid duration of each task and proposed effective heuristic methods to maximize the task coverage and minimize incentive cost. Li et al. [11] proposed a multi-task allocation problem with time constraints, which investigates the impact of time constraints on multi-task allocation and aims to maximize the utility of the MCS platform. Xu et al. [12] presented a universal system model based on reverse auction framework and designed two incentive mechanisms to stimulate more potential users. Cheng et al. [13] formalized a new optimization problem under a traveling budget constraint, namely maximum quality task assignment (MQA), and proposed efficient heuristics to solve it. In addition, some studies focus on location-related tasks. For location-dependent mobile crowdsensing systems, Jiang et al. [14] designed a user recruitment algorithm based on reinforcement learning, which achieved accurate and efficient task allocation while maintaining privacy.

H. Chen () · G. Ma · Y. Huang College of Control Science and Engineering, China University of Petroleum (East China), Qingdao, China e-mail: [email protected]; [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 J. Wu, E. Wang (eds.), Mobile Crowdsourcing, Wireless Networks, https://doi.org/10.1007/978-3-031-32397-3_7

163

164

H. Chen et al.

Wang et al. [15] proposed a demand-based dynamic incentive mechanism that dynamically changes the rewards of sensing tasks at each sensing round in an ondemand way to balance their popularity. Considering the spatiotemporal correlation of heterogeneous tasks, Wang et al. [16] devised a decomposition-and-combination framework to maximize data quality and minimize total incentive budget. Zhang et al. [17] proposed a new multi-task allocation method based on mobility prediction, which jointly considered workers’ and tasks’ spatiotemporal features. In addition, Tao et al. [18] designed a novel privacy mechanism based on hierarchically well-separated trees (HSTs) and investigated privacy protection for online task assignment. However, the above-mentioned studies on task allocation methods with spatiotemporal characteristics either only focus on a single and homogeneous task or do not consider the duration of the user’s task execution. Therefore, we propose two corresponding sensing scenarios and design effective task allocation algorithms respectively. In this chapter, we consider two sensing scenarios. One is that the sensing tasks are not only located in a specific spatial region, but also require mobile users to continuously sense for a sufficient time duration to obtain effective sensing data. In this scenario, the platform divides the large-scale sensing area into multiple subareas, each of which is called an area of interest (AoI). And then, we refine each sensing task’s working time for each mobile user and propose an efficient task allocation algorithm to assign a suitable AoI to each user. In addition, we also consider another scenario in which heterogeneous tasks and users coexist (HTUC). Different tasks have different sensor requirements and spatiotemporal characteristics, and mobile users have different task execution abilities due to different personal preferences and mobile devices. In the HTUC scenario, we design the corresponding user recruitment algorithm, the main idea of which is to make multiple sensing tasks completed by a small number of users.

2 Optimized Allocation of Time-Dependent Tasks for Mobile Crowdsensing In this section, we focus on the task allocation in time-dependent crowdsensing systems and formulate the time-dependent task allocation problem, in which both the sensing duration and the user’s sensing capacity are considered. Furthermore, we propose an efficient task allocation algorithm called OPtimized Allocation scheme of Time-dependent tasks (OPAT), and the extensive simulations are conducted to demonstrate the effectiveness of the proposed scheme.

Spatiotemporal Task Allocation in Mobile Crowdsensing

165

2.1 Problem Statement In this chapter, we mainly focus on the spatiotemporal mobile crowdsensing system, in which the sensing tasks are not only located in a specific spatial region, but also require mobile users to continuously sense for a sufficient time duration to obtain effective sensing data. For example, the crowdsensing platform recruits mobile users to provide traffic monitoring video at the designated intersection and requires that the duration of the video should be sufficient to calculate and predict traffic flow. For noise pollution assessment, mobile users need to submit audio of specific duration at various locations in the city. Based on these data, large-scale and fine-grained noise measurement can be realized. In addition, the time availability of mobile users also needs to be considered in spatiotemporal crowdsensing. Generally, mobile users are with limited time budgets since they are usually part-time employees to perform the sensing tasks in their spare time to earn extra rewards. As different mobile users may have different time budgets, their sensing capacities are also different. Take Fig. 1 as an illustrating example. There are a certain number of sensing tasks distributed in an intersection area, such as traffic monitoring, noise pollution assessment and air quality detection, etc. Each user’s working time for each task needs to meet a requirement of specific time duration. Since the users may want to perform the sensing tasks within dozens of minutes of after-meal walking time or a period of time after work, they need to stay in the intersection area for a certain amount of time for the task performing. In the above spatiotemporal crowdsensing scenario, how to reasonably allocate tasks to maximize the overall benefits is a key problem to be solved.

2.2 System Overview We propose an effective task allocation algorithm, named OPAT [19], to maximize the rewards for the MCS platform. The OPAT considers a crowdsensing platform that leverages the crowd to collect massive sensing data. The publishers will publish Fig. 1 A busy intersection area with a certain number of sensing tasks

166

H. Chen et al.

Platform

1. Area Division

Undivided areas with sensing tasks

uisition 2. Information Acquisition

Divided areas with sensing tasks

AoI Match

4. Task Performing

Fig. 2 The four stages in each sensing round: area division stage, information acquisition stage, task assignment stage, and task performing stage

their sensing tasks on the platform, and the mobile users can apply for the tasks that they are interested in and get rewards after completing the tasks. The platform is responsible for assigning the tasks to appropriate mobile users. The operation of the crowdsensing platform is divided into multiple rounds. As shown in Fig. 2, each round is comprised of four stages, including area division, information acquisition, task assignment, and task performing. In the area division stage, each sensing task issued by the publisher can be divided into multiple identical subtasks by the platform. After that, the platform divides the large-scale sensing area into multiple subareas, each of which is called an area of interest (AoI). In the information acquisition stage, the basic information of new participants will be registered in preparation for task assignment as soon as they join the crowdsensing system. The participants and tasks arriving at the system after the first two stages of the current round will not be processed until task assignment stage of the next round. Then, in the task assignment stage, the crowdsensing system performs task allocation in each AoI separately based on the mobile users’ information. Finally, each participant works on its allocated tasks and uploads the sensing data in the task performing stage. The unperformed tasks in the current round will be added to the task assignment stage of the next round. The above four stages will be repeated until all the tasks have been completed or no new participant joins the system.

Spatiotemporal Task Allocation in Mobile Crowdsensing

167

In the information acquisition stage of each round, each mobile user can upload the necessary information to the platform and select its AoI according to its preference and time budget (e.g., the user is likely to choose the AoI near home or on the way to its destination). Since AoIs are far from each other, the user may not be able to move from one AoI to another within his/her limited time budget, i.e., each user has only one AoI choice. Therefore, there are three assumptions: 1. Compared with the working time for sensing tasks, the travel time among the tasks within an AoI can be negligible as the tasks are highly clustered in the area. 2. Each user selects only one AoI according to his/her own preference and uploads the time budget for task performing within the AoI. 3. Each user is rational, indicating that it will refuse to perform the tasks if the working time exceeds his/her time budget or the gained reward is lower than its cost.

2.3 Problem Formulation   OPAT concentrates on the task allocation within each AoI. Let .T = t1 , t2 , · · · , tm denote the set of sensing tasks in the AoI, where .tj denotes the j th task. Let  .U = u1 , u2 , · · · , un denote the set of mobile users who select the AoI , where .ui is the ith mobile user. In particular, the sensing tasks are time-dependent, which means that each task .tj is associated with a required sensing duration. Specifically, to effectively perform task .tj , user .ui needs to spend a certain amount of working time .W Tij (for example, recording a specific-length video). Let .Tui denote the set of tasks that the platform assigns to user .ui and .W TTui denote the total working time user .ui spends in completing .Tui . Each user .ui has a time budget to perform the sensing tasks, denoted as .Bui . According to user .ui ’s rationality, it will perform .Tui only if .W TTu ≤ Bui . i Generally, task publishers require the platform to divide a task into multiple subtasks (i.e., multiple independent measurements) before publishing it to guarantee the sensing quality, and these subtasks have the same measurement requirements and need to be completed by different mobile users. The division of sensing tasks will be carried out in the area division stage. Assume that each task .tj is divided into .bj independent subtasks with identical characteristics, where .bj is associated with the required sensing quality. That is, each user can perform at most one subtask of task .tj , and task .tj is required to be performed by .bj users. Once the tasks are completed, the platform will get revenues from the publishers and pay rewards to the users. Let .Pij = .rij − pij denote the net profit when user .ui performs task .tj , which is the difference between the revenue .rij that the platform gains from the publisher and the price .pij that the platform pays to user .ui . The crowdsensing platform aims to maximize its net profit by assigning tasks to suitable users. To this end, the time-dependent tasks allocation problem, named TDTA, can be formulated as follows:

168

H. Chen et al.

.

max

y(x) =

 

Pij xij ,

(1)

ui ∈U tj ∈T

  ⎧ Tui = tj | ϑij = 1 , ∀tj ∈ T, ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ W TTui ≤ Bui , ∀ui ∈ U, ⎪ ⎪ ⎨ n  .s.t. ⎪ ϑij ≤ bj , ∀tj ∈ T, ⎪ ⎪ ⎪ ⎪ i=1 ⎪ ⎪ ⎪   ⎩ ϑij ∈ 0, 1 , ∀ui ∈ U, tj ∈ T,

(1a) (1b) (1c) (1d)

where .ϑij is the decision variable, .ϑij = 1 indicates task .tj is assigned to user ui and .ϑij = 0 otherwise. Equation (1b) is to ensure that the total working time required for the assigned tasks will not exceed the user’s time budget. In Eq. (1c), n . i=1 ϑij ≤ bj since task .tj is divided into .bj subtasks, which require .bj users to perform. .

2.4 Task Allocation Algorithm The above task allocation problem is proved to be NP-hard, and an efficient task allocation algorithm, called OPAT, is proposed, which can take full advantage of the sensing capacity of each mobile user. OPAT consists of three steps, which are shown in Algorithm 1. Step 1 We reformulate the TDTA problem. To reduce the error and guarantee the quality of sensing, each task .tj is divided into .bj subtasks with identical characteristics. Let .tj k denote the kth subtask of the j th task, where .j = 1, 2, · · · , m, and .k = 1, 2, · · · , bj . Thus, the total number of subtasks in the system is .M = m j =1 bj . We assume that the platform will gain net profit .zij k when user .ui completes the subtask .tj k , and .zij k = Pij for all .k = 1, 2, · · · , bj . Note that each user can only perform one of the subtasks of each task. Hence, a task allocation scheme is necessary to maximize the platform’s benefits. Consequently, we can get a constrained TDTA problem named C-TDTA that is more complicated than the original TDTA problem, which can be formulated as follows: C − T DT A : max q(v) =

bj m  n  

.

i=1 j =1 k=1

zij k vij k ,

(2)

Spatiotemporal Task Allocation in Mobile Crowdsensing

169

Algorithm 1 OPtimized Allocation scheme of Time-dependent tasks (OPAT) Input: .ui ∈ U, .tj ∈ T, .Bui , .bj , .rij , .pij , .Dj , .Fj , .Qj , .Ei , .Gj . Output: .S F . 1: Step 1: Transform the TDTA problem into the C-TDTA problem; 2: Step 2: Obtain the preliminary task allocation by solving the knapsack problem of each mobile user; 3: Initially, .h = 0; 4: repeat 5: .h = h + 1; 6: Obtain .Th by selecting m subtasks from M tasks according to Eq. (4); 7: Solve the knapsack problem associated with mobile user .uh and get .T Sh ; h according to Eqs. (5) and (6); 8: Define .qh (v) and .zij k 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: 22: 23: 24: 25: 26: 27: 28:



h for next iteration; Obtain .qh (v) and .zij k until .h = n or .qh (v)=0. Step 3: Conflict elimination and task reallocation; Strategy I: for .h = 1 to n do Redetermine .Th by removing and adding; Solve the knapsack problem of user .uh with .TIh ; Update .T Sh ; .h = h + 1; end for The output of Strategy I: .S I Strategy II: for .h = n to 1 do Redetermine .Th by removing and adding; Solve the knapsack problem of user .uh with .TIhI ; Update .T Sh ; .h = h − 1; end for The output of Strategy II: .S I I Compare the profits of .S I and .S I I and return the allocation with more profits.

  ⎧  Tui = tj k | vij k = 1, j = 1, 2, · · · , m, k = 1, 2, · · · , bj , ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ W TTu ≤ Bui , ∀ui ∈ U ⎪ ⎪ i ⎪ ⎪ ⎪ ⎪ b ⎪ j n ⎪  ⎪ ⎪ ⎪ vij k ≤ bj , j = 1, 2, · · · , m, ⎪ ⎪ ⎪ ⎪ i=1 k=1 ⎪ ⎨ n .s.t.  ⎪ ⎪ vij k ≤ 1, j = 1, 2, · · · , m, k = 1, 2, · · · , bj , ⎪ ⎪ ⎪ ⎪ i=1 ⎪ ⎪ ⎪ ⎪ ⎪ bj ⎪  ⎪ ⎪ ⎪ vij k ≤ 1, j = 1, 2, · · · , m, k = 1, 2, · · · , bj , ⎪ ⎪ ⎪ ⎪ k=1 ⎪ ⎪ ⎪   ⎩ vij k ∈ 0, 1 ,

(2a) (2b) (2c)

(2d)

(2e) (2f)

170

H. Chen et al.

where .vij k is the decision variable, .vij k = 1 represents that the platform assigns task tj k to user .ui and .vij k = 0 otherwise. Equation (2d) is to ensure that subtask .tj k can only be assigned to at most one user since the subtask cannot be divided and completed repeatedly. Equation (2e) requires that user .ui can only complete one of the subtasks of task .tj , since the data validity can be guaranteed by multiple sensing data from different users.

.

Step 2 The preliminary task allocation is obtained by solving the knapsack problem of each mobile user. Next, we introduce how to solve the knapsack problem of each user through iterations. Let .qh (v) denote the reward function at the beginning h is the profit associated with .q  (v), i.e., at the beginning of of iteration h and .zij h k iteration one, the reward function is

.

q1 (v) = q(v) =

bj m  n  

zij k vij k =

i=1 j =1 k=1

bj m  n  



1 zij k vij k .

(3)

i=1 j =1 k=1

In each iteration, the reward function will be modified. Let .qh (v) denote the h is the profit associated with .q (v). modified reward function at iteration h and .zij h k In each iteration h, m subtasks are selected from the total M tasks, in which at most one subtask can be selected from its associated task. The subtask selection rule is: for each task .tj , choose subtask .tj k  , where k  = arg

.

max



k=1,2,··· ,bj

h zhj k.

(4)

Let .Th represent the set of tasks selected in iteration h. After that is to solve the knapsack problem of user .uh . Intuitively, the sensing tasks are items, and the time budget of .uh is a knapsack. The platform needs to assign the sensing tasks in .Th to user .uh to maximize profits without exceeding the capacity of the knapsack. The task set assigned to user .uh , denoted as .T Sh , can be obtained by an algorithm of knapsack problem. Next, we define a new reward function .qh (v) at iteration h as qh (v) =

bj m  n  

.

h zij k vij k ,

(5)

i=1 j =1 k=1

where

h .zij k

=



h , i = h ∨ (i > h ∧ t ∈ T S ), zhj jk h k 0, otherwise.

(6)

Then we can get the reward function at the beginning of next iteration, i.e., (h+1) h −zh . The iterative process q(h+1) (v) = qh (v)−qh (v). Accordingly, .zij = zij k k ij k

.

Spatiotemporal Task Allocation in Mobile Crowdsensing

171

terminates when .h = n or .qh (v) = 0, which indicates theknapsack problem of each user is solved and a preliminary task allocation .T S = T S1 , T S2 , · · · , T Sn is obtained. Step 3 Conflict elimination and task reallocation. Since subtask .tj k may be assigned to multiple users in Step 2, which is a conflict and is not allowed, then the conflicts have to be eliminated together with the task reallocation. The following two conflict elimination and task reallocation strategies are put forward. Strategy I Eliminating conflicts and reallocating tasks forward. For user .uh , Strategy I reselects the tasks from the total M tasks. First, redetermine the selected  task set by removing some tasks from .Th . Theremoved task .tj k satisfies .tj k ∈ T S1 , T S2 , · · · , T Sh−1 , T Sh+1 , T Sh+2 , · · · , T Sn , with which there are two scenarios, one is that task .tj k is assigned to .uh and other users, and the other is that task .tj k is assigned to other users except .uh . Then, add some tasks to .Th . The added  task .tj k satisfies .tj k ∈ / T S1 , T S2 , · · · , T Sn , which means task .tj k is not assigned to any user in Step 2. Note that the added tasks may be different subtasks of the  and same task. In view of this, we choose task .tj k with a smaller k, e.g., if task .t62   task .t63 are not selected in Step 2, only task .t62 will be added to .Th . After the above operation, a new selected task set .TIh can be obtained. Then, solve the knapsack problem of user .uh with .TIh and update the task set .T Sh assigned to user .uh . Strategy I eliminates the conflicts and reallocates sensing tasks to users from .u1 to  I I I I .un . Finally, the task allocation with Strategy I, denoted as .S = S , S , · · · , Sn , 1 2 can be obtained. Strategy II Eliminating conflicts and reallocating tasks backward. All the operations of Strategy II are the same as that of Strategy I, except Strategy II eliminates conflicts and reallocates sensing tasks to users from .un to .u1 , which is the reverse of the assignment order in Strategy I. Therefore, we skip the detailed descriptions of Strategy  allocation with Strategy II can be obtained, denoted as  II. Also, the task I I = SI I , SI I , · · · , SI I . .S n 1 2 II At the end of Step 3, the profits of .S I and to get the  .FS Fwill be compared  F allocation with more profits, denoted as .S = S1 , S2 , · · · , SnF .

2.5 Performance Evaluation The simulations are conducted between OPAT and LRBA [20], which demonstrate the advantages of OPAT in terms of the total profits, task completion rate, and average remaining time of users. Figure 3a compares the total profits of the OPAT and LRBA schemes when the number of tasks varies. Figure 3a illustrates that OPAT always achieves more profits than LRBA and the total profits of both the OPAT and LRBA increase with the number of tasks. As OPAT cannot only eliminate the conflicts, but also reallocate the previously unassigned tasks to mobile users, it can obtain more profits than

172

H. Chen et al.

(b)

(a)

Fig. 3 The comparison of OPAT and LRBA on the total profits with different numbers of tasks and users

(a)

(b)

(c)

Fig. 4 Comparison of OPAT and LRBA on the total profits, task completion rate, and average remaining time of users with different time budgets

LRBA. Figure 3b shows an upward trend when the number of users increases, the reason behind which is that with more tasks, both the OPAT and LRBA can recruit more users and assign the tasks to them to obtain more profits. Moreover, OPAT can achieve more profits than LRBA. When the number of tasks is larger than 45, the performance of OPAT is much better than that of LRBA. This is because when the number of tasks becomes larger, there will be more conflicts during the allocation and more tasks will be dropped, in which OPTA demonstrates its advantages by reallocating the dropped tasks to users. Figure 4 shows the impacts of time budget on the performance of the total profits, task completion rate, and average remaining time of users. We fix the number of tasks as .m = 65, then vary .σ from 10 to 45, and set the number of users n to 10, 15, and 20, respectively. Figure 4a illustrates that OPAT always obtains more profits than LRBA. And the total profits increase with the time budget. Figure 4b shows that the task completion rate of OPAT is always higher than that of LRBA, which is consistent with the result that OPAT achieves higher profits. Note that when n is relatively large (i.e., .n ≥ 15), the task completion rate of OPAT can reach .100%,

Spatiotemporal Task Allocation in Mobile Crowdsensing

(a)

(b)

173

(c)

Fig. 5 Comparison of OPAT and LRBA on the total profits, task completion rate, and average remaining time of users with different numbers of tasks

which indicates that all of the tasks are successfully allocated to approximately maximize the profits. Figure 4c illustrates that the average remaining time of users of OPAT is less than that of LRBA as OPAT can make full use of the sensing capacities of users via reallocation. With the increase of time budgets of users, the average remaining time of users increases quickly, since all of the tasks have already been allocated and each user does not need to use up his/her time budget. Figure 5 shows the impacts of the number of tasks on the performance of the total profits, task completion rate, and average remaining time of users. Figure 5 illustrates that the task completion rates of OPAT and LRBA approximate .100% when the number of tasks m is relatively small, in which OPAT and LRBA obtain the similar profits and each user has relatively much remaining time since he/she does not need to use up his/her time budget. The profits of OPAT and LRBA increase with the number of sensing tasks, while the former increases faster since the OPAT can make full use of the time budget of each user, especially when there are sufficient tasks. However, LRBA fails to fully utilize the users’ sensing capacities, which can be validated in Fig. 5b and c. The task completion rate of LRBA decreases significantly with the increase of the number of sensing tasks, while the average remaining time of users decreases more slowly. The reason behind is that when the number of sensing tasks increases, there are more conflicts in task allocation and LRBA cannot successfully allocate all the tasks to the users.

3 Heterogeneous User Recruitment of Multiple Spatiotemporal Tasks In this section, we first put forward the task priority and quantify the task execution ability of users in the HTUC scenario. And then, we define users’ utility function, which is used as the recruitment index of the platform. Finally, we propose the relevant user recruitment algorithms and carry out extensive experimental verification.

174

H. Chen et al.

3.1 Problem Statement As two significant elements of the MCS system, sensing tasks and mobile users have attracted more and more attention. For sensing tasks, on one hand, multiple MCS applications occurring simultaneously may have different requirements for sensors. For example, in modern cities, the residents use cameras to record road information, microphones to detect noise, and particle sensors to measure air quality. On the other hand, different sensing tasks have different spatiotemporal characteristics, and each task needs to be performed at a specific place within a specified time period. It should be noted that the tasks may have spatiotemporal correlations [16]. In other words, there may be multiple tasks that can be executed at the same time and place. Each task with different sensor requirements and spatiotemporal characteristics is called a heterogeneous task. In addition, for mobile users, due to the different mobile devices and personal preferences, the task execution ability of each mobile user is various. For multiple sensing tasks occurring simultaneously in the MCS system, some users can perform all the tasks, while others can only complete some of them. We need to consider the variety of users, which can be classified into multiple categories according to their task execution ability. Such kind of users are called heterogeneous users. Based on the above considerations, we propose a problem of Heterogeneous User Recruitment of multiple heterogeneous Tasks (HURoT) in mobile crowdsensing. In the HURoT problem, we recruit suitable users for each task by considering the users’ execution ability and the attributes of heterogeneous tasks. Obviously, the HURoT problem is a typical optimization problem, and the main goal is to maximize the task completion coverage. In addition, the platform needs to give appropriate rewards to each user who completes sensing tasks. Therefore, another optimization goal is to minimize the total platform payment. Figure 6 shows an example of the HURoT problem. There are three heterogeneous tasks: road information recording .t1 , noise detection .t2 , and air quality measurement .t3 , and note that each point marked in the figure is a separate measurement, which is called a subtask of the corresponding task. In addition, seven types of heterogeneous users are shown, and all users are evenly distributed in the entire sensing area. Note that in different time periods in Fig. 6, the tasks included in different points are different. Due to users’ different task execution ability, each point has a variety of different user recruitment options. In this section, we also made the following assumptions: 1. The sensing process is participatory sensing, that is, the recruited user must unconditionally perform the tasks assigned to him/her. 2. Each subtask requires .κ users to perform to ensure the sensing quality. 3. To avoid the monopoly phenomenon that all subtasks of a heterogeneous task are only completed by a few users, we consider that each user can complete .ξ subtasks of the same heterogeneous task at most.

Spatiotemporal Task Allocation in Mobile Crowdsensing

175

X 

S 11:00

12:00

13:00

14:00

75RDG LQIRUPDWLRQ UHFRUGLQJ

12:00

10:00 11:00 12:00 13:00 14:00

71RLVH GHWHFWLRQ

13:00

14:00

X

X



15:00 

7 $LU TXDOLW\ PHDVXUHPHQW

X

X 

+HWHURJHQHRXV XVHUV 7DVN H[HFXWLRQ DELOLW\

S

X 

X

10:00 11:00 12:00 13:00 14:00 15:00

Fig. 6 An example of the HURoT problem in HTUC scenario

Platform

,QIRUPDWLRQ SURFHVVLQJ 8VHU UHFUXLWPHQW

,QIRUPDWLRQ XSORDGLQJ 10:00 11:00 12:00 13:00 14:00 15:00

7DVN H[HFXWLRQ

11:00

12:00

13:00 14:00

10:00 11:00 12:00 13:0014:00

+HWHURJHQHRXV WDVNV +HWHURJHQHRXV XVHUV

12:00

13:00

14:00 15:00

Fig. 7 The architecture of user recruitment in mobile crowdsensing system

3.2 System Overview In modern cities, it may be expensive and time-consuming to deploy specific devices to complete the large-scale sensing tasks. As an alternative, we can use the idea of crowdsourcing that leverages the power of people to collect sensing data. In the HTUC scenario, users with heterogeneity can travel to the specified location for quick and accurate measurement. The subtasks of different tasks at the same location can all be performed by one user with the required sensors, which can improve the sensing efficiency and reduce the cost. As shown in Fig. 7, the MCS system is divided into five stages, including information uploading, information processing, user recruitment, task execution, and data collection. In the information uploading stage, the task publisher uploads

176

H. Chen et al.

the heterogeneous information of tasks to the platform, and the latter will also register the heterogeneous information of the users in the system before the user recruitment. In the stage of information processing, the platform analyzes and integrates the spatiotemporal characteristics of heterogeneous tasks. The whole sensing process will be divided into multiple sensing cycles, and the set of subtasks at different points will be determined. Then, in the user recruitment stage, the platform adopts efficient strategies to recruit users for each point by considering the global information and records the recruitment results. In the task execution stage, the user executes all tasks assigned to his/her by the platform sequentially according to the recruitment results. Finally, the platform collates the sensing data uploaded by users and submits it to the task publisher in the data collection stage. However, to solve the HURoT problem, there are at least the following challenges that must be overcome. First, the heterogeneity of tasks and users makes the sensing scenario more complicated. Second, how to represent and balance the platform payment model and task coverage ratio model in the new scenario is another challenge. Finally, it is expensive, even infeasible to retrieve an optimal solution in larger solution space, especially for the HURoT problem. Thus, it is necessary to design a suitable user recruitment scheme.

3.3 Problem Formulation Before presenting the HURoT problem in the MCS system, we first define the heterogeneous tasks, heterogeneous users, sensing cycle, and spatial point as follows. Definition 1 (Heterogeneous Tasks) Let .T = {t1 , t2 , · · · , tm } be a set of m heterogeneous tasks, which have different spatiotemporal characteristicsand requirements  for sensors. For each task, say .ti , we utilize a triple, i.e., .ti = Stti , Tti , Srti , to formalize its inherent attribute, where: (1) .Stti = ti1 , ti2 , · · · , tiλi denotes the subtask set of the heterogeneous task  .λi represents the number of subtasks  .ti , where and .tij is the j th subtask; (2) .Tti = ttis , ttie denotes .ti ’s time attribute, where .ttis is the start time and .ttie is the end time; (3) .Srti is the type of sensor required to perform task .ti , e.g., microphone, camera, etc. In addition, .Ltij is subtask .tij ’s location. Definition 2 (Heterogeneous Users) Let .U = {u1 , u2 , · · · , un } be a set of n heterogeneous users who can perform tasks in participatory MCS  system. For each  user, say .uk , we also utilize a triple, i.e., .uk = Luk , Spuk , Huk , to formalize user his/her characteristics, where: (1) .Luk denotes the initial location of user .uk ; (2) .Spuk is .uk ’s traveling speed; (3) .Huk , which is determined by .uk ’s preferences and mobile devices, is a set of .uk ’s executable tasks and reflects user .uk ’s task execution ability. Definition 3 (Sensing Cycle) Let .C = {c1 , c2 , · · · , cw } be a set of w sensing cycles and .cv is the vth cycle. Note that in the proposed HURoT problem, the

Spatiotemporal Task Allocation in Mobile Crowdsensing

177

length of each sensing cycle is not exactly the same, which depends on the states of heterogeneous tasks in different time periods. Definition 4 (Spatial Point) Let .P = {p1 , p2 , · · · , pl } be a set of l spatial points in the MCS system. Each point .pr has a specific location .Lpr . In addition, the inherent property .r of point .pr is defined, which is a set of all whose locations   subtasks  are not further than d away from point .pr , that is, .r = tij |Ltij − Lpr | ≤ d . Based on the above considerations, the HURoT problem will be formulated below. Given a set of m heterogeneous tasks .T = {t1 , t2 , · · · , tm } and a set of n heterogeneous users .U = {u1 , u2 , · · · , un }, each task .ti ∈ T consisting of a set of subtasks, i.e., .Stti = ti1 , ti2 , · · · , tiλi , has independent spatiotemporal characteristics and sensor requirements, and each user .uk ∈ U has different task execution abilities. In addition, a set of l spatial points .P = {p1 , p2 , · · · , pl } is given, and each point .pr with a fixed location .Lpr is composed of one or more subtasks, denoted as .r . The goal is to recruit a certain number of heterogeneous users for each subtask .tij at each point .pr before each task’s deadline, and minimize the total platform payment and maximize the total task coverage ratio in the sensing cycle set .C. Therefore, the proposed HURoT problem can be formulated as follows: ⎧ l  n w  ⎪  ⎪ ⎪ ⎪ min P ay(P) = pay(cv , pr , uk ), ⎪ ⎪ ⎨ v=1 r=1 k=1

.

⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ max

Cov(P) =

l  n w  

(7) cov(cv , pr , uk ),

v=1 r=1 k=1

⎧ Pvu ≤ Bvuk , ∀uk ∈ U, v = 1, 2, · · · , w, ⎪ k ⎪ ⎪ ⎪ ⎪ n w ⎪  ⎪ ⎪ v ⎪ ηkij ≡ κ, i = 1, 2, · · · , m, j = 1, 2, · · · , λi , ⎪ ⎪ ⎪ ⎨ s.t.

.

(7a) (7b)

v=1 k=1

λi ⎪ w  ⎪  ⎪ v ⎪ ⎪ ηkij ≤ ξ, k = 1, 2, · · · , n, i = 1, 2, · · · , m, ⎪ ⎪ ⎪ ⎪ v=1 j =1 ⎪ ⎪ ⎪ ⎩ v ϑki ∈ {0, 1} , ηkij ∈ {0, 1} ,

(7c) (7d)

where .pay(cv , pr , uk ) and .cov(cv , pr , uk ) denote respectively the platform payment paid for .uk and the task coverage ratio contributed by .uk after he/she has performed tasks at .pr in the cycle .cv . Equation (7a) is the time constraint, which means that in each sensing cycle .cv , user .uk ’s time cost .Pvu must not be greater k than his/her time budget .Bvuk . Equations (7b) and (7c) are the other two constraints, where the former indicates that each subtask requires .κ users to perform, and the latter indicates each user can complete .ξ subtasks of the same heterogeneous task

178

H. Chen et al.

v indicate respectively whether the user .u has the ability to at most. .ϑki and .ηkij k perform the task .ti and whether the user .uk has completed the subtask .tij in the cycle .cv .

3.4 Model Analysis In this section, we first present the heterogeneous task priority model, platform payment incentive model, and user-contributing task coverage ratio model. Then, we propose a binary quantization mechanism to quantify the levels of each heterogeneous user and each spatial point.

3.4.1

Heterogeneous Task Priority Model

Although each heterogeneous task is expected to be completed as early as possible, the urgency of different tasks is different and is influenced by many factors. For example, the shorter the duration of a task, the more urgent it is. Another example is that the more subtasks, the more urgent the task is. In addition, task publishers clearly know all the attribute information of heterogeneous tasks they publish. Therefore, the initial evaluation of the publisher is also an important factor. To balance the completion of each task [15], we put forward the concept of heterogeneous task priority that is measured by above three indicators: the duration of heterogeneous task (.xi1 ), the number of subtasks (.xi2 ), and the initial evaluation of the publisher (.xi3 ). Intuitively, the shorter duration of heterogeneous task, the more subtasks, or the higher initial evaluation of publisher, the higher priority of heterogeneous task. Specifically, the priority of task .ti can be formulated as follows: P RIi = ω1 xi1 + ω2 xi2 + ω3 xi3 ,

.

(8)

where .ω1 , .ω2 and .ω3 denote the weights to measure above three indicators and ω1 + ω2 + ω3 = 1. And the three indicators of task priority are quantified by Eq. 9.

.

xi1 = e

.

1−

|Tti | λi |Tt |max , xi2 = e λmax , xi3 ∈ [1, e].

(9)

where .|Tti |, represented as .|Tti | = ttie −ttis , is .ti ’s duration, and .|Tt |max = max(|Tti |) is the maximum duration of all heterogeneous tasks. Further, given the initial reward .Rmin of the task with the lowest priority, the initial rewards of other tasks can be defined according to tasks’ relative priorities. The initial reward of task .ti can be formulated as follows:

Spatiotemporal Task Allocation in Mobile Crowdsensing

Ri =

.

179

P RIi · Rmin , P RImin

(10)

where .P RIi is task .ti ’s priority, and .P RImin = min(P RIi ) is the lowest priority among all tasks. We consider that all subtasks of the same heterogeneous task (e.g., .tia , tib ∈ ti ) have the same initial reward, that is, .Ria = Rib = Ri .

3.4.2

Platform Payment Incentive Model

When mobile users contribute sensing data in the MCS, they will spend not only time but also physical resources to complete sensing tasks. Without an effective incentive, users may not be willing to participate in the MCS system. Therefore, we improve users’ participation motivation by paying rewards. The reward paid by the platform to the user is called the platform payment, which consists of reward payment and travel payment. The former is platform’s reward for the user who has completed tasks, while the latter is platform’s compensation for user’s travel costs. The platform payment paid for .uk at point .pr in cycle .cv will be formulated as follows: v v pay(cv , pr , uk ) = σ · rpkr + (1 − σ ) · tpkr .

.

(11)

For the reward payment, we have the following considerations. First, we adopt a common incentive mode, that is, more pay for more work. Second, the recruited users perform subtasks without extra travel cost, and what they need to do is switch different sensors built-in mobile devices in the HTUC scenario. Therefore, we introduce parameter .α and make appropriate adjustments to the reward payment paid to users. In addition, we define .Svkr as a set of executable subtasks of the recruit user .uk at point .pr and determine which one has the highest priority in .Svkr . The reward payment paid for .uk will be formulated as follows: v v rpkr = ηkij Ri + α ·



.

v ηkef Re ,

(12)

v tef ∈Skr

v

where .tij is the subtask that has the highest priority and .Skr = Svkr /tij represents the set of all subtasks except .tij in .Svkr . For the travel payment, we take the Euclidean distance between the locations of users and the spatial point as the measurement index, which is formulated as follows: v tpkr = β · Ed(L#uk , Lpr ),

.

(13)

where .Ed(L#uk , Lpr ) is the Euclidean distance between user .uk ’s current location # .Lu and the location .Lpr of the point .pr where .uk will travel to. Note that when k

180

H. Chen et al.

user .uk completes subtasks at .pr , the value of .L#uk will be updated to .Lpr and .Lpr will be updated to the location of the next point that .uk will travel to, and so on. In addition, .β is a coefficient used to balance Euclidean distance and platform travel payment.

3.4.3

User-Contributed Task Coverage Ratio Model

In the HTUC scenario, heterogeneous users can perform various subtasks. In other words, each user will make different contributions to the overall task completion coverage of the MCS system at different points or in different cycles. To maximize the overall task coverage ratio, we propose the user-contributing task coverage ratio model, which is formalized as follows: cor(cv , pr , uk ) =

.

κ·

|Sv | kr m

i=1 λi

,

(14)

where .|Svkr | is the length of .Svkr , indicating the amount of sensing data uploaded by user .uk at point .pr in cycle .cv . .λi is the number of subtasks of task .ti and .κ indicates that each subtask requires .κ users to complete. Generally speaking, .cor(cv , pr , uk ) represents the proportion of sensing data uploaded by user .uk to the total data required by the system.

3.4.4

Binary-Based Representation of Level

Previously, we have shown the task execution ability of different users and the attribute characteristics of points. To facilitate the experiment, we propose a binary quantization mechanism to express the levels of different heterogeneous users and spatial points. The detailed description is as follows. Binary Quantization Mechanism For each user in the MCS system, each heterogeneous task has only two execution states, namely executable and non-executable. Similarly, for each spatial point, there are only two relationships between each subtask and the point, that is, included and not included. In both cases, the radix is set to 2, and the basic operators are 0 and 1. Furthermore, according to the binary rule, the bit weight of each heterogeneous task will be expressed, following the idea that the higher the priority, the greater the bit weight. Finally, the level is expressed by the weighted sum of the bit weights of all tasks and the related basic operators. To sum up, the level of each user (e.g., .uk ) and point (e.g., .pr ) can be respectively formulated as follows: Levuk =

m 

.

i=1

ϑki Bwi ,

(15)

Spatiotemporal Task Allocation in Mobile Crowdsensing

Levpr =

λi m  

.

φrij Bwi ,

181

(16)

i=1 j =1

where .Bwi is task .ti ’s bit weight. .φrij is a decision variable, where .φrij = 1 denotes that the subtask .tij belongs to .r , and .φrij = 0 otherwise.

3.5 HURoT Problem-Solving Approaches In this section, we first define the utility function with dual objectives. Then we propose three greedy-based user recruitment algorithms to obtain near-optimal solutions.

3.5.1

Utility Function with Dual Objectives

In the HURoT problem, minimizing platform payment and maximizing task coverage ratio are two contradictory issues. To balance these two objectives, a common idea is to transform them into a comprehensive utility function to jointly optimize. Here the utility function of each user is formulated as follows: v Ukr =

.

cov(cv , pr , uk ) . pay(cv , pr , uk )

(17)

For each user who performs multiple subtasks, we can see that the smaller the platform payment for each sensing data, the higher the utility value of the user.

3.5.2

Utility-Based User Recruitment (UURe)

The basic idea of the UURe algorithm is that the whole sensing process is divided into multiple sensing cycles, and for each point in each sensing cycle, the utility value of each user with executable status is calculated and the user with the greatest utility value will always be recruited. The workflow of UURe will be described in detail as follows. ∗ First, at the beginning of each sensing cycle .cv ∈ C, the candidate  point set .P ∗ of points with executable subtasks will be initialized, i.e., .P = pr |Levpr = 0 . Second, for each point .pr ∈ P∗ , the following recruitment process is iterated. In each iteration, according to .Levpr , the candidate users who can perform at least one subtask at .pr are first picked out and added to the candidate user set .U∗r . And then, the utility value of each user .uk ∈ U∗r is calculated and ranked, in which the user with the highest utility value is recognized as the final recruited user,  and the recruitment result is recorded in .ReListrv . Later, .Contri {uk }, .Need tij , .Levpr ,

182

H. Chen et al.

Algorithm 2 Utility-based User Recruitment (UURe) Input: ti ∈ T, uk ∈ U, pr ∈ P, tij ∈ ti , Tti , Ltij , Huk , Luk , Lpr . Output: ReListPC : Recruited user list. 1: Determine sensing cycle C and the set  of each point;  2: Initially, ReListrv = ∅, Contri {uk } = [0]1×m , N eed tij = κ; 3: for each sensing cycle cv in C do 4: Extract candidate point set P∗ ; 5: for each point pr in P∗ do 6: Extract candidate user set U∗r according to Levpr ; v of each user u ∈ U∗ ; 7: Calculate the utility Ukr k r v v 8: Rankr : rank users in U∗r according to utility Ukr from high to low; 9: s = 0; 10: repeat 11: s = s + 1; 12: Calculate the traveling time  and the time budget B of the user Rankrv [s]; 13: if   B then 14: Rankrv [s] → recruitment list ReListrv ; 15: Update Contri, N eed, L# , B and Levpr ; 16: end if

17: until Levpr = 0 or s = length Rankrv ; 18: end for 19: end for

and other related parameters are updated, and the next iteration of .pr will be carried out until .Levpr = 0 or no candidate users can be recruited. Finally, the abovementioned recruitment process is repeated for each point in each cycle until the end of all sensing cycles. The pseudocode of UURe is shown in Algorithm 2.

3.5.3

Level-First and Utility-Based User Recruitment (L-UURe)

If the amount of sensing data uploaded by a user when performing a heterogeneous task reaches the limit .ξ , we consider that the user no longer has the ability to perform this task. Therefore, if some relatively low-level points (e.g., .Levpr = 3) recruit users earlier than higher-level points (e.g., .Levpr = 7) and always recruit higherlevel users (e.g., .Levuk = 7), there may not be enough high-level users available when high-level points recruit users. The basic idea of L-UURe is that the user recruitment process is carried out according to the level of the point from high to low. The workflow of L-UURe will be described as follows. Similarly, the whole recruitment process is divided into w sensing cycles. For each sensing cycle .cv ∈ C, the candidate point set .P∗ will be initialized first, and then the following recruitment process will be carried out until the end of the cycle. First, the level .Levpr will be calculated for each point .pr ∈ P∗ , and .P∗high is defined as the set of points with the highest level in .P∗ . Second, similar to the UURe algorithm, for each point in .P∗high , all candidate users are added to .U∗r , and the

Spatiotemporal Task Allocation in Mobile Crowdsensing

183

Algorithm 3 Level-first and Utility-based User Recruitment (L-UURe) Input: ti ∈ T, uk ∈ U, pr ∈ P, tij ∈ ti , Tti , Ltij , Huk , Luk , Lpr . Output: ReListPC : Recruited user list. 1: Determine sensing cycle C and the set  of each point;  2: Initially, ReListrv = ∅, Contri {uk } = [0]1×m , N eed tij = κ; 3: for each sensing cycle cv in C do 4: Extract candidate point set P∗ ; 5: repeat 6: Calculate Levpr of each point pr ∈ P∗ and extract the set P∗high ; 7: for each point pr in P∗high do v of each user u ∈ U∗ ; 8: Extract candidate user set U∗r and calculate Ukr k r ∗ v v 9: Rankr : rank users in Ur according to Ukr from high to low; v 10: if exist user in Rankr who meets   B then 11: Extract the user uk with the highest utility and meeting time constraints; 12: User uk → recruitment list ReListrv ; 13: Update Contri, N eed, L# and B; 14: else 15: Remove pr from P∗ ; 16: end if 17: end for 18: until P∗ = ∅; 19: end for

user with the highest utility value will be recruited. Finally, the recruitment result is recorded in .ReListrv , and related parameters are updated. The detailed pseudocode of L-UURe is shown in Algorithm 3.

3.5.4

Global Level-First and Utility-Based User Recruitment (GL-UURe)

In the HTUC scenario, the time attribute of each heterogeneous task is different, which is also the basis for dividing the sensing cycle. That is, the executable states of tasks at the same point may be different in different sensing cycles. Therefore, we propose the GL-UURe algorithm, the main idea of which is to recruit users in the cycle when the level of the point reaches the maximum based on the global information of heterogeneous tasks and users. To facilitate the experiment, .Levmax is defined as a set of the global highest level of each point. The workflow of the GL-UURe algorithm will be described in detail. Before the whole recruitment process is divided into multiple sensing cycles, the set .Levmax is first initialized. For each sensing cycle .cv , the global highest level of ∗ each point in .Levmax is judged whether it is satisfied, and .Pmax is used to denote ∗ the set of candidate points. Afterward, for each point in .Pmax in .cv , the basic idea of ∗ L-UURe algorithm is still used to recruit users. Note that .Pmax here is equivalent to ∗ .P in L-UURe. When the recruitment process of all sensing cycles is over, the set .Levmax is updated, and the above recruitment process is executed circularly until

184

H. Chen et al.

Algorithm 4 Global Level-first and Utility-based User Recruitment (GL-UURe) Input: ti ∈ T, uk ∈ U, pr ∈ P, tij ∈ ti , Tti , Ltij , Huk , Luk , Lpr . Output: ReListPC : Recruited user list. 1: Determine sensing cycle C;   2: Initially, ReListrv = ∅, Contri {uk } = [0]1×m , N eed tij = κ; 3: repeat 4: Extract the global highest level set Levmax ; 5: for each sensing cycle cv in C do 6: for each point pr in P do 7: if Levmax [r] is satisfied in cv then ∗ ; 8: Point pr → Pmax 9: end if 10: end for ∗ do 11: for each point pr in Pmax 12: Recruit users with L-UURe algorithm; 13: Update recruited list ReListrv ; 14: end for 15: end for 16: Update Levmax ; 17: until Levmax = 0 or no users can be recruited;

all tasks are completed or there are no candidate users. The detailed pseudocode of GL-UURe is shown in Algorithm 4.

3.6 Performance Evaluation In this section, we first describe the experiment settings and then make a comparative study of the performance of the three user recruitment algorithms we proposed via extensive experiments.

3.6.1

Experiment Settings

In the experiments, the locations of all heterogeneous users and heterogeneous subtasks are randomly generated in a 20 km .× 20 km area, and the sensing time of each heterogeneous task is a random and continuous time period from 8:00 to 14:00. We assume that the traveling speed Sp of each heterogeneous user is a random value from 2 to 5 m/s, and the payment for movement is set as .β = 0.1$/km. And the time budget .Bvuk of the user .uk in the sensing cycle .cv is equal to the time length of this cycle. The weights to measure the task priority are set as .ω1 = 0.3, .ω2 = 0.3, and .ω3 = 0.4. We also assume that the initial reward for the lowest priority task is .Rmin = 2$ and set .σ = 0.5, .α = 0.5. In addition, according to Eqs. (7d) and (7e), we set .κ = 3 and .ξ = 3. We perform each experiment 500 times and use the average value to demonstrate the performance.

Spatiotemporal Task Allocation in Mobile Crowdsensing 100% 90%

3500 3000 2500 2000 1500 UURe L-UURe GL-UURe

1000

Task coverage ratio

Total platform payment ($)

4000

185

500 50 100 150 200 250 300 350 400 450 500

Number of users (n)

80% 70% 60% 50% 40% 30% UURe L-UURe GL-UURe

20% 10% 0 50

100 150 200 250 300 350 400 450 500

Number of users (n)

(a)

(b)

Fig. 8 Comparison of three algorithms with a different number of users

3.6.2

Experimental Results and Analysis

The optimization goal of the HURoT problem is to minimize the total platform payment and maximize the total task completion coverage ratio. Therefore, the total platform payment and the task coverage ratio are two important indicators to evaluate the performance of the proposed algorithms. In addition, we also record the running time of different algorithms. The Impacts of the Number of Users Figure 8 shows the impact of the number of users on the performance. We fix the number of heterogeneous tasks as .m = 4 and the number of subtasks of each task as 100 and vary the number of heterogeneous users n from 50 to 500. Figure 8a shows the comparison of total platform payments of three algorithms. GL-UURe is obviously superior to L-UURe and UURe, and the more users, the more obvious the advantages are. Moreover, the platform payment increases first and then decreases slowly. This is because when users are insufficient, more users are recruited to complete tasks with the increase of users. And when users are sufficient, the platform tends to recruit the users with higher utility values, which leads to the decrease of the total platform payment. Note that when the number of users is smaller than 120, GL-UURe generates more platform payment because more subtasks are completed. Figure 8b shows the comparison of task coverage ratios. We can see that the task coverage ratio is gradually increasing with the increase of users, and GL-UURe is always better than the other two algorithms. Furthermore, when the number of users is more than 360, the task coverage ratio of GL-UURe and L-UURe can reach .100%, but UURe cannot. The Impacts of the Number of Tasks Figure 9 shows the impact of tasks’ number on the performance. We set the number of heterogeneous tasks m as 2, 3, 4, and 5 and fix the number of heterogeneous users as .n = 200 and the number of subtasks of each heterogeneous task as 100.

186

H. Chen et al. 100% UURe L-UURe GL-UURe

4500 4000

90%

Task coverage ratio

Total platform payment ($)

5000

3500 3000 2500 2000 1500 1000 500

80% 70% 60% 50% 40% 30% UURe L-UURe GL-UURe

20% 10%

0 2

3

4

Number of tasks (m)

5

0 2

3

4

5

Number of tasks (m)

(a)

(b)

Fig. 9 Comparison of three algorithms with a different number of tasks

Figure 9a shows the comparison of total platform payments of the three algorithms. We can see that the total platform payments of three algorithms are growing gradually. This is because as the number of tasks increases, the recruited users will perform more tasks. When the number of tasks is small, the total platform payment of GL-UURe is the smallest, while UURe is the largest. However, when the number of tasks is 5, the platform payment of GL-UURe exceeds that of LUURe, which is caused by the higher task coverage ratio of the former. Figure 9b shows the comparison of task coverage ratios of three algorithms. Note that the performance of GL-UURe is the best, while UURe is the worst, and the advantage of GL-UURe becomes more obvious with the increase of tasks. We can also see that the task coverage ratios of three algorithms decrease in Fig. 9b, which is caused by the limited number of users. The Impacts of the Number of Subtasks Figure 10 shows the impact of the number of subtasks on performance. We fix the number of heterogeneous tasks and heterogeneous users as .m = 4 and .n = 200 and vary the number of subtasks from 50 to 140. With the increase of subtasks, the total platform payments of three algorithms are increasing, while the task coverage ratios are decreasing, which can be verified in Figs. 10a and b, respectively. And the performance of GL-UURe is always better than that of the other two algorithms, since GL-UURe can make full use of the task execution ability of users. Moreover, when the number of subtasks is greater than 140 in Fig. 10a, the platform payment of GL-UURe is higher than that of the other two algorithms, which is caused by the higher task coverage ratio of the GL-UURe. Running Time Figure 11 shows the running time of the three algorithms with different users, tasks, and subtasks. Briefly, UURe has the shortest running time, followed by GL-UURe, and L-UURe is the largest. Next, we will make a further analysis. Figure 11a shows

187

100%

4000

90%

Task coverage ratio

Total platform payment ($)

Spatiotemporal Task Allocation in Mobile Crowdsensing

3500 3000 2500 UURe L-UURe GL-UURe

2000 1500 50

60

70

80

80% 70% 60% 50% 40% 30% UURe L-UURe GL-UURe

20% 10% 0 50

90 100 110 120 130 140

Number of subtasks

60

70

80

90

100 110 120 130 140

Number of subtasks

(b)

(a)

Fig. 10 Comparison of three algorithms with different number of subtasks 6000

2600 2400 2200 2000 1800 1600 1400

4000 UURe L-UURe GL-UURe

5000

Total running time (s)

UURe L-UURe GL-UURe

2800

Total running time (s)

Total running time (s)

3000

4000 3000 2000 1000

1200 1000 50 100 150 200 250 300 350 400 450 500

Number of users (n)

(a)

0 2

3

4

Number of tasks (m)

(b)

5

UURe L-UURe GL-UURe

3500 3000 2500 2000 1500 1000 500 50

60

70

80

90 100 110 120 130 140

Number of subtasks

(c)

Fig. 11 Comparison of three algorithms in running time with different number of users, tasks and subtasks

that with the increase of users, the running time of the three algorithms has different trends. Note that L-UURe decreases with the increase of users, because all tasks can be completed earlier, which reduces the calculation workload of each sensing cycle in the later period. However, as the number of users increases, the running time of UURe first decreases and then increases. The reason for the decrease is similar to that of L-UURe, while the increase of the running time is due to the consideration of the level of points, which increases the workload of later cycles. Different from the above two algorithms, with the increase of users, GL-UURe provides more possibilities to recruit better users for each perception cycle, which is why the running time keeps increasing. Figure 11b and c shows that the running time of the three algorithms increases with more heterogeneous tasks and subtasks, respectively. Different from the increase of users, the increasing number of tasks or subtasks will add more spatial points to the sensing system, which will affect the whole recruitment process.

188

H. Chen et al.

4 Conclusion This chapter introduces the spatiotemporal task allocation problem in mobile crowdsensing and specifically considers two scenarios. One is that the task requires sensing data with a specific duration and mobile users with limited time budgets, and the other is that the spatiotemporal characteristics and sensor requirements of the task are different and the mobile users are heterogeneous (e.g., personal preferences, carrying sensors). For each problem, we first introduce its background and then give a formal definition. Furthermore, we design the corresponding optimization algorithms and conduct extensive comparative experiments on various indicators (i.e., a different number of users and sensing tasks) to validate the effectiveness of the proposed algorithms.

References 1. R.K. Ganti, F. Ye, H. Lei, Mobile crowdsensing: current state and future challenges. IEEE Commun. Mag. 49(11), 32–39 (2011) 2. X. Chen, X. Wu, X. Li, X. Ji, Y. He, Y. Liu, Privacy-aware high-quality map generation with participatory sensing. IEEE Trans. Mob. Comput. 15(3), 719–732 (2016) 3. R. Gao, M. Zhao, T. Ye, F. Ye, G. Luo, Y. Wang, K. Bian, T. Wang, X. Li, Multi-story indoor floor plan reconstruction via mobile crowdsensing. IEEE Trans. Mob. Comput. 15(6), 1427– 1442 (2016) 4. P. Dutta, P.M. Aoki, N. Kumar, A.M. mainwaring, A. Woodruff, Common sense: participatory urban sensing using a network of handheld air quality monitors, in Proceedings of ACM SenSys (2009), pp. 349–350 5. H. Chen, B. Guo, Z. Yu, Q. Han, Toward real-time and cooperative mobile visual sensing and sharing, in Proceedings of IEEE INFOCOM (2016), pp. 1–9 6. X. Lu, Z. Yu, L. Sun, C. Liu, H. Xiong, C. Guan, Characterizing the life cycle of point of interests using human mobility patterns, in Proceedings of ACM UbiComp (2016), pp. 1052– 1063 7. F. Xu, P. Zhang, Y. Li, Context-aware real-time population estimation for metropolis, in Proceedings of ACM UbiComp (2016), pp. 1064–1075 8. Z. Yu, F. Yi, Q. Lv, B. Guo, Identifying on-site users for social events: mobility, content, and social relationship. IEEE Trans. Mob. Comput. 17(9), 2055–2068 (2018) 9. S. Morishiita, S. Maenaka, D. Nagata, M. Tamai, K. Sato, SakuraSensor: quasi-realtime cherry lined roads detection through participatory video sensing by cars, in Proceedings of ACM UbiComp (2015), pp. 695–705 10. L. Wang, Z.Yu, Q. Han, B. Guo, H. Xiong, Multi-objective optimization based allocation of heterogeneous spatial crowdsourcing tasks. IEEE Trans. Mob. Comput. 17(7), 1637–1650 (2018) 11. X. Li, X. Zhang, Multi-task allocation under time constraints in mobile crowdsensing. IEEE Trans. Mob. Comput. 20(4), 1494–1510 (2021) 12. J. Xu, J. Xiang, D. Yang, Incentive mechanisms for time window dependent tasks in mobile crowdsensing. IEEE Trans. Wirel. Commun. 14(11), 6353–6364 (2015) 13. P. Cheng, X. Lian, L. Chen, C. Shahabi, Prediction-based task assignment in spatial crowdsourcing, in Proceedings of IEEE ICDE (2017), pp. 997–1008 14. Y. Jiang, K. Zhang, Y. Qian, L. Zhou, P2AE: preserving privacy, accuracy, and efficiency in location-dependent mobile crowdsensing. IEEE Trans. Mob. Comput. 22, pp. 1–1 (2021)

Spatiotemporal Task Allocation in Mobile Crowdsensing

189

15. Z. Wang, J. Hu, J. Zhao, D. Yang, H. Chen, Q. Wang, Pay on-demand: dynamic incentive and task selection for location-dependent mobile crowdsensing systems, in Proceedings of IEEE ICDCS (2018), pp. 611–621 16. L. Wang, Z. Yu, D. Zhang, B. Guo, C.H. Liu, Heterogeneous multi task assignment in mobile crowdsensing using spatiotemporal correlation. IEEE Trans. Mob. Comput. 18(1), 84–97 (2019) 17. J. Zhang, X. Zhang, Multi-task allocation in mobile crowd sensing with mobility prediction. IEEE Trans. Mob. Comput. 22, pp. 1–1 (2021) 18. Q. Tao, Y. Tong, Z. Zhou, Y. Shi, L. Chen, K. Xu, Differentially private online task assignment in spatial crowdsourcing: a tree-based approach, in Proceedings of IEEE ICDE (2020), pp. 517–528 19. Y. Huang, H. Chen, G. Ma, K. Lin, Z. Ni, N. Yan, Z. Wang, OPAT: optimized allocation of time dependent tasks for mobile crowdsensing. IEEE Trans. Industr. Inform. 18(4), 2476–2485 (2022) 20. S. He, D. H. Shin, J. Zhang, J. Chen, Toward optimal allocation of location dependent tasks in crowdsensing, in Proceedings of IEEE INFOCOM (2014), pp. 745–753

Part IV

Key Technical Components: Data Inference

Joint Data Collection and Truth Inference in Spatial Crowdsourcing Xiong Wang

1 Introduction In this chapter, we propose an architecture of joint data collection and truth inference in spatial crowdsourcing systems. The key gradient to gathering highquality data is assigning crowdsourcing tasks to reliable workers, which is of paramount importance to inferring accurate ground truth. The reliability or expertise of workers is often undisclosed to the crowdsourcing platform, yet it can be estimated along with the discovered truth in truth inference process. The mutual dependence between task allocation and truth inference makes it indispensable to jointly model inference and allocation processes, rather than exploring them separately in previous works [1, 2]. Considering a practical spatial crowdsourcing system in Fig. 1, there are three identified characteristics: location awareness, information dynamics, and quality heterogeneity. Tasks are often tagged with spatial locations, and hence, a worker can execute a task only when the worker happens to be in the vicinity, a.k.a. location awareness. Take noise measurement as an example [3], where the noise map is useful only when workers measure them at required locations. In consequence, task allocation is further restricted by spatial coverage [4]. Due to online task arrivals and moving mobile workers, the crowdsourcing system will keep their dynamic information, namely information dynamics. To deal with the dynamic data, it is essential to establish online truth inference and task allocation. Previous works often explore offline crowdsourcing, where the information of tasks and workers is given in advance [5, 6]. Offline methods are infeasible to online situations, as online crowdsourcing has to handle unavailable future information of dynamic

X. Wang () Huazhong University of Science and Technology, Wuhan, China e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 J. Wu, E. Wang (eds.), Mobile Crowdsourcing, Wireless Networks, https://doi.org/10.1007/978-3-031-32397-3_8

193

194

X. Wang

incoming Task

Worker Mobility

Numerical Task Categorical Task

Allocate Task Spatial Coverage

Task Type

Collect Data

Time

Task Allocation Location-aware

Data Parameter

Truth Inference

Crowdsourcing Platform

Fig. 1 Snapshot of the spatial crowdsourcing system

tasks and workers [7, 8]. Quality heterogeneity indicates that workers have different expertises to different tasks, thereby making it necessary to model various worker expertises. Besides, both numerical and categorical tasks will also coexist in many applications [9]. The accuracy of collected data is hard to guarantee owing to irresponsible workers, lack of sensor calibrations, limited device capability, etc. These factors will cause variety in worker expertise to different types of tasks and further result in heterogeneous data quality. The three characteristics are indispensable in real situations, which should be resolved in both the inference and allocation processes. To this end, we first utilize probabilistic graphical model for numerical and categorical tasks to design an unsupervised inference technique so as to learn the ground truth and worker expertise. Then, an online allocation scheme is proposed, incorporating the estimated truth and expertise, to select highquality workers for data collection. The inter-play between truth inference and task allocation would mutually improve the inference accuracy, i.e., discover reliable information from noisy data.

1.1 Challenges and Motivations The core problem in spatial crowdsourcing is to discover highly accurate results from noisy crowdsourced responses, which, however, is severely impacted by collected data quality. As a result, it is crucial to resolve the issues of truth inference and task allocation [7], where the former is to infer truth from multiple noisy data through appropriate techniques, and the latter aims to assign tasks to workers for data collection. In order to improve data quality and further obtain accurate results [10], one needs to concurrently develop robust methods to infer the unknown ground truth and match tasks to reliable workers for gathering high-quality data. As aforementioned, the characteristics of location awareness, information dynamics, and quality heterogeneity are inherent in most crowdsourcing systems, which, however, are frequently ignored in existing researches pertaining to truth

Joint Data Collection and Truth Inference in Spatial Crowdsourcing

195

inference and task allocation. The absence of joint inference and allocation would severely compromise the accuracy of inferred results, as truth inference alone still suffers from low quality of crowdsourced data. In this chapter, we aim to collect high-quality data to discover accurate truth through a coupled inference and allocation framework. To tackle these problems, we are faced with the following challenges. First, numerical task has continuous value, while categorical task has discrete value. The two different patterns call for distinct treatments, which are tough to be unified. Second, location awareness introduces spatial constraints in matching workers to tasks, thus raising the problem difficulty. Third, assigning tasks to workers involves both historical allocation and current inference results, so that an online formulation is required. Fourth, designing inter-dependent truth inference and task allocation needs concurrent estimations of both ground truth and worker expertise. Though these two topics have been explored, it still remains unresolved to develop a unified framework to characterize the inter-play between inference and allocation processes, not to mention mutually discovering highly accurate ground truth.

2 Model of Truth Inference and Task Allocation We consider a spatial crowdsourcing system, where time is slotted .1, 2, . . . , t, . . . with each time slot length ranging from a few seconds to hours depending on specific applications and practical situations.

2.1 System Overview There are .Kn types of numerical tasks and .Kc types of categorical tasks. Each task .τj has an arrival time .aj , a deadline .dj , and a location .lj . If current time slot t falls into the span .[aj , dj ], we call .τj is active. Denote the active numerical (categorical) task set in time slot t as .n (t) (.c (t)), and .(t) = {n (t), c (t)} is the whole active task set. There are N workers .W = {w1 , w2 , . . . , wi , . . . , wN } in the crowdsourcing system. For each worker .wi , it has an expertise .eik to the kth type numerical or categorical task, representing the quality of its collected data. Since the data quality to different task types, like measuring environmental noise and collecting Wi-Fi fingerprint, is often different due to various sensor calibrations, the worker expertise is also heterogeneous to different types of numerical or categorical tasks accordingly. In current time slot, each worker .wi is located at location .li with a coverage range .R(li ), and .li changes over time because of worker’s mobility. Considering location awareness and spatial constraint, a task .τj ∈ (t) can be assigned to worker .wi if .lj ∈ R(li ). Denote k-th type active numerical (categorical) n (t) (. c (t)), and workers allocated to active tasks assigned to .wi in time slot t as .ik ik

196

X. Wang

task .τj as .Wj (t). When allocated to task .τj , worker .wi will complete the task and communicate the collected data .xij to the platform, where .xij is also called an observation. Due to limited device capability, worker .wi can complete at most .Ci tasks. Our objective is to undertake as many tasks as possible and infer the truth with high accuracy.

2.2 Truth Inference Definition 1 For a task τj , its truth is denoted as zj . Observations from  allocated workers are wi ∈Wj (t) xij . Truth inference is to utilize observations wi ∈Wj (t) xij to infer zj . There are two kinds of tasks, namely numerical task and categorical task. The numerical task asks for a worker to provide a numerical value, like the temperature, so that the ground truth zj is a continuous value. In contrast, the categorical task will ask for a worker to select an answer from a predefined set, such as whether the traffic is congested or not, and hence, the truth zj is a discrete value.

2.2.1

Numerical Task

2 ), where For a task .τj of k-th type, the expertise of worker .wi is .eik = (hik , σik 2 .hik is estimation bias and .σ ik is estimation variance. Estimation bias means that the average error is often not zero, and estimation variance implies that there are variations of errors around the bias. Therefore, the observation .xij is generated according to

xij = zj + nij .

.

(1)

Here, .nij is the data error, which is related to expertise .eik and complies with a 2 ): Gaussian distribution .N(hik , σik .

2 p(nij |hik , σik )= √

 (n − h )2  ij ik . exp − 2 2σik 2π σik 1

(2)

Note that noise .nij to the same task type is in the same range, while observation xij is different due to various ground truth .zj . From Eq. (2), we know that when a worker has high expertise, its estimation bias and variance are small, i.e., the data quality is high with low error. In line with Eqs. (1) and (2), the observation .xij is also a continuous value, which deviates from the ground truth .zj by an error .nij and follows a Gaussian distribution:

.

2 xij ∼ N(zj + hik , σik ).

.

(3)

Joint Data Collection and Truth Inference in Spatial Crowdsourcing Fig. 2 Probabilistic graphical model of numerical tasks

197

ℎ 2

2 2

n

Frequently, the platform has prior information of the ground truth and worker expertise. To capture this, the quantity of truth .zj is assumed to be generated from a Gaussian distribution .N(μj , σj2 ) characterized by the prior belief hyperparameters 2 .μj , σ : j p(zj |μj , σj2 ) = √

.

 (z − μ )2  j j . exp − 2 2σj 2π σj 1

(4)

As described in Eq. (2), worker expertise determines the noise added to .zj . In particular, we model that estimation bias .hik obeys a Gaussian distribution 2 .N(μik , λ ): ik .

p(hik |μik , λ2ik ) = √

 (h − μ )2  ik ik , exp − 2 2λik 2π λik 1

(5)

where .μik , λ2ik are hyperparameters, encoding the prior information on mean and variance of .hik . Consistently, we leverage an inverse Gamma distribution 2 .Inv-Gam(aik , bik ) to restrict the freedom of estimation variance .σ : ik .

2 p(σik |aik , bik ) =

aik  b  bik ik 2 −aik −1 (σik ) exp − 2 , (aik ) σik

(6)

where .aik , bik are hyperparameters of Inv-Gam distribution. Similar prior distributions in Eqs. (5) and (6) are adopted in [11, 12], and one can also employ other distributions when appropriate. On this basis, we propose a probabilistic graphical model to infer the truth, which is illustrated in Fig. 2.

2.2.2

Categorical Task

We focus on the binary categorical task, namely the truth and observation of a task are either 1 or 0. This model can be easily extended to the case of multiple answers. For task .τj of kth type, the expertise of worker .wi is .eik = (sik , fik ), in which .sik and .fik are the sensitivity and specificity, respectively, capturing the accuracy of workers’ observations [1].

198

X. Wang

Fig. 3 Probabilistic graphical model of categorical tasks ′ ′

c

Definition 2 Sensitivity is the probability of truth 1 being claimed as 1. Specificity is the probability of truth 0 being claimed as 0. Therefore, the observation .xij is generated according to 1−xij

x

p(xij |zj = 1, eik ) = sikij (1 − sik )1−xij , p(xij |zj = 0, eik ) = (1 − fik )xij fik

.

, (7)

i.e., given the truth .zj , we calculate the probability when .xij = 0 or .xij = 1. Moreover, we consider that the truth .zj obeys a Bernoulli distribution to enclose the prior knowledge: p(zj = 1) = θj ,

.

(8)

where .θj is drawn from a Beta distribution .Beta(αj , βj ): .

p(θj |αj , βj ) =

1 α −1 θj j (1 − θj )βj −1 . B(αj , βj )

(9)

B(αj , βj ) in Eq. (9) is the normalization constant. .αj , βj are hyperparameters, containing prior belief on .θj . Because of conjugate prior, Beta distribution is well suited for Bernoulli distribution to enable a tractable inference analysis [5, 6]. Analogously, sensitivity .sik and specificity .fik are generated from Beta distribu , β  ), respectively, to restrict their freedoms [1]: tions .Beta(αik , βik ) and .Beta(αik ik

.

p(sik |αik , βik ) =

1 s αik −1 (1 − sik )βik −1 , B(αik , βik ) ik

(10)

  p(fik |αik , βik )=

1  α  −1 fikik (1 − fik )βik −1 ,   B(αik , βik )

(11)

.

.

 , β  ) are hyperparameters, encoding prior information on .s where .αik , βik (.αik ik ik (.fik ). The structure of the truth inference for categorical task is shown in Fig. 3.

Joint Data Collection and Truth Inference in Spatial Crowdsourcing

199

2.3 Task Allocation In every time slot, we infer the truth of each completed task and then assign active tasks to workers. Since device capability of a worker is limited, worker .wi can execute at most .Ci tasks in each time slot. Under these constraints, we define the allocation condition in the following. Definition 3 (Allocation Condition) The condition that a task .τj can be assigned to a worker .wi is listed: • .wi ’s coverage range .R(li ) covers .τj ’s location .lj . • .wi ∈ / Wj (t), i.e., .wi has not been allocated to .τj yet. • In current time slot, the tasks assigned to .wi do not exceed its capability .Ci . Workers who satisfy the allocation condition of .τj are considered as .τj ’s available workers. Also, tasks that meet the condition of .wi are its available tasks. The principle of task allocation is to improve the average inference accuracy; thus, we need to assign tasks that have low precision or high uncertainty when inferred only based on their current claimed observations. In other words, we tend to give less priority  to task .τj that can already be accurately discovered according to observations . wi ∈Wj xij . For example, when .τj is a categorical task and we have obtained that the inferred distribution of .(1, 0) is .(0.99, 0.01), then the truth .zj is very likely to be 1 even no further observations are claimed. In this case, we ought to allocate the rest of worker resources to those tasks with answer distribution near to .(0.5, 0.5). Note that only active tasks can be assigned to their available workers, that is, the tasks whose deadlines are expired will be removed out of active task set.

2.4 Process of Crowdsourcing System In the crowdsourcing system, truth inference and task allocation will proceed alternately. According to Fig. 1, task allocation collects observations that serve as the raw inputs for truth inference, while truth inference estimates worker expertise and task truth that are involved in task allocation. The whole process is described in the sequel, where the warm up phase is the initial stage with random task allocation. Algorithms 1 and 2 are online inference process and online allocation process, respectively, which will be illustrated later. As a result, the alternate allocation and inference can mutually promote each other and lead to more accurate learned results: 1. 2. 3. 4.

Warm up phase. In each time slot, perform  steps 3–8. Obtain observations . wi ∈Wj xij based on task allocation. Call Algorithm 1 to infer truth.

200

5. 6. 7. 8.

X. Wang

Update inferred truth and estimated worker expertise. Join newly arrived tasks and remove expired tasks. Call Algorithm 2 to assign active tasks. Proceed to next time slot.

3 Online Expertise-Aware Truth Inference Truth inference for numerical and categorical tasks is based on Bayes rule, which will be elaborated individually due to their inherent differences.

3.1 Maximum Likelihood Numerical Inference At the beginning, we discuss truth inference for numerical tasks. For notational 2 } as parameters. Let .Z and .X represent truth convenience, we regard .n = {hik , σik set and observation set, respectively. According to Eqs. (3)–(6), we can write out the joint distribution .p(Z, X, n ) as p(Z, X, n ) = p(X|Z, n )p(Z)p(n )    p(zj |μj , σj2 ) = τj ∈n (t)

2 p(xij |zj + hik , σik )



wi ∈Wj (t)

   2 p(σik |aik , bik ) × p(hik |μik , λ2ik ) × wi ∈W k



= .



τj ∈n (t)

 (z − μ )2  j j exp − 2 2 2σ 2π σ j 1

j

 (x − z − h )2  ij j ik  exp − × 2 2σ 2 ik τj ∈n (t) wi ∈Wj (t) 2π σik 

×



1

 b    baik ik ik 2 −aik −1 (σik ) exp − 2 (aik ) σik

wi ∈W k

×

  wi ∈W k



 (h − μ )2  ik ik . exp − 2 2λ 2 ik 2π λik 1

Denote .L(Z, X, n ) as the logarithm of .p(Z, X, n ), and then

(12)

Joint Data Collection and Truth Inference in Spatial Crowdsourcing

τj ∈n

+ τj

 1 (xij − zj − hik )2  2 − ln σik − 2 2 2σik (t) w ∈W (t)



L(Z, X, n ) = ln p(Z, X, n ) =

.

201

i

j

 (zj − μj )2  − 2σj2 ∈ (t) n

  (hik − μik )2  bik  2 + −(aik + 1)ln σik − + − 2 + const. 2 σik 2λik w ∈W k w ∈W k i

i

(13) We use maximum a posteriori to estimate .Z, n based on .X. Specifically, we compute .Z, n that can maximize .L(Z, X, n ). To achieve this goal, take the 2: derivations of .L(Z, X, n ) with respect to .zj , hik , σik μj − zj xij − hik − zj ∂L(Z, X, n ) = + , 2 wi ∈Wj (t) ∂zj σik σj2 .

xij − hik − zj ∂L(Z, X, n ) μik − hik = + , n (t) 2 τ ∈ ∂hik j σik λ2ik ik  (xij − zj − hik )2  aik + 1 1 bik ∂L(Z, X, n ) − 2 + − = + 2 . n 2 2 )2 2 τj ∈ik (t) ∂σik 2σik 2(σik σik (σik )2 By setting the derivations to 0, it yields

zj =

.

hik =

.

2 .σik

1 μ σj2 j

+



1 wi ∈Wj (t) σ 2 (xij ik

1 σj2 1 μ λ2ik ik

+

+





=



,

1 wi ∈Wj (t) σ 2 ik

1 n (t) 2 (xij τj ∈ik σik

1 λ2ik

bik +

− hik )

+

− zj ) ,

n (t)| |ik 2 σik

n (t) τj ∈ik

aik + 1 +

(14)

(xij −zj −hik )2 2 n (t)| |ik 2

,

(15)

(16)

where .| · | is the set cardinality, and k is the involved type of task in the active numerical task set .n (t). From Eqs. (14)–(16), one can observe that the solutions are not closed form 2 and .z . Therefore, we should first initialize due to the coupling between .hik , σik j 2 .zj , hik , σ , and then iterate Eqs. (14)–(16) until convergence. ik

202

X. Wang

3.2 Expectation Maximization Categorical Inference We now present the truth inference for categorical tasks. Similarly, regard .c = {θj , sik , fik } as parameters for categorical tasks. Denote .X as the observation set, and .Z as the truth set where truth .zj is the latent variable to be discovered. Since we consider binary truth inference, .zj and .xij are either 1 or 0. Using Eqs. (7)–(11), we compute the joint probability: p(Z, X, c ) = p(X|Z, c )p(Z|c )p(c )   p(zj |θj )p(θj |αj , βj ) = τj ∈c (t)

.

 

×



p(xij |zj , sik , fik )



(17)

wi ∈Wj (t)   p(sik |αik , βik )p(fik |αik , βik ),

wi ∈W k

where k is the type of task in the active categorical task set .c (t). In each time slot, we have to estimate unknown variables including latent truth .zj and parameters .θj , sik , fik according to observations .X. The estimation of truth and parameters is through maximizing their joint probability .p(Z, X, c ) based on the fact that .zj is 1 or 0. To achieve the goal, we devise an iterative method based on Expectation Maximization (EM) that consists of E-step and M-step. E-step In the E-step, we calculate the expectation of the latent variable, namely truth .zj , given estimated parameters .rc , where the superscript r implies the rth iteration. Denote .zj,1 and .zj,0 as the posterior probabilities .p(zj = 1|X, rc ) and .p(zj = 0|X, rc ), respectively. Based on Bayes rule, the posterior probability r .p(zj |X, c ) is derived: p(zj , X|rc ) p(zj , X|rc )

= r p(X|rc ) zj p(zj , X|c )  ∝ p(X|zj , rc )p(zj |rc ) = p(xij |zj , rc )p(zj |rc ).

p(zj |X, rc ) = .

wi ∈Wj (t)

(18) Particularly, we have zj,1  p(zj = 1|X, rc ) ∝ θj



x

sikij (1 − sik )1−xij ,

wi ∈Wj (t) .

zj,0  p(zj = 0|X, rc ) ∝ (1 − θj )



1−xij

(1 − fik )xij fik

wi ∈Wj (t)

(19) .

Joint Data Collection and Truth Inference in Spatial Crowdsourcing

Denote .Aj = θj 1−x fik )xij fik ij .



xij wi ∈Wj (t) sik (1

− sik )1−xij and .Bj = (1 − θj )

203



wi ∈Wj (t) (1

−

Hence, we obtain .

zj,1 =

Bj Aj , zj,0 = . Aj + Bj Aj + Bj

(20)

M-step In the M-step, we re-estimate parameters .c given posterior probabilities zj,1 , zj,0 acquired in the r-th iteration. The re-estimated parameters are denoted as r+1 .c , which are attained by maximizing the following objective function: .

cr+1 = arg max f (c , rc ) = Q(c , rc ) + ln p(c ),

(21)

.

c

where .Q(c , rc ) = EZ|X,rc [ln p(Z, X|c )]. As for .ln p(Z, X|c ), it is expressed: ln p(Z, X|c ) =

.

 ln p(zj |c ) + τj ∈c (t)

 ln p(xij |zj , c ) .



(22)

wi ∈Wj (t)

We also write out each term of .Q(c , rc ) below: Q(c , rc )

⎡

 ⎣zj,1 ln θj +



=

τj ∈c (t)

wi ∈Wj (t)



+zj,1



 xij ln sik ⎤

(1 − xij )ln (1 − sik )⎦

wi ∈Wj (t)

⎡

.

+



⎣zj,0 ln (1 − θj ) +

τj ∈c (t)

+zj,0

wi ∈Wj (t)

Furthermore,



⎤ xij ln fik ⎦ .

wi ∈Wj (t)

xij ln (1 − fik )



(23)

204

X. Wang



ln p(c ) =

p(θj |αj , βj ) +

wi ∈W

τj ∈c (t)



=

  [p(sik |αik , βik ) + p(fik |αik , βik )] k

[(αj − 1)ln θj + (βj − 1)ln (1 − θj )]

τj ∈c (t) .

+

[(αik − 1)ln sik + (βik − 1)ln (1 − sik )]

wi ∈W

k

wi ∈W

k

(24)

  [(αik − 1)ln fik + (βik − 1)ln (1 − fik )] + const. +

To acquire the estimation of each parameter in .c , we take the derivations of f (c , rc ) over .c :

.

zj,0 αj − 1 βj − 1 zj,1 ∂f − + − , = θj 1 − θj θj 1 − θj ∂θj x 1 − xij  αik − 1 βik − 1 ∂f ij + − − , = zj,1 sik 1 − sik sik 1 − sik . ∂sik c τj ∈ik (t)

∂f = ∂fik



zj,0

1 − x

ij

fik

c (t) τj ∈ik

−

 −1 β − 1 xij  αik + − ik . 1 − fik fik 1 − fik

Set the derivations to 0, re-estimations of .c in the .(r + 1)-th iteration are derived: θj =

.

zj,1 + αj − 1 , zj,1 + zj,0 + αj + βj − 2

sik =

.

c (t) zj,1 xij τj ∈ik

c (t) zj,1 τj ∈ik

+ αik − 1

+ αik + βik − 2

(25)

,

(26)

fik =

.

 c (t) zj,0 (1 − xij ) + α τj ∈ik ik − 1

.   c (t) zj,0 + α τj ∈ik ik + βik − 2

(27)

Since both latent variables .zj,1 , zj,0 in Eq. (20) and parameters .θj , sik , fik in Eqs. (25)–(27) are not closed form, E-step and M-step are processed alternatively until convergence.

3.3 Algorithm Design for Truth Inference We now describe the process of truth inference for both numerical tasks and categorical tasks together, which is shown in Algorithm 1. The convergence

Joint Data Collection and Truth Inference in Spatial Crowdsourcing

205

condition is that the change of parameters between two adjacent iterations is lower than a predefined threshold, where the computation complexity of each iteration is .O(|(t)|N ). This complexity result is because inferring the truth of any task .τj ∈ (t) needs to traverse all the observations from the worker set .Wj (t), whose cardinality is .O(N ). Algorithm 1: Online truth inference 1 2 3 4 5 6

Input: Task set (t), worker set W and observations {xij } Output: Estimated truth {zj }, estimated parameters n (t), c (t) Obtain the numerical tasks n (t) and categorical tasks c (t); Initialize truth {zj }; Initialize parameters n , c ; // For numerical tasks while not converge do Calculate zj according to Eq. (14); Calculate n according to Eqs. (15)–(16);

// For categorical tasks 7 while not converge do // E-step 8 Calculate the probability of zj being 1 and 0 according to Eq. (20); // M-step 9 Calculate c according to Eqs. (25)–(27); 10 return {zj }, n , c ;

4 Online Location-Aware Task Allocation Truth inference is based on the observations claimed by mobile workers, which are determined by task allocation. The objective of task allocation is to match tasks to workers for collecting high-quality data. To this end, we propose two pivotal criteria to make allocation decisions: probability improvement and entropy reduction, which reflect the truth inference from two different perspectives.

4.1 Probability Improvement-Based Allocation In time slot t, we first calculate the accuracy probability of each inferred truth based on current available observations and then assign tasks to workers in order to maximize the improvement of accuracy probability, namely probability improvement. Recall from the allocation condition, worker .wi can perform task .τj when .τj ’s location .lj is in .wi ’s coverage range .R(li ). Moreover, the total number of assigned numerical and categorical tasks should not exceed .wi ’s capability .Ci .

206

X. Wang

Therefore, we can formulate the allocation problem of maximizing probability improvement in time slot t as follows: max





pjn +

τj ∈n (t) wi ∈Wj (t)

.





pjc

τj ∈c (t) wi ∈Wj (t)

(28)

s.t. allocation condition, where .Wj (t), the worker group newly allocated to .τj , is to be determined. .pjn (.pjc ) is the probability improvement of a numerical (categorical) task if assigned to a new worker, which will be elaborated in later subsections. After the allocation, we will add each .Wj (t) to .Wj (t) to obtain their observations.

4.1.1

Numerical Tasks

For a numerical task, we aim to infer its truth as close to the real value as possible. ∀τj ∈ n (t) of k-th type, Eq. (3) states

.

2 p(xij − hik |zj , σik )= √

.

 (x − h − z )2  ij ik j . exp − 2 2σik 2π σik 1

(29)

Equation (29) tells that if we take off estimation bias from an observation, the result obeys to a Gaussian distribution with ground truth .zj as the mean value and 2 as the variance. When .σ 2 is small, .x − h is highly likely estimation variance .σik ij ik ik to be close to truth .zj . Hence, the objective is transformed into obtaining a truth estimation with small inference variance. Since .zj ∼ N(μj , σj2 ), we have p(μj |zj , σj2 ) = √

.

 (μ − z )2  j j . exp − 2σj2 2π σj 1

(30)

Denote .zj as the inferred truth with current available observations. Combining

1 μj + w ∈W (t) 12 (xij −hik ) i j σik σj2

1 1 + wi ∈Wj (t) σ 2 σj2 ik

Eq. (14), i.e., .zj =

, and Eqs. (29)–(30), we obtain

 (z − z )2  1 j j , exp − p(zj |zj , σ 2j ) = √ 2 2σ j 2π σ j

.

(31)

where σ 2j =

.

1 σj2

+



1 1 wi ∈Wj (t) σ 2 ik

.

(32)

Joint Data Collection and Truth Inference in Spatial Crowdsourcing

207

The result of .σ 2j is due to the linearity of Gaussian distribution. Equation (31) implies that the intermediate inferred truth follows a Gaussian distribution with the ground truth as the mean value. To improve the inference accuracy, we need to reduce the intermediate variance .σ 2j . For a task .τj , if the error between inferred truth .zj and ground truth .zj is bounded by a threshold . , we consider the inferred truth is accurate. Correspondingly, the bounded probability is called accuracy probability. As aforementioned, we can increase accuracy probability by reducing .σ 2j that is allocating high-expertise workers to the task. However, since the number of workers and their capability are limited, we ought to allocate a worker to a task that can bring a huge gain in accuracy probability to promote allocation efficiency. In consequence, the key idea of online task allocation is to choose a worker–task pair with the highest probability improvement each time, until the device capability is met or no tasks are left. Moreover, we will give more priority to tasks with high uncertainty. With a sight abuse of notations, we use .σ 2j to denote the estimated variance based  2 on current observations . wi ∈Wj (t) xij , and .σ  j to denote the estimated variance if a new worker is allocated. From previous analysis Eq. (31), we know that .zj ∼ 2

N(zj , σ 2j ), z j ∼ N(zj , σ  j ), where .zj , z j are corresponding inferred truths. The accuracy probability of .zj is .

pnj  p(|zj − zj | < ) =

       − 1− = 2 − 1. σj σj σj

(33)

Analogously, the accuracy probability after new allocation is n

.

p j  p(|z j − zj | < ) =

  σ j

     − 1− = 2 − 1. σ j σ j (34)

If we allocate a new worker to a task .τj , the probability improvement is n

pjn = p j − pnj .

.

(35)

In general, we first choose one of each worker’s available tasks with the largest .σ 2j (highest uncertainty) under allocation condition and calculate probability improvement .pjn . We then select the worker–task pair with the highest .pjn as a numerical candidate pair, which will be compared to the probability improvement of categorical tasks. Finally, we choose the numerical candidate pair if it has higher probability improvement in the comparison.

208

4.1.2

X. Wang

Categorical Tasks

Considering the discrete categorical truth, accuracy probability for categorical task is defined as the probability that the inferred truth equals to the ground truth. Since .zj,1 (.zj,0 ) is the posterior probability of .zj = 1 (.zj = 0), if one of them approaches 1, the inferred truth is more likely to be accurate. Obviously, the worst case is .zj,1 = zj,0 = 0.5, i.e., we could only randomly guess the truth from .{1, 0}. We can maximize the absolute difference .|zj,1 − zj,0 | to avoid this ambiguous case, thus maximizing the accuracy probability. Different from linear denominator of .σ 2j (shown in Eq. (32)) in numerical task, .zj,1 , zj,0 of categorical task are composed of complex formulations in Eq. (20). Hence, it is challenging to compute .zj,1 , zj,0 via selecting worker–task pair sequentially. Besides, updating .zj,1 , zj,0 also involves newly collected observations that are unknown in the allocation phase. To bypass the two problems, we first come up with a simpler representative expression, while maximizing the representative expression is equivalent to maximizing .|zj,1 − zj,0 |. We next use the expectations instead of exact observation values in the task allocation. By doing this, it is tractable to gradually allocate workers to tasks with low computation. Denote .mj as the ratio between .zj,1 and .zj,0 . Based on Eq. (19), we have  x θj wi ∈Wj (t) sikij (1 − sik )1−xij zj,1 .mj = , =  1−x zj,0 (1 − θj ) wi ∈Wj (t) (1 − fik )xij fik ij

(36)

and .mj ∈ (0, +∞). We can rewrite .zj,1 and .zj,0 as .

zj,1 =

mj 1 . , zj,0 = mj + 1 mj + 1

(37)

m −1

Intuitively, .|zj,1 − zj,0 | = | mjj +1 |. One can validate that .|zj,1 − zj,0 | decreases in m −1

(0, 1) and increases in .[1, +∞) with regard to .mj . It is improper to regard .| mjj +1 | as the representative expression due to its non-differentiability. Therefore, we have to find a differentiable expression that has the same monotonicity as .|zj,1 − zj,0 | over .mj . With these requirements, we construct the following representative expression:

.

f (mj ) = mj +

.

1 . mj

(38)

The most difficult part in maximizing .f (mj ) is that updating .mj depends on the new observations that are unknown beforehand. A feasible way is to use the worker expertise to calculate the expectations of .mj and .f (mj ). We suppose that worker .wi is newly allocated to .τj and denote .rij as x

rij =

.

1−xij

sikij (1 − sik

)

1−x (1 − fik )xij fik ij

.

(39)

Joint Data Collection and Truth Inference in Spatial Crowdsourcing

209

The definition of .rij is consistent with .mj in Eq. (36), and .rij only depends on the newly allocated worker. To allow for a gradual worker–task selection, the ideal updating of .mj is .mj ← mj rij if .wi is allocated. Then we can compute .f (mj ) and .zj,1 , zj,0 . Similar to .mj , .rij also relies on .xij that is unknown a priori. To circumvent this issue, we leverage the expectations of .mj and .rij (denoted as .E[mj ] and .E[rij ]) to update parameters and further conduct task allocation. Since .f  (mj ) > 0, .f (mj ) is convex. According to Jensen’s inequality, we have E[f (mj )] ≥ f (E[mj ]).

.

(40)

1 as the lower bound to Therefore, we can use .f (E[mj ]) = E[mj ] + E[m j] approximate .E[f (mj )]. Next, we discuss how to obtain .E[mj ] utilizing worker expertise .eik = (sik , fik ). Recall from Definition 2 of the sensitivity and specificity .sik , fik , the probabilities of .xij = 1 and .xij = 0 are

p(xij = 1) = θj sik + (1 − θj )(1 − fik ), p(xij = 0) = θj (1 − sik ) + (1 − θj )fik . (41)

.

Combining with Eq. (39), we have .rij = rij =

.

1−sik fik

sik 1−fik

with probability .p(xij = 1), and

with probability .p(xij = 0). Hence, the updating of .E[mj ] is

E[mj ] ← E[mj rij ] = E[mj ]E[rij ]  . = E[mj ] p(xij = 1)

1 − sik  sik . + p(xij = 0) fik 1 − fik

(42)

Acquiring updated .E[mj ], we are able to approximate .E[|zj,1 − zj,0 |] utilizing Eq. (37) with .mj replaced by updated .E[mj ]. In total, we conclude the allocation process for categorical tasks in each time slot: 1. For each task .τj , calculate its current .f (E[mj ]). 2. Since .E[f (mj )] has the same monotonicity as .E[|zj,1 − zj,0 |], we allocate new workers to the tasks with small .E[f (mj )] (namely high uncertainty) to eliminate their ambiguities. Specifically, we choose one of each worker’s available tasks with the smallest .f (E[mj ]) under the allocation condition, which serves as the lower bound of .E[f (mj )] in Eq. (40). 3. Compute the expectation of the absolute probability difference .E[D] = E[|zj,1 − zj,0 |] for each chosen task .τj based on .E[mj ]. If a new worker is allocated, it updates as .E[D  ] based on Eq. (42). Select the worker–task pair that can bring the highest positive probability improvement .pjc : pjc = E[D  ] − E[D].

.

Mark this worker–task pair as a categorical candidate pair.

(43)

210

X. Wang

4. Update .E[mj ] according to Eq. (42) if .pjc is higher than .pjn of the numerical candidate pair. 5. Repeat steps 1–4 until the device capability is met or there are no available categorical tasks. Remark We compare probability improvements of numerical and categorical candidate pairs and choose the higher one so as to achieve the largest gain of Eq. (28) in each selection.

4.2 Entropy-Reduction-Based Allocation In this section, we further study entropy-reduction-based allocation where entropy is used to measure the uncertainty of each task. The intuition is to eliminate task uncertainty by reducing its entropy when assigning tasks under the allocation condition. max Ennj + Encj τj ∈n (t) wi ∈Wj (t) τj ∈c (t) wi ∈Wj (t) . (44) s.t. allocation condition, where we decide newly allocated workers .Wj (t), and .Ennj (.Encj ) is the entropy reduction of a numerical (categorical) task if assigned to a new worker. After the allocation, we still add each .Wj (t) to .Wj (t) to collect new observations.

4.2.1

Numerical Tasks

As for numerical tasks, we know that the inferred truth obeys the Gaussian distribution according to Eq. (31), where ground truth is the mean value and the 2 variance is calculated in Eq. (32). Still denote .σ 2j and .σ  j as the currently inferred variance and the variance when a new worker is allocated, respectively. We obtain the entropy before a new worker is allocated: 

n

Enj = −

+∞

.

 =−

=

−∞ +∞ −∞

N(zj |zj , σ 2j )ln N(zj |zj , σ 2j )dzj  (z − z )2   1  (z − z )2  1 j j j j   ln dzj exp − exp − 2 2 2 2 2σ 2σ 2π σ j 2π σ j j j

1 1 ln (2π σ 2j ) + . 2 2

(45)

Joint Data Collection and Truth Inference in Spatial Crowdsourcing

211

The entropy after a new allocation is n

En j =

.

 1 1  2 ln 2π σ  j + . 2 2

(46)

Hence, we can compute the entropy reduction: n

n

Ennj = Enj − En j .

.

(47)

In general, we first choose one of each worker’s available tasks with the largest σ 2j (highest uncertainty) under the allocation condition and calculate its entropy reduction .Ennj . We select the worker–task pair with the largest .Ennj as a numerical candidate pair, which is then compared with the categorical candidate pair to make the final allocation decision.

.

4.2.2

Categorical Tasks

For categorical tasks, we first leverage probabilities .zj,1 , zj,0 to obtain the entropy of each task and then design an allocation scheme to reduce task entropy.  The entropy of .τj based on observations . wi ∈Wj (t) xij is Encj = −zj,1 ln zj,1 − zj,0 ln zj,0 .

.

(48)

Since .zj,0 = 1 − zj,1 , the entropy can be re-written as .Encj = −zj,1 ln zj,1 − (1 − zj,1 )ln (1−zj,1 ), so that .Encj increases in .(0, 0.5) and decreases in .(0.5, 1) over .zj,1 . zj,1 . Expressing .zj,1 through Still denote the ratio between .zj,1 and .zj,0 as .mj = zj,0 c .mj as in Eq. (37), we know that .En increases in .(0, 1] and decreases in .(1, +∞) j over .mj . Due to the complexity of .Encj with regard to .mj , we use a representative expression .f (mj ) = mj + m1j to determine the matches between workers and tasks instead. One can prove that .Encj has the opposite monotonicity to .f (mj ) with respect to .mj , i.e., we could increase .f (mj ) by reducing task uncertainty .Encj . Besides, tasks with smaller .f (mj ) have higher priority because their truths are more obscure. Since observations are inaccessible in the allocation phase, we need to use expectations of .mj , f (mj ), Encj to select worker–task pairs. The process of updating and computing those expectations has already been depicted in Sect. 4.1.2. The main difference lies in that we calculate the expected entropy instead of the expected accuracy probability. To save space, we only present the allocation flows here: 1. For each task .τj , calculate its current .f (E[mj ]). 2. Choose a worker’s available task with the smallest .f (E[mj ]) under the allocation condition, and calculate the entropy reduction .Encj .

212

X. Wang

3. Select the worker–task pair that has the highest positive entropy reduction as a categorical candidate pair. 4. Update .E[mj ] if .Encj is higher than .Ennj . 5. Repeat steps 1–4 until the device capability is met or there are no available categorical tasks. Note that we still need to compare .Encj with .Ennj so as to achieve the largest gain of Eq. (44).

4.3 Algorithm Design for Task Allocation Based on the former analysis, we illustrate the algorithm design for task allocation of both numerical and categorical tasks. For ease of exposition, we will show the process of probability-improvement-based allocation in Algorithm 2 and then indicate the difference when designing algorithm for entropy-reduction-based allocation. According to Algorithm 2, we first obtain the available tasks for each worker in Lines 2–3. After that, we check whether tasks can be assigned to worker .wi in Lines 8–9. As for numerical tasks, we select tasks with high uncertainty and calculate their probability improvement in Lines 10–12. Next, we compute the probability improvement for categorical tasks in Lines 13–15, where we are prone to choosing less accurate tasks. Finally, select the task that has the highest probability improvement, and update the parameters in Lines 16–25. The computation complexity of each round, namely Lines 5–25, is .O(|(t)|N) since the allocation will traverse all workers and their available tasks. Now we point out the differences if leveraging entropy reduction to assign tasks. Because entropy is utilized to measure the uncertainty, Line 12 should be “Obtain .Ennj according to Eq. (47),” while Line 15 should be “Obtain .Encj  .” The worker–task selection will change accordingly, in which we choose worker– task pair with the highest entropy reduction .Ennj , Encj  .

5 Performance Evaluation In this section, we will show the results for truth inference and task allocation and compare our proposed algorithms with other baseline methods.

Joint Data Collection and Truth Inference in Spatial Crowdsourcing

213

Algorithm 2: Online task allocation Input: Task set (t), task parameters {θj }, worker set W , worker expertise {eik }, device capability {Ci } n (t)}, { c (t)} Output: New task allocation {Wj (t)}, {ik ik 1 Obtain numerical and categorical tasks n (t), c (t); 2 for wi ∈ W do 3 Obtain available numerical and categorical tasks Ain (t), Aic (t); 4 while 1 do 5 if W = ∅ then 6 break; 7 8 9

for wi ∈ W do if Ci == 0 or Ain (t), Aic (t) == ∅ then W = W \ wi , continue; Obtain σ j , ∀τj ∈ Ain (t) according to Eq. (32); Select τj with the largest σ j ; Obtain the pjn according to Eq. (35); Obtain f (E[mj  ]), ∀τj  ∈ Aic (t); Select τj  with the smallest f (E[mj  ]); Obtain the pjc  according to Eq. (43);

10 11 12 13 14 15 16 17 18 19

Select worker-task pair (wi , τj ) with largest improvement pjn ; Select worker-task pair (wi , τj  ) with largest improvement pjc  ; if no available tasks then break;

20 21 22

if pjn > pjc  then Wj (t) = Wj (t) ∪ wi , Ci = Ci − 1; n (t) =  n (t) ∪ τ , update σ ; ik j j ik

23 24 25

else Wj  (t) = Wj  (t) ∪ wi , Ci = Ci − 1; c (t) =  c (t) ∪ τ  , update E[m  ]; ik j j ik

n (t)}, { c (t)}; 26 return {Wj (t)}, {Wj  (t)}, {ik ik

5.1 Dataset and Settings 5.1.1

Dataset

A proper dataset should satisfy: (1) inclusion of both numerical and categorical tasks; (2) workers and tasks being associated with locations; (3) tasks having arrival time and deadline, namely time-varying active tasks and collected data; (4) involvement of workers’ trajectories to model their mobility. Sur Dataset We utilize 10 smartphones carried by 10 users to gather their trajectories and sensory data around the surroundings, including: altimeter, magnetic field, light intensity, and noise level. We regard the aggregated data of 2 well-calibrated smartphones as the ground truth and also compute a noise set by comparing the data of one smartphone with the ground truth. Due to the impossibility of aligning

214

X. Wang

workers’ trajectories in practice, we generate data errors by first sampling a value from the calculated noise set and then adding various random offsets for other smartphones with unaligned trajectories. Task’s location is also the aggregated location of the 2 smartphones, that is, their halfway points. LTE Dataset We obtain LTE measurements from [13], where a person uses a smartphone to measure LTE signals as well as GPS information. The trajectories are divided into 160 pieces to stimulate 160 different workers, and each worker has 4 kinds of measurements (crowdsourced data): RSRQ (Reference Signal Received Quality), RSSI (Received Signal Strength Indicator), GPS height, and GPS satellite. RSRQ and RSSI are from LTE signals, while GPS height and GPS satellite are the information of current height and GPS serial number from GPS sensor. We consider the collected data of the first worker as the ground truth and acquire a noise set by comparing the data of another similar-trajectory worker with the ground truth. Therefore, we still first sample a value from the acquired noise set and add random offsets to stimulate other workers’ data errors because of their unaligned trajectories. Besides, task’s location is set to the first worker’s location since his measurements are used as the ground truth. We regard the time a piece of data sampled as the arrival time, and a random later time as the deadline since there is no specific deadline contained in the dataset. In this way, each task of the two datasets has a ground truth, arrival time, deadline, and location. According to the task allocation, only the observations from allocated high-expertise workers will be used to infer the truth. As for Sur dataset, we consider the tasks for altimeter and magnetic field as numerical tasks and the tasks for light intensity and noise level as categorical tasks. Since all the raw data are continuous values, we directly utilize altimeter and magnetic field and convert light intensity and noise level into binary values by setting thresholds. The total rounds are 216 s, with 4 pieces of data and locations every second. We apply the same procedure to LTE dataset, so we get RSRQ and RSSI as numerical tasks and GPS height and GPS satellite as categorical tasks based on the same threshold method. As the collected observations have various qualities due to the sensor calibration and worker’s effort, we will use them to infer ground truth and estimate worker expertise.

5.1.2

Parameter Settings

In line with the sampling frequency of the Sur and LTE datasets, that is 1 Hz, the time slot length is set to 4 s, and there are total .T = 216 4 = 54 time slots for both two datasets. Besides, .Kn = Kc = 2 as each dataset has 2 types of numerical tasks and categorical tasks, respectively. In every time slot, .4 × 2 = 8 numerical tasks and .4 × 2 = 8 categorical tasks arrive at the crowdsourcing system. The arrival time of a task is the time slot when it is sampled, while the deadline is the arrival time delaying a random integer value in .[4, 7]. The location of each task consists of a longitude and a latitude. Worker’s location in each time slot is the position at the starting second of that time slot.

Joint Data Collection and Truth Inference in Spatial Crowdsourcing

215

Table 1 Parameter setting Parameter T .Kn , Kc 2 .(μj , σj ) 2)

Value/range 54 2 (102.013,1.256) for altimeter, (52.218,12.576) for magnetic field

.(μj , σj

.(−7.726, 3.048)

.(αj , βj )

(19,2) for light intensity, (8,13) for noise level (30,3) for GPS height, (14,7) for GPS satellite rectangle region, within 0.0001/0.005 of .li for Sur/LTE integer in .[5, 8] for Sur and in .[4, 6] for LTE .[−3, 3], .[0, 2] .[Ci − 1, Ci + 2] .[8, 12], .[2, 5] .[8, 12], .[2, 5]

.(αj , βj ) .R(li ) .Ci

2

.μik , λik .aik , bik .αik , βik





.αik , βik

for RSRQ, .(−58.855, 30.165) for RSSI

In Sur dataset, we assign the value to each hyperparameter using statistical information of collected observations. For altimeter, .μj , σj2 are set to 102.013 and 1.256. For magnetic field, the values of .μj , σj2 are 52.218 and 12.576. Regarding light intensity, .αj , βj are set to 19 and 2. As for noise level, .αj , βj are 8 and 13. With regard to tasks in LTE dataset, .μj , σj2 for RSRQ and RSSI are .(−7.726, 3.048) and .(−58.855, 30.165), respectively. .αj , βj for GPS height are 30 and 3, while those for GPS satellite are 14 and 7. As for worker .wi , the coverage range .R(li ) includes the rectangle region whose centroid is .wi ’s current location .li . Considering trajectory sparsity, longitude and latitude ranges of .R(li ) are within 0.0001 of .li for Sur dataset and within 0.005 of .li for LTE dataset. The device capability .Ci is an integer valued in .[5, 8] for Sur dataset and in .[4, 6] for LTE dataset. In line with the noises in both two datasets, 2 is drawn from .[0, 2]. .a , b are integers in the .μik is drawn from .[−3, 3], while .λ ik ik ik range .[Ci −1, Ci +2]. .αik is sampled in .[8, 12] and .βik is valued in .[2, 5]. Similarly,   .α ik draws from .[8, 12] and .βik is valued in .[2, 5]. For convenience, Table 1 lists the parameter settings.

5.1.3

Comparison Algorithms

We introduce four baseline algorithms for comparisons: • AVE: compute the average value. For task .τj , take the mean value of observations

wi ∈Wj (t) xij

as the inferred truth. • MV:

Majority voting. Answer .zj (0 or 1) that gets the highest votes . wi ∈Wj (t) 1zj =xij is the inferred truth of task .τj . • TF: Truth finder proposed in [14]. TF iteratively updates truth confidence and worker trustworthy, which are dependent on each other. Result with the .

|Wj (t)|

216

X. Wang

highest confidence is considered as the inferred truth. Iterative updating involves implication score that implies similar observations will support each other. Set the implication score between .xij and .xi  j as .exp(−|xij − xi  j |/20) for numerical tasks, and implication score between .xij and .xi  j as 1 if they are the same and .−1 otherwise for categorical tasks. • OPT: Optimization-based algorithm designed in [9]. OPT is to find inferred truth by minimizing its weighted distance to observations. Especially, we implement Euclidean distance for both numerical and categorical tasks. For brevity, denote MLNI and EMCI as our proposed Maximum Likelihood Numerical Inference and Expectation Maximization Categorical Inference, respectively. In general, AVE, TF, OPT, MLNI are evaluated for numerical tasks, while MV, TF, OPT, EMCI are specified for categorical tasks. AVE and MV are the most basic inference algorithms and represent the methods that treat workers equally without considering worker expertise. TF and OPT can be applied to both numerical and categorical tasks, and they are able to differentiate workers via trustworthy and weight, separately, enabling a similar function to the worker expertise here. Therefore, we choose these four algorithms as benchmarks.

5.1.4

Evaluation Metric

We use RMSE and Err as the metrics to evaluate numerical and categorical tasks, respectively. Let .zj and .zj be the inferred truth and the ground truth, respectively.  RMSE =

.

1 Mn (zj − zj )2 , j =1 Mn

(49)

where .Mn is the total number of completed numerical tasks.

Mc Err =

.

j =1 1zj =zj

Mc

,

(50)

where .Mc is the total number of completed categorical tasks. .1zj =zj = 1 if .zj = zj , and .1zj =zj = 0 otherwise. If RMSE or Err is smaller, the inference result is more appreciated.

5.2 Results of Truth Inference Under settings described above, we run each evaluation for five times and show the average results in the following.

Joint Data Collection and Truth Inference in Spatial Crowdsourcing Table 2 Sur: results of truth inference. Bold values indicate the best achieved results

Probability-improvement-based allocation Methods AVE TF OPT MLNI 2.2699 1.5296 1.4456 1.0153 RMSE TF OPT EMCI Methods MV Err 0.1975 0.1249 0.1264 0.0840 Entropy-reduction-based allocation Methods AVE TF OPT RMSE 2.1170 1.5062 1.3968 TF OPT Methods MV 0.2136 0.1342 0.1405 Err

Table 3 LTE: results of truth inference. Bold values indicate the best achieved results

217

MLNI 1.0177 EMCI 0.0799

Probability-improvement-based allocation Methods AVE TF OPT MLNI 2.0181 1.7830 1.5982 1.1143 RMSE Methods MV TF OPT EMCI Err 0.2105 0.1061 0.1214 0.0601 Entropy-reduction-based allocation Methods AVE TF OPT 1.9773 1.5942 1.4595 RMSE TF OPT Methods MV Err 0.2235 0.1128 0.1375

MLNI 1.1515 EMCI 0.0674

Results of truth inference are demonstrated in Tables 2 and 3. One can see that our proposed two allocation schemes lead to slightly different inference outcomes. For numerical tasks, RMSE of our proposed MLNI is the smallest for both probability-improvement-based allocation and entropy-reduction-based allocation on two datasets. The same conclusion also holds for categorical tasks in terms of Err. Therefore, our algorithm performs better than the baseline methods. The reason behind is that the baseline methods ignore diverse worker expertise to different types of tasks. Besides, they cannot capture the prior information of both tasks and workers. Now we discuss the influence of worker number N on truth inference. We vary N from 3 to 10 for Sur dataset and from 20 to 160 for LTE dataset and record the results of truth inference. Figure 4 exhibits the RMSE for numerical tasks and Err for categorical tasks under probability-improvement-based allocation. Similarly, results for entropy-reduction-based allocation are depicted in Fig. 5. From the figures, our proposed MLNI and EMCI can achieve lowest RMSE and Err for both two task allocation schemes on both Sur and LTE datasets. As N increases, all results will fluctuate and tend to decrease, which are especially apparent for MLNI and EMCI. This is because a task will receive more observations when N is larger, and hence, a more accurate inference is accessible. We next explore the impact of device capability .Ci , i.e., the amount of tasks a worker can complete at most during one time slot. Device capability ranges from

218

X. Wang

(a)

(b)

(c)

(d)

Fig. 4 Probability-improvement-based allocation, performance, vs. the number of workers. (a) Sur: RMSE vs. N . (b) Sur: Err vs. N . (c) LTE: RMSE vs. N . (d) LTE: Err vs. N

(a)

(b)

(c)

(d)

Fig. 5 Entropy-reduction-based allocation, performance, vs. the number of workers. (a) Sur: RMSE vs. N . (b) Sur: Err vs. N . (c) LTE: RMSE vs. N . (d) LTE: Err vs. N

Joint Data Collection and Truth Inference in Spatial Crowdsourcing

219

Fig. 6 Probability-improvement-based allocation, performance, vs. device capability. (a) Sur: RMSE vs. .Ci . (b) Sur: Err vs. .Ci . (c) LTE: RMSE vs. .Ci . (d) LTE: Err vs. .Ci

[3, 6] to .[7, 10] for Sur dataset and from .[2, 4] to .[6, 8] for LTE dataset, with an increase of 1 in the lower and upper bounds. The results are shown in Figs. 6 and 7. MLNI and EMCI still have the best performance with regard to RMSE and Err. Both RMSE and Err for all the algorithms will fluctuate with the increase of device capability because the number of completed tasks is increasing dramatically that may lead to a decrease in the average number of observations for a task. Even so, we can see that RMSE and Err still decline over device capability in most times on both two datasets.

.

5.3 Results of Task Allocation We compare inferred results of our proposed allocations with those of the following two benchmarks: (1) random allocation, which will assign tasks to workers randomly; (2) greedy allocation, which myopically allocates high-expertise workers to tasks with currently lowest inferred accuracy so as to improve the average performance. Similarly, each simulation runs for five times to acquire the average results. We first show the influence of worker number N, where the results obtained from two datasets are exhibited in Fig. 8. The figure indicates that probability improvement and entropy-reduction-based allocations can always achieve lower

220

X. Wang

Fig. 7 Entropy-reduction-based allocation, performance, vs. device capability. (a) Sur: RMSE vs. (b) Sur: Err vs. .Ci . (c) LTE: RMSE vs. .Ci . (d) LTE: Err vs. .Ci

.Ci .

(a)

(b)

(c)

(d)

Fig. 8 Comparison results of allocation, performance, vs. the number of workers. (a) Sur: RMSE vs. N . (b) Sur: Err vs. N . (c) LTE: RMSE vs. N . (d) LTE: Err vs. N

Joint Data Collection and Truth Inference in Spatial Crowdsourcing

221

Fig. 9 Comparison results of allocation, performance, vs. device capability. (a) Sur: RMSE vs. (b) Sur: Err vs. .Ci . (c) LTE: RMSE vs. .Ci . (d) LTE: Err vs. .Ci

.Ci .

RMSE and Err compared to the random and greedy algorithms when N increases from 3 (20) to 10 (160) for Sur (LTE) dataset, and random allocation yields the poorest performance. In addition, the results varying with device capability are illustrated in Fig. 9, in which device capability still ranges from .[3, 6] (.[2, 4]) to .[7, 10] (.[6, 8]) for Sur (LTE) dataset. One can see that our proposed schemes distinctly outperform both random and greedy allocations in terms of RMSE and Err. The number of completed tasks is also an important criterion since we want to undertake as many tasks as possible. To this end, we examine how the number of completed tasks varies over the number of workers N and the capability .Ci . Figure 10 shows the influences of N and .Ci for both Sur and LTE datasets. As the number of workers or device capability increases, one can observe that the number of completed tasks also increases. Besides, the number of completed tasks in Sur dataset is smaller than that in LTE dataset. This is because on the one hand tasks in LTE dataset have higher uncertainty so that they are more likely to be executed, and on the other hand there are more workers considered in LTE dataset who will complete more tasks.

222

X. Wang

Fig. 10 Completed number of tasks, performance, vs. the number of workers and device capability. (a) Sur: influence of N . (b) Sur: influence of .Ci . (c) LTE: influence of N . (d) LTE: influence of .Ci

5.4 Results of Running Time We implement the methods in Python and run the experiments on a Windows machine with 8 GB RAM, Intel core i5 3.20 GHz for five times to obtain the average time. The whole running time of probability-improvement-based allocation and entropy-reduction-based allocation during 54 time slots is now discussed. The running time over the number of workers N and the capability .Ci is drawn in Fig. 11. The figure shows that the running time will grow over N or device capability since more tasks are completed in each time slot. Besides, the running time using LTE dataset is much higher than that using Sur dataset because more workers are modeled in LTE dataset. Though the running time is increasing, it is still in polynomial time of the total time slots T for the two allocation schemes on both datasets.

5.5 Results on Larger Dataset Lastly, we show that our inference and allocation framework also has good performances on a larger synthetic dataset S-LTE, considering there are no off-theshelf datasets available. Specifically, we construct S-LTE from LTE dataset in this way: for each worker, we add random noises to its trajectory and the collected data

Joint Data Collection and Truth Inference in Spatial Crowdsourcing

223

Fig. 11 Running time, performance, vs. the number of workers and device capability. (a) Sur: influence of N . (b) Sur: influence of .Ci . (c) LTE: influence of N . (d) LTE: influence of .Ci Table 4 S-LTE: results of truth inference. Bold values indicate the best achieved results

Probability-improvement-based allocation TF OPT MLNI Methods AVE RMSE 1.7653 2.1764 1.5694 1.2099 Methods MV TF OPT EMCI Err 0.1184 0.0841 0.1267 0.0395 Entropy-reduction-based allocation TF OPT Methods AVE 1.5766 2.0042 1.3586 RMSE TF OPT Methods MV Err 0.1269 0.0095 0.1321

MLNI 1.1217 EMCI 0.0434

to generate another two items including the trajectories and claimed observations. Hence, the S-LTE dataset would entail .N = 480 workers. Similarly, the results of truth inference are shown in Table 4. Intuitively, our algorithms perform better than the baseline methods, as we can achieve more accurate inferred ground truth.

6 Chapter Summary In this chapter, we investigate joint task allocation and truth inference in a spatial crowdsourcing system. We fully consider the location awareness, information dynamics, and quality heterogeneity of workers and tasks, which more conform

224

X. Wang

to real situations. We first propose probabilistic graphical model to infer truth for numerical and categorical tasks. Then, we design probability-improvementand entropy-reduction-based online task allocation schemes to enable collections of high-quality data. Finally, we perform extensive evaluations to compare our proposed algorithms with existing inference methods and allocation schemes. Results show that our inference algorithms can achieve much lower RMSE and Err for numerical and categorical tasks, respectively, and our allocation policies can collect high-quality data to improve inference accuracy.

References 1. B. Zhao, B.I.P. Rubinstein, J. Gemmell, J. Han A Bayesian approach to discovering truth from conflicting sources for data integration, in Very Large Data Base Endowment, vol. 5 (2012), pp. 550–561 2. P. Cheng, X. Lian, L. Chen, J. Han, J. Zhao, Task assignment on multi-skill oriented spatial crowdsourcing. IEEE Trans. Knowl. Data Eng. 28, 2201–2215 (2016) 3. R. Rana, C.T. Chou, N. Bulusu, S. Kanhere, W. Hu, Ear-phone: a context-aware noise mapping using smart phones. Pervasive Mob. Comput. 17, 1–22 (2015) 4. L. Kazemi, C. Shahabi, GeoCrowd: enabling query answering with spatial crowdsourcing, in ACM International Conference on Advances in Geographic Information Systems (2012), pp. 189–198 5. R.W. Ouyang, M. Srivastava, A. Toniolo, T.J. Norman, Truth discovery in crowdsourced detection of spatial events. IEEE Trans. Knowl. Data Eng. 28, 1047–1060 (2015) 6. A. Augustin, M. Venanzi, A. Rogers, N.R. Jennings, Bayesian aggregation of categorical distributions with applications in crowdsourcing, in International Joint Conference on Artificial Intelligence (2017), pp. 1411–1417 7. X. Zhang, Y. Wu, L. Huang, H. Ji, G. Cao, Expertise-aware truth analysis and task allocation in mobile crowdsourcing, in IEEE International Conference on Distributed Computing Systems (2017), pp. 922–932 8. Y. Zheng, G. Li, R. Cheng, DOCS: a domain-aware crowdsourcing system using knowledge bases, in Very Large Data Base Endowment, vol. 10 (2016), pp. 361–372 9. Q. Li, Y. Li, J. Gao, B. Zhao, W. Fan, J. Han, Resolving conflicts in heterogeneous data by truth discovery and source reliability estimation, in ACM Special Interest Group on Management of Data (2014), pp. 1187–1198 10. S. Yang, F. Wu, S. Tang, X. Gao, B. Yang, G. Chen, On designing data quality-aware truth estimation and surplus sharing method for mobile crowdsensing. IEEE J. Sel. Areas Commun. 35, 832–847 (2017) 11. R.W. Ouyang, L.M. Kaplan, A. Toniolo, M. Srivastava, T.J. Norman, Aggregating crowdsourced quantitative claims: additive and multiplicative models. IEEE Trans. Knowl. Data Eng. 28, 1621–1634 (2016) 12. B. Zhao, J. Han, A probabilistic model for estimating real-valued truth from conflicting sources, in QDB (2012) 13. W. Zhang, H. Huang, X. Tian, Gaussian process based radio map construction for LTE localization, in IEEE International Conference on Wireless Communications and Signal Processing (2017) 14. X. Yin, J. Han, P.S. Yu, Truth discovery with multiple conflicting information providers on the web. IEEE Trans. Knowl. Data Eng. 20, 796–808 (2007)

Cost-Quality Aware Compressive Mobile Crowdsensing Yong Zhao, Zhengqiu Zhu, and Bin Chen

1 Background The rapid development of the Internet of Things (IoT) and mobile computing technologies [1] promotes the emergence of intelligent, open, and large-scale sensing mechanisms, allowing citizens to effectively collect and share real-time information, and enabling innovative urban computing solutions to tackle city-level challenges such as noise [2], traffic congestion [3], and infrastructure status [4]. With the help of sensor-rich smartphones, mobile crowdsensing (MCS) [5, 6] plays an increasingly important role in urban computing for satisfying various urban-scale monitoring needs. Though the sensed information can be used to improve people’s daily life, the practical application of MCS systems is restricted by the availability of participants and the required large sensing and communication costs. To provide high-quality services with less sensing costs, current studies investigated the inherent correlations embedded in the sensory data and proposed the sparse MCS/compressive crowdsensing (CCS) approach to tackle this issue. In CCS, only few cells are selected to sense, and the missing data in unsensed cells are inferred based on the sensory data via compressive sensing method. Therefore, the organizer needs to carefully decide where to sense. In the cell selection process, data quality and cost are two primary concerns. First, data of different cells may involve diverse spatio-temporal correlations, and thus selecting different cells may lead to diverse quality of the data inference. To find the cells more

Y. Zhao · Z. Zhu () · B. Chen National University of Defense Technology, Changsha, Hunan, China e-mail: [email protected]; [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 J. Wu, E. Wang (eds.), Mobile Crowdsourcing, Wireless Networks, https://doi.org/10.1007/978-3-031-32397-3_9

225

226

Y. Zhao et al.

conducive to data inference, the compressive sensing method is also used during or before cell selection process to assess the “importance” of each cell. Second, the cost of sensing in different cells can be diverse due to many factors, such as the condition of the sensing device, location, and routing distance [7]. Since the total budget of a MCS campaign is often limited, a sophisticated cost-aware cell selection strategy may be able to collect data from more cells. Various sampling strategies have been proposed to perform the cell selection in CCS systems. For instance, randomized sampling strategy was employed in many studies [8, 9], in which the sampling is conducted independently by each cell flipping a coin according to the probability assigned to the nodes. Typical example is the cost-aware compressive sensing (CACS) [8]. The CACS aims to achieve a good tradeoff between minimizing the recovery accuracy and the total sampling cost. To predict the recovery accuracy, the CACS proposed the regularized column sum (RCS) deduced from the restricted isometry property (RIP). Then, the optimal randomized sampling strategy with the lowest cost can be produced by convex optimization. Moreover, several studies showed that the number of samples taken based on these randomized sampling methods is significantly larger than needed [10]. Therefore, the active sampling strategy carefully selecting cells to provide more information with lower sensing cost was developed [11, 12]. The active sampling strategy usually estimates the error or uncertainty of the unsensed cells by leveraging data inference algorithms and then actively selects the cell with a larger estimated error or uncertainty. The CCS-TA [11] (compressive crowdsensing task allocation) is a typical active sampling strategy that selects cells in multiple steps with a sampling stop condition controlling the cost. In this framework, the query by committee (QBC) method was employed to assess the uncertainty (importance) of different cells. Liu et al. [12] proposed a deep Qnetwork-based cell selection strategy that can approximate the global optimal strategy relying on sufficient data training. Based on those works, Zhu et al. [13] further considered diverse sensing cost of different cells. However, there is no direct relationship between the estimated uncertainty and the quality of the recovery of missing data. Besides, the uncertainty calculation of the missing data relies on the pre-recovery of the missing data. Therefore, the uncertainty calculation would suffer from a large error when the pre-recovery is not accurate. Different from the above-mentioned approaches, Xie et al. [10] proposed a bipartite-graph-based sensing scheduling scheme, called active sparse MCS (ASMCS), to actively determine the sample locations. Specifically, the bipartite graph is leveraged to denote the environment matrix with well-defined vertexes and edges, and the sensing cost is also incorporated into the graph. Then, the low-cost cell selection problem is modeled as an edge selection problem. In the next sections of this chapter, we will see how the cost–quality-aware compressive crowdsensing problem is formulated and the different advanced solutions for the problem.

Cost-Quality Aware Compressive Mobile Crowdsensing

227

2 System Model and Problem Statement In this section, the typical cost–quality-aware compressive crowdsensing system is firstly presented, which contains five parts: importance assessment, cost assessment, cell selection, quality assessment, and data inference. The cell selection will be described later, and the rest four parts are introduced in this section. Finally, the problem is formulated.

2.1 System Model A typical CCS scenario begins with a sensing task launched by its organizer to obtain fine-grained urban context results, e.g., humidity over a large-scale target area for a long time, as shown in Fig. 1. To provide high-quality sensing services, the target sensing area is divided into m cells according to the organizer’s requirement. In the meantime, the whole sensing campaign is also split into n equal sensing cycles. For instance, the organizer needs to update the full humidity sensing map once every hour (sensing cycle), and in each sensing cycle, the data quality requirement is that the mean absolute error for the whole area should be less than 1.5% (humidity). To meet the data quality requirement under the constraint of task budget B, the organizer needs to carefully select cells to make a tradeoff between the data quality and the sensing cost in a cell. After meeting the quality requirement in a sensing cycle, the humidity values of the remaining cells are deduced through data inference approaches. Through this crowd-powered subset sensing approach, the organizer can obtain sufficient data based on the sensing requirement and costs.

2.2 Data Inference In CCS, we often leverage the historical and the current sensory data to infer the data of the remaining unsensed cells. Compressive sensing, as a good choice for inferring the full ground data matrix from the partially collected sensing values with convincing theoretical deviation, has shown its effectiveness in several scenarios [9]. ˆ m×n based on the low-rank property is recovered as The full ground data matrix .G follows: .

ˆ min rank(G)

(1)

ˆ ◦ S = M, s.t., G where .◦ denotes the element-wise multiplication and each entry .Sij denotes whether the cell i at cycle j is selected for sensing. Thus, .Sij equals 0 or 1. M denotes the actual measurement data matrix. Based on the singular value decom-

228

Y. Zhao et al.

Fig. 1 The general process of the system model

position (SVD) and compressive sensing theory, we convert the above nonconvex optimization problem as follows: .

min λ(||L||2F + ||R||2F ) + ||LR T ◦ S − M||2F .

(2)

This optimization changes the rank minimization problem (minimize the rank of ˆ to minimizing the sum of L and R’s Frobenius norms. And .λ allows a tunable G) ˆ tradeoff between rank minimization and accuracy fitness. To get the optimal .G, alternating least squares procedure is leveraged to estimate L and R iteratively. ˆ is calculated as follows: Finally, the inferred data matrix .G

.

ˆ = LR T . G

.

(3)

2.3 Importance Assessment Before selecting cells, the CCS system would like to know the importance of each cell or the information provided by each cell. In traditional active learning, if the model is less certain about the prediction on an instance, then the instance is considered to be more informative for improving the model and will be more likely to be selected for label querying [14]. Inspired by this idea, a reward criterion or a value criterion can be leveraged to estimate the informativeness of an entry, i.e., a spatio-temporal cell in the matrix G. The challenge here is how to quantify the informativeness of an instance in a cell for recovering the entries in other cells.

Cost-Quality Aware Compressive Mobile Crowdsensing

229

Obviously, the entropy-based (e.g., QBC) or mutual-information-based method (e.g., Gaussian-process-based mutual information) belongs to a reward criterion, which indicates what is good in an immediate sense, while the value-function-based method (reinforcement-learning-based method) considers what is good in the long run. Though mutual-information-based method and value-function-based method can better quantify which cell is more informative, they require sufficient historical data to compute the informativeness of a cell. Also, they are unable to be applied in a fresh task when no sufficient data are acquired. Therefore, a simple but general method is introduced here, i.e., QBC as the importance assessment method. As for other suitable approaches, interested readers can refer to [10, 15]. QBC originates from such an idea that if the variance of an entry is large, it implies that the entry cannot be certainly decided by the algorithm and thus may contain more useful information to recover the estimated full sensing matrix. QBC framework quantifies the prediction uncertainty based on the level of disagreement among an ensemble of compressive sensing algorithms. Specifically, a committee of matrix completion algorithms is applied to the partially observed data matrix to impute the missing values. The variance of prediction (among the committee members) of each missing entry is taken as a measure of uncertainty of that entry. In general, the committee consists of several commonly used inference algorithms, including compressive sensing, spatio-temporal compressive sensing, K-nearestneighbors (KNN), and SVD. Assume that the committee includes a set of H inference algorithms. In a sensing cycle j, the already selected cells with measurements in this cycle are denoted by .Sj (.Sj ∈ V). The sensor measurements in these selected locations are represented by ˆ j ) = Rh (xSj ). .χSj = xSj . By using one of the inference algorithms, we have .G(:, For an unsensed cell .υ ∈ / Sj , the informativeness of this cell can be formulated as Iυ,j =

.

H h=1

2

ˆ h (υ, j ) − G(υ, ¯ (G j )) /H,

(4)

where .Iυ,j represents information (importance) of unsensed cell .υ in time cycle ¯ ˆ h (υ, j ) is the j; .G(υ, j ) denotes the average value predicted by the committee; .G predicted value of the h-th matrix completion algorithm in cell .υ.

2.4 Cost Assessment In practical CCS applications, an activity organizer has to consider the diversity of sensing costs for several reasons: (1) Sensors possessed by mobile users are inherently diverse, and measurement accuracy largely depends on the sensors. Generally, data reports with high precision should be offered with higher rewards; (2) The cost of reporting a sensory data to the organizer varies based on the network condition, distance to the nearest cell tower, cellular data plan, or other concurrent activities on the device; (3) Prior work also found evidence that the final cost may

230

Y. Zhao et al.

also be affected by the subjective perception of participants. For instance, a user would ask for a higher reward when he is running out of battery. In brief, different conditions can result in vastly different costs in crowdsensing activities. However, it is difficult to assess sensing cost. Since real costs are hard to obtain, a cost estimation function is often required. Due to the cost diversity, a simple cost estimation function is hard to estimate the value of sample cost accurately. Though in present practices, multi-factor regression models are trained to estimate the current cost of the operation and the influence of prior cost observations is also considered, and this design is still far from practical values. How to design an estimator to give an entirely accurate estimation of the sensing cost is not the focus of the CCS system. Thus, the synthetic cost maps generated by the cost function are widely used [8], for example, i.i.d. with dynamic cost map, monotonic with dynamic cost map, and spatial correlated with dynamic cost map.

2.5 Quality Assessment The quality assessment is significant in CCS, since we need to know whether the selected cells can provide enough information to recover the missing data under the data quality constraint. In randomized sampling strategies, the RIP is widely used to prove theoretical guarantees on recovery accuracy for a given sampling strategy [8]. In addition, Xie et al. [10] assessed the data quality from the perspective of solvable linear system. To make the linear system solvable, it is required that the unknown vertex is connected with at least k known vertexes through k edges. In this subsection, leave-one-out statistical analysis (LOO-SA) is introduced to assess the inference quality. First, a leave-one-out resampling mechanism is implemented to obtain the set of (inferred, true) data pairs. Then, by comparing the inferred data with the corresponding true collected data, Bayesian inference or Bootstrap analysis is leveraged to assess whether the current data quality can satisfy the predefined .(, p) quality requirement. Leave-one-out is a popular resampling method to measure the performance of many prediction and classification algorithms. Suppose that we collect sensory data from .m out of all the m cells, the idea of LOO is that for each time, we leave one observation out and infer it based on the rest .m − 1 observations by using compressive sensing algorithms. After running this process for all .m observations, we get .m predictions accompanying with the .m true observations, as follows: x = x1 , x2 , . . . , xm  , y = y1 , y2 , . . . , ym  .

.

(5)

Based on the .m (inferred, true) data pairs, we can use Bayesian inference or Bootstrap analysis to estimate the probability distribution of the inference error .ε, e.g., mean absolute error (MAE) to help quality assessment. Actually, satisfying the .(, p) quality can be converted to calculate the probability of .εk ≤ , i.e., .P (εk ≤ ), for the current cycle k. If .P (εk ≤ ) ≥ p can hold for every cycle k, then .(, p)

Cost-Quality Aware Compressive Mobile Crowdsensing

231

quality is expected to be satisfied as a whole. Different from Bayesian inference (require the error metric is normal distribution), the advantage of Bootstrap is that no assumption on the distribution of the observations needs to be made. Detailed information about Bayesian or Bootstrap analysis can be found in reference [16].

2.6 Problem Formulation Based on the previous models, we define the general research problem of compressive crowdsensing and focus on the cell selection. The cell selection problem can be formulated as Eq. (6): Given the CCS task with m cells and n cycles, the total all , the CCS organizers attempt to select a subset of cells that sensing budget .Bcost consume less cost and bring less inferred error in the recovery of missing data under the cost and quality constraints.   min fq (S) , fc (S)   s.t., cq (S) , cc (S) . .

(6)

The S is the cell selection matrix that records which cells are selected. .fq (S) and .fc (S) denote the optimization objectives in terms of data quality and cost, respectively. The objectives can be specified as

.

fc (S) :

m i=1

ˆ G) fq (S) : error(G, n S[i, j ]C[i, j ], j =1

(7)

ˆ G) denotes the inferred error between where C denotes the cost map and .error(G, ˆ the inferred matrix .G and ground truth matrix G. .cc (S) indicates the constraint of sensing cost that can be specified as m n .

i=1

j =1

all S[i, j ]C[i, j ] ≤ Bcos t.

(8)

The .cq (S) indicates the constraint of data quality of selected cells. When the LOO-SA is employed to assess the quality, the .cq (S) can be specified as |{k|εk ≤ , 1 ≤ k ≤ n}| ≥ n · p .

ˆ k], G[; , k]). where εk = error(G[:,

(9)

As we cannot foresee the ground truth matrix .Gm×n for a compressive crowdsensing task, it is impossible to obtain the optimal cell selection matrix .Sm×n in

232

Y. Zhao et al.

reality. To overcome the difficulties, the advanced cell selection methods will be elaborated in the following section.

3 Advanced Cell Selection Strategies in CCS 3.1 Randomized Sampling Strategy Random sampling is one basic and popular approach used in compressive sensing due to its stability in performance and simplicity in practice. By employing the basic compressive sensing method into crowdsensing, Xu et al. [10] designed the cost-aware compressive sensing (CACS) by taking the cost diversity of cells into consideration. The key technical component of CACS is to formulate the regularized column sum (RCS) as a practical metric for predicting recovery accuracy. Based on the RCS, CACS devised a convex-optimization-based method (RCS-constraint Optimization) for obtaining the randomized sampling scheme with the least cost using relaxation. The details will be introduced in the following section.

3.1.1

Recovery Accuracy Prediction Based on Regularized Column Sum

To balance the recovery accuracy and sampling cost, the two factors need to be quantitatively analyzed. The existing compressive sensing works used the restricted isometry property (RIP) to provide provable recovery lower bounds. However, the classical method has proven to lack practicality [8] because it is NP-hard to verify the RIP condition. Besides, the RIP only provides an insufficiently loose lower bound on recovery accuracy. To conquer the limitation of RIP, CACS proposed a probability version of RIP, that is statistical RIP (StRIP), which is more computationally efficient with a good performance. To verify StRIP, the bound column sum condition needs to be satisfied, which can be reformulated in terms of RCS. Definition 1 (Regularized Column Sum) Let . ⊆ {1, · · · , n} denote the index set of a sample instance, and .|| = m. Let .F = {f (j )} denote the partial Fourier matrix containing the rows indexed by .. The regularized column sum (RCS) of .F is  2     .σ (F ) = fi,j  , max logm    j =2,··· ,n

(10)

i∈

where .fi,j is the element in .i th column, .j th row of .F . By using Lemma 1 and Theorem 1 in [8], it can be proved that a sensing matrix with a smaller RCS has better chance to satisfy RIP. Moreover, for a given sparsity

Cost-Quality Aware Compressive Mobile Crowdsensing

233

k, if .σ (F ) is small enough, with an overwhelming probability, the recovery error in .l2 -norm is upper bounded.

3.1.2

CACS via Convex Optimization

Based on the RCS constraint, CACS employs the convex-optimization-based method to find a satisfying sampling matrix with the least cost. Given a certain level of recovery accuracy requirement, a constant .α can always be found such that the required recovery accuracy is met with high probability when .σ (F ) ≤ α. However, the computational complexity of determining the sampling matrix . with least cost satisfying the constraint .σ (F ) ≤ α is exponential. Thus, the constraint is modified as .E [σ (F )]  α, and the convex optimization problem is formulated as following, to find an optimal randomized sampling strategy .π under the RCS constraint with least cost. min C T π s.t., 1T π = m 2 T 2 2 (Re(F·j )T π ) + (I m(F·j ) π )  α .

0  πi  1, i = 1, . . . , n ,

(11)

where m is the expected sample size and C is the cost map. .Re(F·j ) and .I m(F·j ) indicate the real and imaginary components of the .j th column in F , respectively.

3.2 Active Sampling Strategy with Multiple Steps Different from randomized sampling strategy, the active sampling strategies aim to select the cells containing more information actively. In this subsection, a use case study is firstly presented to illustrate the basic idea of active sampling strategy with multiple steps, then the sensing cost factors are analyzed, and the estimation method of diverse cost is also presented. Finally, the cost–quality beneficial cell selection approach is introduced.

3.2.1

Use Case Study

Figure 2 shows the basic idea of the cell selection process in a sensing cycle. Suppose the target area contains six cells and the sixth sensing cycle start currently; at first, no sensory data are collected in the sixth cycle (Fig. 2(1)). The cost–quality cell selection method works as follows:

234

Y. Zhao et al.

Fig. 2 An example of the cell selection process (6 cells and 6 cycles)

(1) First, under the given cost budget (e.g., 2 CNY) of a selection, the strategy lists all possible solution combinations (i.e., here we have 7 possible solutions within the cost budget, and they are single cells 1 to 6 and the combination of cells 1 and 6), compares the solutions by considering the tradeoff between information and sample costs, selects the first two salient cells (cell 1 and cell 6) under the budget, i.e., 2, and allocates the sensing tasks to two mobile users in cell 1 and cell 6, respectively. Mobile users perform the sensing tasks and return the sensory data to the MCS server (Fig. 2(2)). (2) Then, given the collected sensory data, the quality assessment module decides if the data quality satisfies the predefined .(, p) quality requirement. If the data quality does not meet the quality requirement, we have to select more cells for sensing (choose cell 5 to sense in Fig. 2(3)). In this way, the strategy continues to allocate tasks to new cells and collects sensory data (illustrated in step (1)) and more details will be introduced in the section of cell selection, until the quality of the collected sensory data satisfies the predefined requirement. (3) Given the collected sensing values, the compressive sensing module infers the values of unsensed cells.

3.2.2

Cost Estimation

Since the cost of obtaining a specific sample in practical CCS systems depends highly on the location, time, condition of the device, human expectation, density

Cost-Quality Aware Compressive Mobile Crowdsensing

235

of participants, moving distance, and many other factors, the cost diversity in the cell selection process needs to be considered. In the following, many types of costs are discussed, including routing cost, measurement cost, perception cost, and their combination. Then, a new cost function is introduced to estimate the sensing cost.

Cost Factors In practice, different types of costs often occur in CCS systems, including but not limited to: (1) routing cost; (2) measurement cost; (3) perception cost. Routing Cost Consider such a scenario, when mobile users in some cells are insufficient, so both the costs of moving from the present location to the target location and that of making measurements at cells need to be considered.The cost function .c(V) is used to denote such a cost, defined as .c(V) = cR (V) + υ∈V cυ , where .cR (V) is the cost of the shortest walk to visit each selected cells in .V at least once. Note that .cR (V) is generally non-submodular and cannot be exactly computed in polynomial time. It will be discussed in detail in the following subsection. Measurement Cost The .cυ is used to denote the cost of collecting measurements in each cell .υ ∈ V. Generally, this kind of cost consists of energy consumption and data consumption on sensing devices as well as data management cost. Devices consume energy in both measuring and reporting a sample, e.g., locate a GPS signal and report position. This cost depends on the location as well as the status of the device. The reporting cost may depend on the network, i.e., Wi-Fi, 3G or 4G, the signal strength, variability to the network, and the congestion level. Meanwhile, the reporting may incur cellular data cost when using cellular networks. Also, the submitted data stored in Cloud or network and the quality control of data incur a management cost. Perception Cost Finally, mobile users may have different perceptions of a given cost. In other words, this cost is a subjective evaluation of the provided services. For example, a user carrying on a smartphone with a full battery may not care about the energy consumption for GPS locating to be a large cost, whereas other users may be more sensitive to the same amount of energy usage when they are running out of the battery of smartphones. Such perception-based cost adjustments .cb should be considered as they are important to user experience in MCS applications.

Cost Function Here a cost function is introduced to estimate the value of sample costs when the actual cost is hard to acquire. To estimate the measurement cost, outdoor measurements of GPS energy consumption can be conducted using smartphones at hundreds of subareas across the target region, and record the data consumption of 3G/4G/5G cellular network at those locations in the meantime. Further, the

236

Y. Zhao et al.

Fig. 3 Three different types of cost maps on the traffic dataset. (a) i.i.d with dynamics cost map (Cost Type 1, CT1). (b) Monotonic with dynamics cost map (Cost Type 2, CT2). (c) Spatial correlated with dynamics cost map (Cost Type 2, CT2)

remaining battery level of a device is considered naturally as a type of “perception cost”: the lower the remaining battery, the more valuable it is, the higher cost it should be assigned. The .cb = Bu 1−b is defined as the perception-based cost for the remaining battery, where b is the ratio of the remaining battery, and .Bu is a constant. In particular, as b goes to zero, the cost is high and approaches .Bu quickly. The intuition is that when b is large, mobile users are not sensitive and thus the measurement cost dominates. On the other hand, when b is small, users are sensitive, and thus this factor will contribute a lot to the final cost. Therefore, we choose a multiplication combination of measurement cost and perception cost plus routing cost as the overall sensing cost function .cυ · cb + .c(V). Prior study [8] proved that synthetic cost maps based on the overall cost estimation function are also feasible for performance evaluation when it is hard to conduct practical measurement and estimation. Therefore, three synthetic cost maps are generated based on the proposed cost function, and they are i.i.d. with dynamic cost map, monotonic with dynamic cost map, and spatial correlated with dynamic cost map. Here, we take the final dynamic cost map on traffic dataset as an example and show the cost distribution on different subareas over time (in a day) in Fig. 3. In the rest of the paper, we use CT1, CT2, and CT3 to refer to i.i.d. with dynamic cost map, monotonic with dynamic cost map, and spatial correlated cost map, respectively. As we see in Fig. 3, the three sampling cost maps present different changing characteristics (randomness in CT1, monotonicity in CT2, and spatial correlation in CT3), with darker color indicating larger cost.

3.2.3

Cost–Quality Beneficial Cell Selection

With the cost and informativeness assessment method above, in this subsection, the diversity of sample cost is incorporated into the cell selection process, and two selection strategies, namely GCB-GREEDY and POS, are proposed to balance the two objectives at meantime: minimize the sensing cost and maximize the

Cost-Quality Aware Compressive Mobile Crowdsensing

237

informativeness in the collected cells. The detailed strategies of cell selection are formulated as follows: (1) The generalized cost–benefit greedy selection strategy (GCB-GREEDY) Generally, the cell selection problem tries to select a subset .Sj (salient cells) from the cell set .V with an objective function .fq (S) (information function) and a constraint of the subset size (select one by one) in each cycle j. Therefore, the previous cell selection problem can be formalized as arg max fq (Sj )

.

Sj ⊆V

s.t. |Sj | ≤ Bsize ,

(12)

where .| · | denotes the size of a set; .Bsize is the maximum number of selected elements (the stopping criterion is decided by LOO-SA). But in a cost–quality beneficial selection, the constraint of subset size should be transformed into the one . At the core of the GCB-greedy algorithm budget constraint as .fc (Sj ) ≤ Bcost is the following heuristic: in each iteration k, add to the set .Sj an element .υk such that: υk ← arg maxυ∈V\S k−1

.

fq (Sjk−1 ∪ υ)

j

fc (Sjk−1 ∪ υ)

,

(13)

where .Sj0 =∅ and .Sjk ={υ1 , . . . ,υk}. Since the routing cost is usually ignored due to the sufficient participants assumption in the cost function, the cell selection problem is transformed into a problem of maximizing a monotone submodular function .fq with a monotone approximate cost constraint .fc . The corresponding GCB-greedy algorithm is shown in Algorithm 1. It iteratively selects one subarea .υ to sense such that the ratio of the marginal gain on .fq and .fc by adding .υ is maximized. Algorithm 1 The GCB-greedy-based cell selection algorithm Require: A monotone objective function, fq ; A monotone approximate cost function, fc ; one . The budget constraint, Bcost Ensure: one . The solution Sj ⊆ V with fc (Sj ) ≤ Bcost 1: Let Sj = ∅ and V = V ; 2: repeat f (S ∪υ)−f (S ) 3: υ * ← arg maxυ∈V fqc (Sjj ∪υ)−fqc (Sjj ) ; one then S = S ∪ υ * end if; 4: if fc (Sj ∪ υ * ) ≤ Bcost j j   * 5: V = V \{υ }; 6: until V = ∅ one fq (υ); 7: Let υ * ← arg maxv∈V;fc (υ)≤Bcost 8: return .arg maxS  j ∈{Sj , υ * } fq (S  j ) and S  j .

238

Y. Zhao et al.

(2) The Pareto optimization selection strategy (POS) Inspired by the solutions in [17], the cell selection problem in Eq. (12) can be reformulated as optimizing a binary vector. We introduce a binary vector .s ∈ {0, 1}m to indicate the subset membership, where .si = 1 if the i-th element in .V is selected in a sensing cycle, and .si = 0 otherwise. So the cell selection problem can be formulated as a bi-objective minimization model:

.

  arg mins∈{0,1}m −fq (s), fc (s) , one + 1 −∞ ifs = {0}m or fc (s) ≥ Bcost fq (s) =   , i j Sij · Iij otherwise.   fc (s) = i j Sij · Cij

(14)

where .|s| denotes the number of 1s in .s; .Sij denotes the entry in cell selection matrix .Sm×n ; .Iij represents the information of cell i in sensing cycle j; .Cij is one is the cost budget the approximate sample cost of cell i in sensing cycle j; .Bcost in one selection, which is set as the maximal cost value of unsensed cells; .fq one + 1 is set to .−∞ to avoid trivial or over-cost solutions. We use the value .Bcost one instead of .Bcost in the definition of .fq as this gives the algorithm some look one would ahead for larger constraint bounds. However, every value of at least .Bcost work for the theoretical analysis. The only drawback would be a potentially larger population size that influences the runtime bounds. The bi-objective optimization model performs active selection to maximize the informativeness and meanwhile to minimize the sample costs of the selected cells. A recently proposed Pareto optimization for monotonic constraints (POMC) algorithm [18] is employed to solve this problem. POMC is an evolutionary style algorithm that maintains a solution archive and iteratively updates the archive by replacing some solutions with better ones. It is also known as Global SEMO in the evolutionary computation literature [19], shown in Algorithm 2.

3.3 Active Sampling Strategy Based on Bipartite Graph To minimize the sensing cost while accurately recovering the missing data, the active sparse MCS (AS-MCS) scheme was proposed, which uses bipartite graph to represent the matrix factorization and devises the sampling strategy based on the graph. In this section, the bipartite-graph-based matrix factorization is firstly introduced. Then, the sampling strategy aiming at forming a complete and robust linear system is presented.

Cost-Quality Aware Compressive Mobile Crowdsensing

239

Algorithm 2 The POMC-based cell selection algorithm Require: A monotone objective function, fq ; A monotone approximate cost function, fc ; one ; The budget constraint, Bcost The number of iterations, T. Ensure: one . The solution s ∈ {0, 1}m with fc (s) ≤ Bcost 1: Let s = {0}m and P = {s}; 2: Let t = 0; 3: while t < T do 4: Select s from P uniformly at random; 5: Generate s  by flipping each bit of s with probability 1/m; 6: if ∃z ∈ P such that z  s  then 7: P = (P \{z ∈ P |s   z}) ∪ {s  }; 8: end if 9: t = t + 1; 10: end while one fq (s) and s. 11: return arg maxs∈P ;fc (s)≤Bcost

3.3.1

Representing Matrix Factorization Based on Bipartite Graph

In CCS systems, the matrix factorization and completion are used to infer the missing data as shown in Sect. 2.2. To further investigate the relationship between the entries in factor matrix L and .R T , the bipartite graph is employed, as shown in Fig. 4. .li and .rj denote the i-th row and the j-th row of matrix L and R, respectively. As shown in the figure, each sampled data can form an equation as .mij = li rjT . Therefore, the bipartite graph can be built as .BG = {Vl , Vr , E}, in which .Vl and .Vr denote the row vector of factor matrices. If the entry .mij is sampled, an edge connecting the .li in .Vl and .rj in .Vr is added into the set E. Thus, each sample corresponds to one edge in the graph. As shown in Fig. 4, the rank of the monitoring matrix is set as .k = 2; thus the factor matrix has 2 columns, and the sensing campaign has already performed three cycles. The row vectors in the L and .R1:3 are known since enough historical data have been obtained. In the upcoming cycle .t4 , the CCS system needs to select several cells to sense and uses the collected data to calculate the new vector .r4 in R. Since there are two unknown parameters in .r4 , at least two samples are required to be obtained in cycle .t4 , so the linear system containing two equations can be formed to solve the two unknowns, as shown in Eq. (15). .

mi4 = li1 r41 + li2 r42 mi  4 = li  1 r41 + li  r42

(15)

240

Y. Zhao et al.

Fig. 4 Matrix factorization and the representation based on bipartite graph

where the two taken samples are assumed as .mi4 and .mi  4 . In terms of the bipartite graph, the new vertex .r4 needs to have at least two edges connected to other vertex in .Vl . The linear system in Eq. (15) can be further converted into matrix form, as shown in Eq. (16).

.

li1 li2 li  1 li  1



r41 r42



=

mi4 mi  4



L∗ r ∗ = m ∗ .

(16)

The coefficient matrix .L∗ consists of the row vectors of all selected cells, .r ∗ is the unknown vector, and .m∗ contains the samples. Considering the cost diversity of different cells, each edge in BG can be associated with a weight to reflect the cost on sensing the cells. Then, to determine the unknown vertex .r4 with minimal cost, the vertexes in .Vl can be sorted in ascending order according to their weight and the top-k vertexes will be selected. However, this intuitive strategy may lead to an incomplete and unstable linear system, which fails to provide stable and unique solution. In next section, two sampling strategies will be introduced to form a complete and robust linear system.

3.3.2

Sampling to Form a Complete and Robust Linear System

∗ Given a linear system as .L∗ r ∗ = M ∗ , to determine ∗ the unknown vector .r with k unknown components, the condition .rank L = k should be satisfied, so a complete linear system can be formed. Therefore, to construct a complete linear

Cost-Quality Aware Compressive Mobile Crowdsensing

241

system while minimizing the number of selected samples, the first sampling strategy is designed as follows: Sampling Selection Strategy 1 For a candidate cell, it would be selected if its addition can increase the rank of the corresponding matrix .L∗ of the linear system built so far. The strategy 1 is easy to implement since the cell selection only needs to know part of the factor matrices corresponding to historical data rather than the sampling value. In the sensing process, the sample data may suffer from noise and even loss due to the flaw of sensing devices or unreliable transmission link. The small perturbation of the sample may lead to significant difference of the data inference. Thus, ASMCS proposed another sampling strategy to select the cells that can form a robust linear system. From the geometric view of the linear system, when the cosine values between the row vectors of the coefficient matrix .L∗ are small, the linear system is robust to the data with error. Therefore, the second sampling strategy is designed as follows: Sampling selection strategy 2 For a candidate cell, it can be selected if the similarity between cosine value of the new candidate cell with all cells already selected is smaller than a preset threshold. Combining the two sampling strategies and considering the cost diversity, the complete solution is given as follows: (1) To minimize the cost of sensing, the candidate cells are sorted according to their associated weight in ascending order, and the cell that has small degree of its corresponding vertexes in .Vl is ranked ahead among the cells with same weight. (2) The candidate cells are checked one by one to see if the corresponding linear system to be complete and robust until the k cells are selected, according to sampling selection strategies 1 and 2.

4 Evaluation In this section, the evaluation of the proposed method in [13] is briefly introduced on four real-life datasets and three cost maps.

4.1 Experimental Setup 4.1.1

Datasets

Four real-life sensory datasets are adopted, Birmingham-Parking [20], DataFountain competitions, TaxiSpeed [21], and SensorScope [22] to evaluate the applicability of

242

Y. Zhao et al.

Table 1 Statistics of four evaluation datasets Project City Cycle length Duration Cell size Number of cells Mean .± Std.

Parking BirminghamParking Birmingham 1h 76 days Car parks 30 0.5155 .± 0.2597

Flow DataFountain competitions Beijing 1h 12 days 200 m .× 200 m 68 8.84 .± 13.68

Traffic TaxiSpeed

Humidity SensorScope

Beijing 1h 4 days Road segments 100 13.01 .± 6.97 m/s

Lausanne 0.5 h 7 days 50 m .× 30 m 57 84.52 .± 6.32 %

active sampling strategy with multiple steps. The datasets contain various types of sensory data in representative IoT applications, such as parking occupancy rate, flow index, traffic speeds, and humidity. The detailed statistics of the four datasets are shown in Table 1. Parking (occupancy rate): The car park occupancy rate readings are sensed in the Birmingham-Parking project, collected from 32 different car parks for two months and 16 days with a sensing cycle at 60 min. Since the occupancy rates are only recorded for eight hours every day, a target area that has 30 car parks with valid readings (from 8:00 am to 4:00 pm) is leveraged. In this scenario, we take the car parks as the cells. Flow (passenger index in a region): The dataset is provided by the DataFountain competitions for predicting the future passenger index in Beijing. Specifically, the flow index readings are sensed during the outbreak of the COVID-19 from 100 different types of key regions. The sensing lasts for 30 days from 2020-01-17 to 2020-02-15 with a collecting cycle at 60 min, and the target region is initially divided into 997 cells with an equal size at .200 m × 200 m. Only 68 cells with valid values for 12 days are leveraged here. Traffic (speed): The speed readings of taxis are collected for road segments in the TaxiSpeed project in Beijing. The project lasted for 4 days from 2013-0912 to 2013-09-15. Specifically, this dataset contains more than 33,000 trajectories collected by GPS on taxis. And each sensing cycle lasts for 60 min. According to [21], we consider the road segments as the cells, and a target area that has 100 road segments with valid sensed values is selected. Humidity: The humidity readings are sensed in the SensorScope project, collected from the EPFL campus with an area about .500 m × 300 m for 7 days (from 2007-07-01 to 2007-07-07). Each sensing cycle lasts for 30 min. In the experiments, the target area is divided into 100 cells with each cell size .50 m × 30 m. Since only 57 cells are deployed with valid sensors, we just utilize the sensory data at the cells with valid readings.

Cost-Quality Aware Compressive Mobile Crowdsensing

4.1.2

243

Baselines

The proposed cell selection strategies are compared with two baselines: SIMPGREEDY and QBC. SIMP-GREEDY: Since there is typically a conflict between the informativeness and sample cost in a cell, the most straightforward strategy is to simply divide the informativeness by the sample cost. Thus, we can have the selection strategy as .

arg maxυ∈V\Sj fq (υ)/fc (υ).

(17)

This strategy transforms a bi-objective problem into maximizing the single objective .fq (υ)/fc (υ) in each selection, which provides a simple solution for cost– quality beneficial selection, but it may fail when one of the two factors dominates the other [14]. Hence, SIMP-GREEDY is considered as a baseline. QBC: Previous works [11, 16] have proven the feasibility of applying QBC into cell selection process. Some “committee members” are contained in QBC to determine which salient cell to sense in the next task. More specifically, the “committee” is formed by different data inference algorithms, such as spatialtemporal compressive sensing, compressive sensing, KNN, and SVD. Finally, it chooses the cell where the inferred data of various algorithms have the largest variance as the next selection for sensing without considering the cost diversity. In other words, QBC tries to minimize the total costs of selecting cells by selecting the unsensed cells with the largest variance. Therefore, QBC is suitable as a baseline.

4.2 Experimental Results 4.2.1

Errors of Inferred Value

The average inference error is firstly compared, i.e., MAE brought by different cell selection strategies while changing the number of selected cells for each cycle without considering .(, p) quality. As exhibited in Fig. 5, similar tendencies are observed over four types of sensing tasks. As the increment of the number of selected cells in each sensing cycle, the average inference errors drop rapidly. The fact implies that more information brought by the increasing selected cells promotes the accuracy of data inference. Note that the information modeling of the proposed strategies, i.e., POS and GCB-GREEDY and the baseline SIMP-GREEDY, is based on QBC, and thus they share the comparable error levels theoretically. This fact is also confirmed by the experimental results though the inference error of the proposed strategies is better than that of the baselines in many circumstances. Next, the performances of the proposed cell selection strategies will be evaluated and discussed by considering .(, p) quality, which is more practical in real-world applications.

244

Y. Zhao et al.

Fig. 5 Average inference error of different sensory datasets under the condition of a fixed number of selected cells for each cycle without considering .(, p) quality. The X-axis denotes the fixed number of selected cells, while the Y-axis represents the corresponding average inference errors. As shown in this figure, the proposed strategies and baselines share the similar error levels since we do not leverage the quality assessment module, and these strategies are all based on the same information modeling methods. (a) Parking. (b) Flow. (c) Traffic. (d) Humidity

4.2.2

The Number and Total Costs of Selected Cells

Then we focus on analyzing the research objective—how much sample costs could the proposed algorithms save while further obtaining more informativeness to reduce the inference errors? The proposed strategies are compared to the baselines from three aspects: costs, selected cells, and inference errors on four real-life datasets. On the Parking occupancy rate sensing, for the predefined .(, p) quality, we set the error bound . as 0.1 and p as 0.9 or 0.95. In other words, we require the inference error smaller than 0.1 for around 90% or 95% of cycles. The average number of selected cells for each cycle is shown in Fig. 6a, where the baseline QBC always selects the fewest cells on three different cost maps, while GCB-GREEDY and POS can select 0.5%–4.2% (on average 2.8%) and 0.6%–5.2% (on average 3.5%) more subareas than QBC, respectively. Except for the circumstance of CT1 (95%), SIMP-GREEDY also selects a bit more cells (0.21%–0.9%, on average 0.5%). Note that in CT1 (i.i.d. with dynamics cost map), the proposed strategies select more cells. The phenomenon can be explained by the statistics of cost maps since CT1 has more sample cost with small values. So the proposed algorithms may choose more than one cell in a selection. In general, GCB-GREEDY and POS only need to select on average 6.67 (7.23), 6.74 (7.26) out of 30 cells for each sensing cycle to

Cost-Quality Aware Compressive Mobile Crowdsensing

245

Fig. 6 The number and total costs of selected cells for Parking, Flow, Traffic, and Humidity under the condition of considering .(, p) quality. The X-axis denotes different types of cost map; the Y-axis represents the number of selected cells in (a), (c), (e), (g), while the Y-axis represents total costs of selected cells in (b), (d), (f), (h)

246

Y. Zhao et al.

ensure the inference error below 0.1 in 90% (95%) of cycles, respectively. Though more cells are selected by the proposed strategies, the total costs of the proposed strategies outperform those of the baselines. Generally, QBC costs the most, while POS saves the most costs, as shown in Fig. 6b. Specifically, GCB-GREEDY and POS spend 1.6%–9.1% (on average 4.7%), 2.1%–11.2% (on average 5.7%) fewer costs than QBC. Meanwhile, the proposed strategies perform better than SIMPGREEDY on cost saving in practically all circumstances. Especially in CT1, the proposed algorithms can achieve the best performance. Due to the simple greedy heuristic, SIMP-GREEDY cannot ensure a full superiority over QBC. Note that in the case of CT3 (90%), it even spends more cost than QBC. For the Flow and Traffic dataset, we observe a similar tendency in Fig. 6c–f. It is noteworthy that the proposed strategies achieve better performance than the baselines since they leverage a more complex mechanism to balance the sample cost and information. Specifically, POS and GCB-GREEDY select more cells and save more costs compared to those in parking sensing tasks since the average number of selected cells in a time cycle becomes larger. Also, the inference error of the proposed strategies is obviously reduced. On the Humidity dataset shown in Fig. 6g and h, the proposed strategies POS (GCB-GREEDY) can reduce inference errors by 6.1% to 10.1% (5.7% to 8.5%) compared with QBC, and 0.8% to 2.7% (0.2% to 2.2%) compared with SIMPGREEDY. Also, the proposed strategies POS (GCB-GREEDY) reduce the sample cost by 1.8% to 15.2% (1.4% to 15.02%) compared with QBC, and 1.0% to 8.5% (0.6% to 7.4%) compared with SIMP-GREEDY. The proposed cell selection method explicitly outperforms the baselines, with two strategies (Pareto optimization and generalized cost–benefit greedy) from three aspects: less sample cost, more selected cells with sensing values, and less inference error, since a complex mechanism is leveraged to minimize the total cost and maximize the informativeness. Besides, the results on different cost map implicate that the proposed cell selection strategy is able to handle various kinds of cost inconstancy, especially when the cost map has a bigger range and standard deviation.

5 Summary In this chapter, the compressive crowdsensing as a promising sensing paradigm is introduced. In CCS, only few cells are selected to sense, in which the cost and data quality are two major concerns. Therefore, three kinds of advanced cell selection strategies in CCS are presented: randomized sampling strategy, active sampling strategy with multiple steps, and active sampling strategy based on bipartite graph. In those strategies, the compressive sensing method is not only used in the data recovery after sampling but also used in the cell selection to assess the importance of cells. Finally, the evaluation on four real-life datasets demonstrates the feasibility and advantages of the cost–quality-aware cell selection strategy.

Cost-Quality Aware Compressive Mobile Crowdsensing

247

References 1. H. Sun, W. Shi, X. Liang, Y. Yu, IEEE Internet Things J. 7(2), 800–817 (2019) 2. E. D’Hondt, J. Zaman, E. Philips, E.G. Boix, W. De Meuter, in Proceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing (2014), pp. 727–738 3. Z. Liu, S. Jiang, P. Zhou, M. Li, IEEE Trans. Intell. Transp. Syst. 18(10), 2851 (2017) 4. H. Aly, A. Basalamah, M. Youssef, IEEE Trans. Mob. Comput. 16(10), 2712 (2016) 5. J. Wang, L. Wang, Y. Wang, D. Zhang, L. Kong, IEEE Internet Things J. 5(5), 3747 (2018) 6. J. Wang, Y. Wang, D. Zhang, F. Wang, H. Xiong, C. Chen, Q. Lv, Z. Qiu, IEEE Trans. Mob. Comput. 17(9), 2101 (2018) 7. S. Devarakonda, P. Sevusu, H. Liu, R. Liu, L. Iftode, B. Nath, in Proceedings of the 2nd ACM SIGKDD International Workshop on Urban Computing (2013), pp. 1–8 8. L. Xu, X. Hao, N.D. Lane, X. Liu, T. Moscibroda, in Proceedings of the 14th International Conference on Information Processing in Sensor Networks (2015), pp. 130–141 9. X. Hao, L. Xu, N.D. Lane, X. Liu, T. Moscibroda, in 2017 16th ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN) (IEEE, 2017), pp. 29–40 10. K. Xie, X. Li, X. Wang, G. Xie, J. Wen, D. Zhang, in Proceedings of the 2019 International Conference on Management of Data (2019), pp. 195–210 11. L. Wang, D. Zhang, A. Pathak, C. Chen, H. Xiong, D. Yang, Y. Wang, in Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous Computing (2015), pp. 683–694 12. W. Liu, L. Wang, E. Wang, Y. Yang, D. Zeghlache, D. Zhang, Comput. Netw. 161, 102 (2019) 13. Z. Zhu, B. Chen, W. Liu, Y. Zhao, Z. Liu, Z. Zhao, IEEE Internet Things J. 8(5), 3831 (2020) 14. S.J. Huang, M. Xu, M.K. Xie, M. Sugiyama, G. Niu, S. Chen, in Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (2018), pp. 1571– 1579 15. W. Liu, E. Wang, Y. Yang, J. Wu, in Proc. of the IEEE International Conference on Computer Communications (INFOCOM 2022), vol. 33 (2022) 16. L. Wang, D. Zhang, Y. Wang, C. Chen, X. Han, A. M’hamed, IEEE Commun. Mag. 54(7), 161 (2016) 17. C. Qian, Y. Yu, Z.H. Zhou, in Advances in Neural Information Processing Systems (2015), pp. 1774–1782 18. C. Qian, J.C. Shi, Y. Yu, K. Tang, in IJCAI, vol. 17 (2017), pp. 2613–2619 19. C. Qian, Y. Yu, Z.H. Zhou, Artif. Intell. 204, 99 (2013) 20. A. Camero, J. Toutouh, D.H. Stolfi, E. Alba, in International Conference on Learning and Intelligent Optimization (Springer, 2018), pp. 386–401 21. J. Shang, Y. Zheng, W. Tong, E. Chang, Y. Yu, in Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2014), pp. 1027–1036 22. F. Ingelrest, G. Barrenetxea, G. Schaefer, M. Vetterli, O. Couach, M. Parlange, ACM Trans. Sens. Netw. 6(2), 1 (2010)

Part V

Key Technical Components: Security and Privacy

Information Integrity in Participatory Crowd-Sensing via Robust Trust Models Shameek Bhattacharjee and Sajal K. Das

1 Introduction Rapid advances in smartphones, IoT devices (smart vehicles, watches), and the adoption of Android, iOS apps., have led to the emergence of Mobile CrowdSensing [1] (MCS) paradigms. Specifically, in MCS paradigms, a crowd of users known as participants submit observations termed as reports (viz., event, image, audio, analog sensor values) about any phenomenon in their immediate environment, to a crowd-sensing server (MCS server). The MCS server synthesizes such reports to infer a meaningful summary statistic known as an event [2]. Such events are either directly published on the MCS App as “actionable information” for other non-participant users (consumers) or used indirectly to influence decisions in the physical world that improve the quality of life in smart cities. Intertwining relationships of the MCS paradigm with the physical world create complex dependencies with cyber-physical systems (CPS) in smart cities [3]. The MCS paradigm can be broadly classified into two sub-domains: (i) Opportunistic (implicit) MCS and (ii) Participatory (explicit) MCS. Opportunistic Vehicular MCS In O-MCS, a user subscribes to a crowd-sensing provider and enters into an agreement that her device(s) will act as a surrogate sensor that automatically captures sensing data “without” the need for explicit human involvement. In this domain, most commonly, the crowd-sensed inputs are

S. Bhattacharjee () Department of Computer Science, Western Michigan University, Kalamazoo, MI, USA e-mail: [email protected] S. K. Das Department of Computer Science, Missouri University of Science and Technology, Rolla, MO, USA e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 J. Wu, E. Wang (eds.), Mobile Crowdsourcing, Wireless Networks, https://doi.org/10.1007/978-3-031-32397-3_10

251

252

S. Bhattacharjee and S. K. Das

continuous-valued analog inputs. Therefore, this domain is analogous to a sensor network. Consequently, research in this area uses statistics, machine learning, and Bayesian inference for estimating the true value of a certain physical quantity of interest. This process is known as truth discovery. In the context of smart transportation, the opportunistic vehicular crowd-sensing automatically collects ambient speeds, direction, and time-stamped geo-location from either hand-held mobile devices or from TMC sensors that are deployed on road segments of the city. This is not the focus of this work. Participatory Vehicular MCS (P-MCS) P-MCS requires explicit human involvement, where the concerned participant manually enters contributions in the form of reports, posts, or any piece of information, into dedicated vehicular crowdsensing mobile app. (e.g., Google Waze) that senses the state of the physical world. For example, Google Waze is a participatory MCS for vehicular event discovery, where participants manually report various observed traffic incidents to an MCS server via a smartphone app in exchange for certain “incentives” [3]. The MCS server aggregates these reports and takes decisions on what event has happened according to a decision-making scheme. According to the type of event, it issues recommended actions (e.g., reroute recommendations, travel times, gas prices, the presence of stalled vehicles and obstacles, accidents, weather hazards, and crime scene). Such information allows consumers to improve driving experiences through dynamic route planning in busy cities [4] and improves operations of the physical transportation infrastructure. Therefore, the above abstraction is analogous to a modern cyber-physical system. In such participatory MCS apps, one cannot share/forward other’s reports but can offer a rating/feedback on the perceived usefulness of the other user’s contributions. The participatory MCS is this work’s focus. Contributions This chapter specializes in event trust models and user reputation scoring models that are designed to mitigate vulnerabilities as well as the cold start problem in a mobile crowd-sensing platform that is participatory in nature. We also identify some design requirements that event trust and user reputation models should have to prevent event generation and feedback weaponizing attacks. In this chapter, we show how we can introduce design changes in existing models of trust and reputation that are well suited to defend against challenges and attacks in mobile crowd-sensing platform. We also explain how these design changes can mitigate the impact of false reporting and feedback weaponizing attacks on participatory crowdsensing applications. Our design changes are inspired from cognitive neuroscience, commonsense reasoning, prospect theory, and logistic regression, and we show that a unique mix of the above can be unified into one common framework. We also use a vehicular crowd-sensing as a proof-of-concept to elucidate the concepts behind the vulnerabilities and solution approach to trust and reputation scoring methods.

Information Integrity in Participatory Crowd-Sensing via Robust Trust Models

253

2 Architecture for Participatory MCS Here we introduce some terminologies related to the design of participatory MCS. More details can also be found in our previous works [2, 4, 5]. We use a proof-ofconcept for vehicular application. The architecture of the participatory MCS and our solution’s framework architecture for event trust and user reputation for preserving information integrity are summarized in Fig. 1. Physical World is the phenomena being observed. In the case of participatory MCS in a vehicular context, it is usually a selection from a set of possible events that indicates a particular event type. For example, accident, jam, crime, civic unrest, hazard, police presence, etc. In other contexts, it could be any event that will affect the aggregate statistic/decisions taken by the crowd server. Reporters These are users reporting a particular event of interest. Our observations with Waze indicate that a certain number of reports are required (depending on the location/city) before Waze decides to publish or broadcast the event. This is aligned with existing works that look for a correlation between observations from various reports, at least in the cold start phase. However, as we shall see later, such a correlation does not necessarily imply that a published event is actually truthful. Feedback Monitoring Apparatus This includes a human Rater, drones, and mobile trusted participants that act as a “verifier” of the event indicated from a set of reports. Based on the feedback as an evidence, the CS provider can build the event truthfulness score for events. However, such event truthfulness scoring should be aware of feedback weaponizing attacks and should be able to mitigate them.

Fig. 1 Architecture of Quality and Quantity (aka QnQ) Unified Trust Framework

254

S. Bhattacharjee and S. K. Das

MCS Servers This is the entity that processes all reports and converts them to events, publishes events publicly for the feedback monitoring apparatus, and deploys all models of trust, reputation, and incentive assignment. In Fig. 1, the MCS server receives reports, and rating counts as inputs from various user IDs and runs the QoI model and User Reputation Models, and the inputs and intermediate results are stored in databases. More details on Fig. 1 are added when we discuss the solution framework at various places in Sect. 4. QnQ Framework Our framework is named as QnQ since it can unify both Quantity (degree of participation) aNd Quality of information (accuracy) offered by users. Reputation scores learned by our framework from user behaviors can be used for selfish and malicious user identification, proportionally fair reputationbased incentive disbursement, and how much weight future reports from each user should be used by any weighted decision-making model for robust decision-making.

3 Security Threats and Challenges In this section, we first describe the types of dishonest behaviors that plague a participatory MCS platform. Understanding dishonest motivations are important to create a complete threat model for participatory MCS. Second, we show that operational and AI-based vulnerabilities lead to broadly two different classes of attacks called Event Generation Attacks and Feedback Weaponizing Attacks. Under each class of attack, we discuss the sub-type of attacks. Third, we discuss the cold start challenge that is not incorporated in most existing works that try to establish event trustworthiness and user reputation.

3.1 Types of Dishonest Behaviors While classifying dishonest behaviors, there could be types: (1) Selfish and (2) Malicious. Selfish users are those who generate fake events to gain something in return for themselves, whereas malicious users only aim in creating operational damage in the domain that uses the MCS. Selfish behaviors in an MCS setting can be divided into 2 subgroups of behaviors: (1) Individual Selfish Intent The most common motivation is getting more incentives/rewards from the MCS provider. In all MCS Apps, the incentives depend on the degree of participation, a requirement to keep churn. Hence, even when there is no event, these users tend to report an event to maximize rewards. This was an observation seen from our study on Waze dataset [4]. However, when there is an event, these users report truthfully as was verified in retrospect by Waze. Hence, they have a mixture of true and false behaviors. However, regardless of the degree of fake reports, such behaviors are risky for both the users and providers of MCS and need to be identified. Thus, a level of risk

Information Integrity in Participatory Crowd-Sensing via Robust Trust Models

255

aversion is required in the design if such users are identified. We will see later how this can be done in our QnQ design. (2) Collaborative Selfish Intent This could be understood by a real-life example. In Los Angeles, USA, news reports surfaced that a group of residents in a residential area colluded on the Waze App to generate fake jam reports, such that Google Waze would not reroute traffic through their neighborhood [6]. This is however only possible when selfish users are located in that particular area. Suppose a smartphone user goes out for work in the downtown area where she spends 10 h of her day. She could only false report during the evening when she is in the uptown residential area. At other times of the day, she has no selfish incentive to report false events. In most cases, the selfish user is reporting truthfully. Conversely, for a user who works from or stays at home throughout, it makes complete sense to generate numerous fake reports on jam because these reports are auto-GPS stamped. Occasionally, when she goes out, she does not produce fake reports since she does not have any incentive to produce false reports at other locations. Thus, for the same selfish objective in the same area, different users could have widely different quantities of false reports, which makes the classification of user intent challenging. But again when this person is not in that strategic area she reports truthfully. (3) Malicious Intent: These users do not have any incentive to report truthfully, and their only goal is to create some operational impact on the MCS. For example, these may be users that are being controlled by an organized and motivated adversary (that recruits malicious users/or hacks various apps) and collude to give misleading reports whenever they choose to report or rate because their only gain is to inflict maximum operational damage. For the vehicular context, an orchestrated attack can be used to create a fake jam that creates a strategic void in a certain location. Of course, all reports agree on the event type, location, and time epoch to make the attack believable. This would lead others to avoid this area. Such strategic voids could be used for criminal or other illegal activity, or a strategic void may cause unwarranted traffic reroutes and can create a traffic block in another strategic part of the city (this would depend on the road network). Malicious intent is defined by behaviors where the users do not seek any incentive or benefit for themselves apart from creating havoc in the domain for which this MCS is applied.

3.2 Cold Start Problem and Other Challenges If event truthfulness is known, it can be used to calculate user reputation. If prior user reputations are known, they can be used to calculate an event’s truthfulness. However, there are a few practical problems with these approaches in participatory MCS, which we discuss here.

256

S. Bhattacharjee and S. K. Das

Event Truthfulness Scoring Model Most common event truthfulness scoring models assume the knowledge of the prior reputation base for all MCS users is available. A survey of such methods can be found in [3]. It is true that once there is an established user reputation learned, the event truthfulness can be calculated as a weighted function of the the prior reputation of those users who reported this event [7]. However, a practical question to pose is, how does the MCS reach a stage where it possesses a reliable reputation base across a wide geographical area, after being launched? Clearly, incentives in place can generate selfish as well as malicious behaviors. When an MCS app is launched, it does not have an existing user base and hence prior reputation is unknown. This issue is the cold start problem. Hence, any event truthfulness calculated from a biased reputation of selfish and malicious users during cold start will be misleading. Therefore, again it is reinforced that a feedback apparatus is necessary to first assess event truthfulness during the cold start and then improve the reputations of those users who have contributed to truthfulness events (quality) and a significant number of truthful events (quantity). Additionally, there are methods that use spatio-temporal correlation [8], votingbased agreements [9], or consensus models [10]. A detailed survey was done in our previous work [3]. While, this works well for opportunistic MCS but does not work for participatory MCS. This is because of a feature that it does not make sense to report anything on a participatory MCS if no event has happened. Therefore, any organized selfish or malicious behavior that generates fake reports will have contextual similarities, but it does not mean that the event is true. Therefore, again we see the importance of the feedback apparatus that could help in quantifying truthfulness. One important sacrifice required is that during the cold start fake events will have to be published along with true events initially for the feedback to be on them. However, over time we can use it to refine and learn the reputation of users. After cold start, the MCS platform does not publish events blindly but only uses user reputation, context, prior likelihood to decide whether to publish an event or not [2, 7]. Reputation Scoring Models A second class of solutions finds the reputation of users, by assuming that the ground truth is available either in real time or in retrospect. Such knowledge is then used as a basis to compare the individual user’s reports with the ground truth. Any mismatch is punished, while matches are rewarded. Punishments and rewards are updated and aggregated over time to produce the final trust/reputation score of the MCS users. However, the assumption of the availability of ground truth has two practical problems: (i) getting the ground truth is often not feasible that is why one needs to rely on MCS. This undermines the true benefit of crowd-sensing; (ii) in certain cases, the ground truth may not be possible. (iii) ground truth may be too expensive to find. Naturally, in such cases, no information on the reputation of users would indicate how honest or dishonest they are. When an MCS system is launched, there is no idea whether an event indicative from the user reports is indeed, truthful.

Information Integrity in Participatory Crowd-Sensing via Robust Trust Models

257

Again we see the need for a feedback apparatus as a proxy for the ground truth’s unavailability. However, the feedback apparatus itself is vulnerable to attacks that target weaknesses in the AI-based trust scoring models that use feedback to determine event trust scores. In this chapter, we are going to give an exposition of this idea, some associated challenges, and a guide to mitigating this vulnerability.

3.3 Categories of Vulnerabilities and Attack Types: (1) Event Generation Attacks: These are attacks that happen during the reporting of events and most commonly discussed threats. There can be two possible types: • Fake Event Generation: A vulnerability of a participatory event discovery MCS is that the honest participants need not report anything in the absence of an event, and hence by default a high correlation and agreement among false reports from the adversary is implicitly guaranteed. However, this does not indicate that the event is true. Therefore, we suggested in [4] the need for rating feedback apparatus on crowd-sensing apps. This is a feature that is now found in Waze Apps. • Alter Event Type: The objective is to make the CS arrive at the wrong event type, even if an event did happen. In this case, the adversaries induce an “incorrect action” that exacerbates the consequences of the event that did happen. For example, if it is a roadblock but falsely reported as an accident, then law enforcement and paramedic vehicles will be sent out, worsening the congestion. Furthermore, paramedics will be delayed/unavailable in places that actually need them. (2) Feedback Weaponizing Attacks: The introduction of the feedback apparatus itself (for finding event truthfulness) leads to another vulnerability. The adversary weaponizes the feedback apparatus to target the AI approach that uses the feedback to calculate event truthfulness. The adversary deploys a set of users that blend themselves as part of the rating user base and give misleading feedback to the published events at the cold start phase. This is termed as “feedback weaponizing” since the design feature of feedback is used as a weapon by the adversary, to poison the body of evidence that should have helped to identify selfish or malicious participants and false events. In participatory MCS, the adversary can recruit a large number of fake raters from companies such as BestReviewApp [11] or MoPeak [12]. Additionally, the total volume of watchdogs/anchors/mobile trusted agents is typically sparse in MCS [13, 14], or sub-regions may have spatio-temporal sparseness of rater availability, allowing adversaries to dominate the feedback apparatus with lesser attack budgets:

258

S. Bhattacharjee and S. K. Das

• Ballot stuffing: A rating user gives positive feedback to an false published event generated by dishonest reporters. • Obfuscation stuffing: A rater submits uncertain feedback to a (false) published event (generated by dishonest reporters). • Bad mouthing: A rater submits negative feedback to a legitimate published event (generated by honest reporters).

4 Quality and Quantity Unified Architecture for Secure and Trustworthy Crowd-sensing In this chapter, we focus on two main contributions, the robust QoI model that is event-centric and the reputation scoring model that is user-centric. The QoI model runs for every event, while the reputation score is built based on behavior observed over multiple events on a defined time epoch over which multiple events have happened.

4.1 Robust Quality of Information Model The QoI model gives the degree of event truthfulness that assigns a trust score for each event based on the feedback received against each event. The QoI model should be designed in a way such that it can inherently mitigate the effect of feedback weaponizing attacks viz. ballot stuffing, obfuscation, ballot stuffing, without assuming a prior reputation base. The event truthfulness module has three sub-modules: (i) Posterior Estimation of Probability Masses, (ii) Non-linear Dual Weighing Module, and (iii) Prospect Theoretic Activation.

4.1.1

Posterior Estimation of Probability Masses

Regardless of the number of categories in the rating state space, a common practice for mathematical tractability is to map multiple categories into a ternary state space of belief, disbelief, and uncertainty. The rationale is that feedback systems mainly indicate three mutually exclusive states of preference: (1) approval (positive), (2) disapproval (negative), and (3) unsure (undecided). Jøsang Et. Al. [15] popularized the terms belief (b), disbelief (d), and undecided (u) to denote the posterior probability masses for respectively a positive, negative, or uncertain preference in the total evidence. The calculation is achieved through a maximum likelihood estimate with a non-informative prior. Therefore, the first step is to assign which rating categories in the feedback apparatus contribute to belief .α, disbelief .β, and undecided .μ, respectively. As an

Information Integrity in Participatory Crowd-Sensing via Robust Trust Models

259

example, in a 5-star feedback system, ratings 4 and 5 contribute to positive, ratings 1 and 2 to negative, while the rating 3 to undecided. In the Yelp app, three ratings could be provided to each review, namely useful (positive), funny (negative), and cool (uncertain). Then the total rating counts under positive .ηα , negative .ηβ , and uncertain interactions .ημ are added up such that the total number of ratings for a given k-th event is .Nk = ηα + ηβ + ημ = N. Using Bayesian inference, one can η +1 η +1 +1 ; .d = Nβ+3 ; .u = Nμ+3 , when a show that the posteriori belief masses are .b = ηNα+3 non-informative prior is used for each of the possibilities. The theoretical proof of the above can be found in our previous work [16]. Existing Works In most existing works, the event truthfulness is modeled via a Beta Reputation (for binary feedback) System [17] or a Dirichlet Reputation System [18] (for multinomial feedback) and uses the Jøsang’s Belief Model. Below we first describe the Jøsang’s belief model and then understand some of its limitations in preventing feedback weaponizing attacks and also being unable to reflect subtle differences in quality and quantity of evidence. QoI via Jøsang’s Belief Model Jøsang’s belief model [15] uses an Expected Bayesian belief (.E J ) that is given as E J = b + a.u,

.

(1)

t+1 s+1 r+1 ; .u = r+s+t+3 ; and r, s, and t denote the number ; .d = r+s+t+3 where .b = r+s+t+3 of positive, negative, and uncertain ratings received for a certain event/entity. The value .a = 0.5 is the relative atomicity that is equal to the reciprocal of the cardinality of inference state space .{true, f alse} [15]. Since the inference state space is a true or false event, it should be set as .a = 0.5. The limitation of the Jøsang’s belief is that how much uncertainty should contribute to the truthfulness score of an event and does not change with the number of ratings available.

However, in our previous work [2], we summarized that if these models are used for quantifying event truthfulness, then these will give biased QoI scores in the presence of feedback weaponizing attacks. Specifically, in our previous works, we introduce two design requirements that are missing in the above models, which is the root cause for the vulnerabilities. These two design requirements include: RQ1: Embed Quantity of Feedback Confidence Property is a requirement that says that the QoI score should be a function of not only b, and u, but the number of ratings that produce the b and u, should play a role in the QoI Scoring. Jøsang’s belief model (details in Sect. 5) fails to capture the differences in the confidence of the feedback community, thereby making the resultant expected belief (QoI in our case) more vulnerable to manipulation by malicious raters who provide positive ratings to false events (Ballot Stuffing attack) and vice versa (Bad Mouthing attack). This biases the QoI score of false events in favor of the adversaries. As shown in Table 1, each event is denoted as .E : N, r, s, t, where N is the total number of received ratings, while .r, s, t are positive, negative, and undecided ratings, respectively, and the Jøsang-based QoI values are calculated according to Eq. (1). For event E1, 3

260

S. Bhattacharjee and S. K. Das

Table 1 Limitations of Jøsang’s belief model Issues Confidence of community Not null invariant

Examples E1:.7, 3, 2, 2 E2:.70, 30, 20, 20 E3:.105, 5, 0, 100 E4:.25, 5, 0, 20

Jøsang’s QoI 0.55 0.57 0.51 0.53

out of 7 feedback are good, whereas for the event E2, 30 out of 70 are good. The Jøsang’s belief model generates almost the same QoI in both examples. From an adversary’s perspective, it is easy to compromise or manage 3 good raters in E1 and maintain the same fraction of positive ratings as E2. However, it is harder to maintain the same fraction when the crowd is large (as in E2), in which case the adversary has to manipulate 30 raters. Hence, given the same fraction of positive feedback, any event with more feedback should be considered as more trustworthy. If this feature is not incorporated, the QoI becomes more vulnerable [4]. RQ2: Embed No Null Invariance Property Jøsang’s belief models do not offer the required null invariance property, which results in unwarranted increases in the QoI scores due to a high number of undecided ratings. Such a high number of uncertain ratings could be deliberate (Obfuscation Stuffing attack) or be a result of legitimate uncertainty. In either case, such an event should not have a high QoI score. For example, event E3 in Table 1 has 100 uncertain feedback out of 105. However, it achieves almost the same QoI as event E4 which in contrast has only 20 uncertain ratings. For most services, it is common sense to think that it is not appropriate for E3 to have the same score as E4 since E3’s evidence contain a huge amount of uncertainty. Thus, the QoI scoring model needs a mathematical provision for being null invariant. Given the above two requirements, we need to introduce design changes to Jøsang’s Belief Model. In the following, we first discuss the details of QoI via Jøsang’s Belief Model. Then, we introduce our QoI model, in a way such that how the requirements are embedded on top of the traditional Jøsang’s model, becomes apparent to the reader.

4.1.2

Non-linear Weighing of Probability Masses:

In this section, we will show how QnQ’s QoI model is different from the previous methods and how it mitigates the effect of feedback weaponizing attacks by embedding both quality and quantity of the “total feedback” received against each event. QnQ’s QoI In a series of works [2, 4], we had proposed the QnQ QoI that has the following general form:

Information Integrity in Participatory Crowd-Sensing via Robust Trust Models

261

Fig. 2 Inside the QoI Module

τk = (wb ).b + (wu ).u,

.

(2)

where .0 < {wb , wu } < 1. Hence, .0 < τk < 1. The most novel part of our event truthfulness module is that it applies two nonlinear functions; belief function .wb and uncertainty function .wu , to the degrees of belief and uncertainty that is calculated from the feedback received on a published event. These functions produce weights depending on the value of N and act as knobs that regulate how much the observed b and u should contribute to the final trustworthiness of the event. Thus, in this way, the actual weights applied directly depend on the number of ratings N received for an event. The above flow is pictorially depicted in Fig. 2. Belief’s Function The functional form of .wb should be such that it intrinsically has properties to reduce the effect of ballot stuffing and bad mouthing, while .wu has properties to intrinsically reduce the effect of obfuscation stuffing. For the functional form of .wb , we recommend the use of a function that is similar to the Generalized Richard’s Curve ([19]) for this purpose. By common sense reasoning, if a fewer number of raters endorse the same event, as opposed to the one endorsed by more raters, the second one should carry more weight to the belief even if their probabilities are the same. The problem with Jøsang’s [15, 17] and Dempster–Shafer [20] beliefs is that the probability masses are ratios only. Hence, using the number of ratings as an input parameter to the Richard’s curve (.wb (N)) to the belief probability mass makes sense. Apart from N, three parameters are embedded in Richard’s curve—initial value, growth, and inflection parameters—which play a role in mitigating the effect of ballot stuffing and bad mouthing attacks. The generalized logistic curve is parameterized by an independent variable N and is given as wb (N) = L +

.

U −L , (1 + Ab .e−Bb .N )1/ν

(3)

262

S. Bhattacharjee and S. K. Das

Fig. 3 Anatomy of how belief weight mitigates ballot stuffing

where L and U are the lower and upper asymptotes, the B is the growth parameter, ν is the displacement parameter that controls at what value of N does the .wb (N ) enter into the exponential growth phase, and A is the starting value when .N = 0 that acts as prior confidence on the b. The displacement parameter .ν allows to embed “minimum enough ratings” to start believing in the observed belief; i.e., at what number of ratings should the observed belief start to contribute to the truthfulness score? The answer to enough ratings can be found through common sense reasoning as realistically how many unique raters an adversary can control after neutralizing Sybil attacks given the population of the neighborhood. From Fig. 3, the right side picture shows that the larger the value of .ν, the smaller is the number of ratings at which the function .wb starts to take enter the exponential growth phase. Therefore to be risk averse, one needs to have a smaller value of .ν, which would require a more number of ratings to be given to allow the belief mass to contribute to actual trust. The growth rate parameter .Bb allows to embed the idea of “at what number of ratings is it sufficient to completely trust the observed belief mass for a given event.” The initial value does not play too much of a role but needs to be high for .wb to start from a low baseline. In Fig. 3, it can be noted that the smaller the value of the .Bb , the larger the number of ratings it takes to reach the maximum possible value of .wb , i.e., equal to 1. Hence, for any number of ratings where .wb < 1, the observed degree of b is not fully contributing to the event truthfulness since .wb < 1. This gives the explainability behind why it is able to thwart effect of ballot stuffing. It makes sense to have the .Bb more small in a risk-averse system, since a larger number of ratings will be required to completely allow the belief from the feedback evidence to influence the learnt event truthfulness score and consequently the user reputation score as shown later.

.

Information Integrity in Participatory Crowd-Sensing via Robust Trust Models

263

The traditional machine learning approach to finding parameters of a weighing function is to take a supervised approach where the idea is to select a parameter value that minimizes a loss function. However, a large amount of data are required for accurate training, and the availability of labels is an issue during the cold start phase. The benefit of our approach is that since we use common sense reasoning for parameter selection, it bypasses this initial problem. Even though every city and every locality can have unpredictable or completely different patterns, the parameter selection can be done contextually by a system administrator based on the local patterns of how many ratings are typical in a given time context. Uncertainty Function The uncertainty function’s functional form is a combination of Richard’s Curve and Kohlrausch Relaxation Function [19, 21] which is again inspired by cognitive psychology models. How uncertainty influences human beliefs hinges around a knot point. When the quantity of evidence is not sufficient, you tend to give a benefit of the doubt (i.e., the degree of uncertainty). However, if the evidence is sufficient whatever doubt is within that evidence does not get any benefit, i.e., it stops contributing to the trustworthiness (a positive state of mind). A functional form that represents the above was required for our purpose. For a functional form that models reasoning before the knot point, Richard’s curve is used. For discounting the effect of uncertainty on the event truthfulness score, we use the Kohlrausch relaxation function [22] that is used in physics to model the behavior of systems that converge to an equilibrium after a sudden trigger. In our proposed model, the trigger is the knot point where .N = Nthres (refer to Fig. 4). After the

Fig. 4 Anatomy of reasoning under obfuscation stuffing

264

S. Bhattacharjee and S. K. Das

knot point, the .wu is calculated using the Kohlrausch relaxation function, whose parameter .0 < ϕ < 1 controls the rate of decay in .wu with increasing N. Larger values of .ϕ cause more rapid decrease from the knot point and vice versa, and this would take care of the scenario where a high number of ratings have been given, and even if most of them are uncertain, since the .wu 1, .θ1 > 0, and .0 < φ1 < 1, and .τk is the raw event truthfulness of the k-th event in the cold start phase. The advantage of using the above Eq. (8) is apparent with the reputation scoring of the selfish group. Some selfish users may contribute to the quantity of good event reports but give bad reports when it is absolutely required. To prevent such selfish

268

S. Bhattacharjee and S. K. Das Honest Selfish Malicious

Score

1 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 -1 Qu120 ali 100 ty 80 (T 60 20 0 rue 40 60 40 100 80 tal Events) Ev 20 120 en 0160140uantity (To ts) Q

(a)

1 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 -1

1 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 -1 Qu 70 ali 6050 ty 40 (T 30 10 0 rue 20 30 20 50 40 l Events) 60 Ev 10 en 0 90 80u70 ty (Tota ts) Q anti

Score

Honest Selfish Malicious

1 0.8 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 -1

(b)

Fig. 7 Differences in user reputation score .R i with varying link functions. (a) Our prospect theoretic link function. (b) Traditional logit link function

users from gaining a higher reputation than other selfish or low-activity honest users, this novel link function achieves a separate cluster in the user reputation plane that is distinct from most honest or malicious users regardless of the strength hidden in their selfish behavior quantities. This is important for fairness, lowering losses in the incentive disbursement, and discouraging opportunistic behavior, which is not well achieved by the traditional logits used in machine learning. The difference in the user reputation score distribution and clusters between the logit and prospect theoretic link function can be seen in Fig. 7 demonstrating the user reputation scoring distribution as a function of the quantity (the number of reports) and quality (the number of actual correct reports). Figure 7a shows three distinct tiers for malicious, selfish, and honest users using the prospect theoretic link function. In contrast, Fig. 7b implementing the logit link function cannot separate subtle variations in selfish behavior prevalent within the selfish group. Therefore, one selfish group is shifted more to the honest side.

4.2.2

One-Hot Encoded Sum

This step is required to get the aggregate quality of participation over the quantity of participated events for the i-th user, which is the cornerstone of quality and quantity unification in our user reputation score. There are three main steps involved in giving this unified aggregate sense of reputation as described below:   Transformed QoI Vector Let all event QoI values for all events be kept in a . Q 1XK vector from the cold start phase, for each event .k ∈ 1, K where K is the total number of events. Each .Qk is a score of .Qk ∈ [−λ1 , 1], corresponding to the k-th published event.

Information Integrity in Participatory Crowd-Sensing via Robust Trust Models

269

User Participation Encoding Vector For each user i, we need to generate a onehot encoding on whether the i-th user was involved in reporting the k-th event. If involved, the encoding gets a value of 1, or else zero. This encoding is retrieved from the user event association database that contains the

record of all users that reported a given event. At the end, we get a vector . I (i) 1XK per user, where each element .I (i) (k) is  where,

.

(i)

I

(k) =

1, 0,

If i reported event k Otherwise

(9)

User Participation Encoding Vector Now, the encoding for every user is multiplied with the .Qk vector containing the link function transformation of the event truthfulness. This process achieves a very important thing of combining both quality and quantity in the user reputation scoring model. Note that, if an user did not participate in reporting an event, the number of zeroes will be higher in his encoded vector I of dimensions .1 × K. Hence, multiplying it elementwise with .QT of dimensions i .K × 1 will yield a scalar score, which we call as .S such that:   S i = I (i)

.

1×K

  QT

K×1

(10)

such that the aggregated reputation score .Si obtained from Eq. (10) is a real number in the interval .[−∞, +∞].

4.2.3

Output Activation and Classification Criterion

To make it intuitive and consistent with the definition of trust metrics, and varying degrees of participation in different users, we need to map the value of the .S i into the same range such that differences between the three classes of user intent are distinguishable, and we use the logistic distribution function to map its values in the interval .[−1, +1]. Therefore, the final reputation score (.Ri ) of a reporter i is given as ⎧

⎪ 1 ⎪ , if Si > 0 ⎪ ⎪+ S −μ+ ⎪ − i +s ⎪ C ⎪ 1+e ⎪ ⎪ ⎪ ⎨

(11) .Ri = , 1 − , if Si < 0 ⎪ ⎪ |Si |−|μ− s | ⎪ − ⎪ ⎪ C− 1+e ⎪ ⎪ ⎪ ⎪ ⎪ ⎩ 0, if Si = 0 − where .μ+ s and .μs correspond to the mean reputation scores for the set of positive .Si and negative .Si , respectively. Then there is another standardization parameter

270

S. Bhattacharjee and S. K. Das √

3σ

√

3σ

C + = π s and .C − = π s , where .σs + and .σs − are the standard deviations for the set of positive .Si and negative .Si , respectively, over all users who had reported in the cold start phase. Thus, during the cold start phase .R i produces the reliable user reputation base by using the level of reputation as a weight that can give proportional importance to the future events they report, the incentives they receive, etc. The .R i scores can be seen in Fig. 7 as a function of the quantity and quality of reports for all users in a Waze App. In Fig. 1, the flow diagram describes that such user reputation scores can serve as inputs to any reputation-based incentive calculation model and any reputation-based decision-making scheme that would work more accurately incrementally after the cold start phase has passed. In this way, our method can prevent any attempt by an adversary to poison the accuracy of learning the reputation scores during the cold start phase, when prior user reputation and behaviors are unknown.

.

+

−

5 Analytical Case Study In this section, we introduce an analytical study that did not appear in the experimental results of our research papers, but we believe it will give insights necessary to adapt or improve our model in other contexts. To better explain the five important aspects of QoI and reputation scoring in QnQ, we consider two different scenarios: (i)sparsely crowded location and (ii) densely crowded location. Let the total number of ratings received be 50 and 300 for the sparse and the dense locations, respectively. The event published by the CS administrator based on the reports can be either true or false. Thus, there could be four possible scenarios: (a) false event/sparse location, (b) false event/dense location, (c) true event/sparse location, and (d) true event/dense location. The densities keep changing all the time. Under the four cases, let the adversary have a budget to manage a fixed number of raters (say 20), in every location, who can pose the following four types of threats: (i) false ratings to true events (bad mouthing); (ii) true ratings to false events (ballot stuffing); (iii) deliberate undecided ratings to boost up expected truthfulness of false events (obfuscation stuffing); and (iv) combine false and undecided ratings (mixed attack). Let the parameters of our system take the following values: .Ab = Au = 20, .ν = 0.25, .ϕ = 0.2, .Nthres = 60, and max = 0.5. Only the parameters .B and .B are adjusted in the runtime of QoI .wu b u scoring. For cases (a) and (c), which correspond to a sparse location that has a low number of feedback, the growth rate parameter is adjusted .Bb = Bu = 0.08. For cases (b) and (d), which correspond to a dense location, .Bb = Bu = 0.04. Note that the only difference is in the growth parameter B that mostly depends on the expectation that more ratings will be available over time. For cases (a) and (b), where the event is false, we do not consider the bad mouthing attack as it is not relevant, as such attacks are meant to manipulate

Information Integrity in Participatory Crowd-Sensing via Robust Trust Models

271

Table 3 Case-a: False event at a sparse location Threat No threat Ballot stuffing Obfuscation Mixed

.r, s, t .5, 42, 3 .25, 22, 3 .5, 22, 23 .15, 22, 13

b/d/u 0.11/0.81/0.075 0.49/0.43/0.07 0.11/0.43/0.45 0.3/0.43/0.26

Jøsang 0.128 0.525 0.335 0.43

QnQ 0.036 0.147 0.093 0.12

Jøsang 0.111 0.176 0.205 0.163

QnQ 0.087 0.15 0.09 0.12

Jøsang 0.635 0.46 0.6 0.53

QnQ 0.177 0.128 0.156 0.164

Table 4 Case-B: False event at a dense location Threat No threat Ballot stuffing Obfuscation Mixed

.r, s, t .25, 260, 15 .45, 240, 15 .25, 240, 35 .35, 240, 25

b/d/u 0.085/0.86/0.052 0.15/0.79/0.052 0.085/0.79/0.12 0.12/0.79/0.085

Table 5 Case-c: True event at a sparse location Threat No threat Bad mouthing Obfuscation Mixed

.r, s, t .30, 15, 5 .30, 35, 5 .30, 15, 25 .30, 25, 15

b/d/u 0.58/0.3/0.11 0.42/0.49/0.08 0.42/0.22/0.36 0.42/0.36/0.22

true events into false by deliberate fake (negative) ratings. We represent the rating distributions for each threat barring bad mouthing by a tuple .r, s, t. The expected truthfulness scores generated by Jøsang’s model and QnQ are depicted in Table 3. It is evident from Table 3, for sparse location, the truthfulness score assigned by QnQ to false events is much less compared to that given by the Jøsang’s model. Moreover, our model can readily detect the obfuscation attack, assigning it the lowest value. For case (b), the truthfulness comparison is presented in Table 4. If the location is densely populated, the QoI score assigned by Jøsang’s model is relatively less compared to sparse locations. However, it is still on the higher side compared to the scores generated by QnQ, which computed the coefficients as .wb = 0.99 and .wu = 0.05. Like (A), here also QnQ is particularly able to be severe on obfuscation attack. For cases (c) and (d), as the event is true, a ballot stuffing attack is not practical since that is meant to manipulate a false event into true by deliberate fake (positive) ratings. Tables 5 and 6 give the comparison of truthfulness values computed by the two models in these two cases. For case (c), it is evident that Jøsang’s model assigns high QoI even. However, QnQ refrains from assigning higher truthfulness value even to true events unless it receives a substantial number of ratings. The truthfulness value given by QnQ is lowest under bad mouthing attack, which shows that our model is less robust if rogue raters give deliberate false ratings to true events. The values of the coefficients computed here are .wb = 0.28 and .wu = 0.14(normal), 0.2(other threats).

272

S. Bhattacharjee and S. K. Das

Table 6 Case-D: True event at a dense location Threat No threat Bad mouthing Obfuscation Mixed

.r, s, t .180, 90, 30 .180, 110, 30 .180, 90, 50 .180, 100, 40

b/d/u 0.597/0.3/0.102 0.56/0.34/0.09 0.56/0.28/0.16 0.56/0.31/0.13

Jøsang 0.648 0.605 0.64 0.625

QnQ 0.596 0.558 0.562 0.56

Unlike the other three cases, the truthfulness scores assigned in (d) by both models are on the higher side and are at par with each other. This is because, the event is true and substantial ratings have been received, which led to the generation of high scores. Summarizing the results depicted in Tables 3, 4, 5, and 6, we draw three important observations: (i) QnQ is resilient to ballot stuffing attack by preventing false events to be portrayed as true ones for both low and high number of ratings; (ii) unlike Jøsang’s model, QnQ is completely null invariant and thwarts the threat of obfuscation, and (iii) the proposed model is robust against bad mouthing attacks if a substantial number of ratings are available.

6 Conclusion Our framework is unable to provide robustness against targeted feedback weaponizing attacks, where the majority of the ratings received for given events are compromised. The vulnerability against bad mouthing is more pronounced than ballot stuffing and obfuscation stuffing. However, the final number of ratings received is not entirely under the control of the adversary. Hence, to evade an attack 100% of the time, the adversary needs to get lucky that other raters do not participate in the rating process. We believe that promoting more feedback is important but should not carry incentives, since the selfish intent will become practical in the feedback apparatus as well. Acknowledgments The research is supported by National Science Foundation of USA grant numbers SATC-2030611, SATC-2030624, OAC-2017289.

References 1. J.A. Burke, D. Estrin, M. Hansen, A. Parker, N. Ramanathan, S. Reddy, M.B. Srivastava, Participatory sensing, in Center for Embedded Network Sensing (2006). 2. S. Bhattacharjee, N. Ghosh, V. Shah, S. K. Das, QnQ: a quality and quantity based unified approach for secure and trustworthy mobile crowdsensing. IEEE Trans. Mob. Comput. 19(1), 200–216 (2020)

Information Integrity in Participatory Crowd-Sensing via Robust Trust Models

273

3. F. Restuccia, N. Ghosh, S. Bhattacharjee, S.K. Das, T. Melodia, Quality of information in mobile crowdsensing: survey and research challenges. ACM Trans. Sens. Netw. 13(4), 1–43 (2017) 4. S. Bhattacharjee, N. Ghosh, V.K. Shah, S.K. Das, QnQ: a reputation model for securing mobile crowdsourcing systems from incentive losses, in IEEE Communications and Network Security (2017). (To appear) 5. S. Bhattacharjee, N. Ghosh, V.K. Shah, S.K. Das, W2Q: a dual weighted QoI scoring mechanism in social sensing using community confidence, in IEEE PerCom Workshops (2017), pp. 375–380 6. Irate Homeowners Are Spoofing Waze To Reroute LA Traffic. https://jalopnik.com/iratehomeowners-are-spoofing-waze-to-reroute-la-traff-1660192011 7. Y. Li, Q. Li, J. Gao, L. Su, B. Zhao, W. Fan, J. Han, Conflicts to harmony: a framework for resolving conflicts in heterogeneous data by truth discovery. IEEE Trans. Knowl. Data Eng. 28(8), 1986–1999 (2016) 8. X.O. Wang, W. Cheng, P. Mohapatra, T. Abdelzaher, Enabling reputation and trust in privacypreserving mobile sensing. IEEE Trans. Mob. Comput. 13(12), 2777–2790 (2014) 9. J. Pasternack, D. Roth. Knowing what to believe (when you already know something), in Proceeding of Computational Linguistics (2010), pp. 877–885 10. K.L. Huang, S.S. Kanhere, W. Hu, On the need for a reputation system in mobile phone based sensing. Ad Hoc Netw. 12, 130–149 (2014) 11. https://www.bestreviewapp.com/ 12. https://mopeak.com/buy-ios-reviews/ 13. C. Meng, H. Xiao, L. Su, Y. Cheng, Tackling the redundancy and sparsity in crowd sensing applications, in ACM Conference on Embedded Network Sensor Systems (ACM SenSyS) (2016), pp. 150–160 14. R. Zhang, J. Zhang, Y. Zhang, C. Zhang, Secure crowdsourcing-based cooperative spectrum sensing, in IEEE INFOCOM, (2013) 15. A. Jøsang, An algebra for assessing trust in certification chains, in NDSS (1999) 16. S. Bhattacharjee, M. Chatterjee, Trust based channel preference in cognitive radio networks under collaborative selfish attacks, in IEEE PIMRC (2014), pp. 1502–1507 17. A. Jøsang, R. Ismail, The beta reputation system, in Bled eConference (2002), pp. 41–55 18. A. Josang, J. Haller, Dirichlet reputation systems, in The Second International Conference on Availability, Reliability and Security (ARES’07) (2007), pp. 112–119 19. N. Ram, K. Grimm, in Handbook on Child Psychology, Development Science and Methods: Growth Curve Modeling and Longitudinal Factor Analysis (Wiley, London, 2015), pp. 758– 785 20. B. Yu, M.P. Singh, An evidential model of distributed reputation management, in ACM AAMAS (2002), pp. 294–301 21. J.R. Eiser, M.P. White, A psychological approach to understanding how trust is built and lost in the context of risk, in Conference on Social Contexts and Responses to Risk (2005) 22. R.S. Anderssen, S.A. Husain, R. Loy, The Kohlrausch function: properties and applications. Anziam J. 45, 800–816 (2004) 23. F. Restuccia, S.K. Das, FIDES: a trust-based framework for secure user incentivization in participatory sensing, in IEEE WoWMoM (2014), pp. 1–10 24. J.S. Cramer, in The Origins and Development of the Logit Model (Cambridge University Press, Cambridge, 2003) 25. R.P. Barnwal, N. Ghosh, S.K. Ghosh, S.K. Das, Enhancing reliability of vehicular participatory sensing network: a Bayesian approach, in IEEE SMARTCOMP (2016), pp. 1–8 26. D. Kahneman, A. Tversky, Prospect theory: an analysis of decision under risk, in Econometrica (World Scientific, Singapore, 1979), pp. 263–291

AI-Driven Attack Modeling and Defense Strategies in Mobile Crowdsensing: A Special Case Study on Fake Tasks Didem Cicek, Murat Simsek, and Burak Kantarci

1 Introduction Mobile crowdsensing (MCS) is an unprecedented mass data sensing and collection activity that does not require expensive sensing infrastructure and leverages volunteer crowds to contribute data through smart devices equipped with various types of sensors. MCS emerges as an important contributor to the creation of the Internet of Things (IoT) and enabling the concept of smart city where cyber and physical objects interact with each other. As local authorities strive to employ the information and communication technologies to improve their public services or to tackle urbanization problems, they look for cost-effective and ubiquitous data collection methods. MCS stands out as an effective technique in this regard that can collect real-time data of wide coverage. A typical MCS system consists of a service provider, a sensing platform, and participant nodes who send sensed data to the platform. A high number of participants are required for a MCS application to be effective, and it is a big challenge to recruit the required number of participants. The service provider wants benign participants in the system who contribute with legitimate tasks, and at the same time, the participants need to feel secure and that their privacy is respected. A successful MCS system can only operate in a secure, privacy-preserving, and trustworthy environment. However, as MCS systems are open (i.e., non-dedicated [9]) and mobile, and communicate through wireless channels, they are prone to security threats. Fake task submissions appear as a severe and understudied security threat in MCS systems. They can be initiated by external attackers or by internal sensing

D. Cicek · M. Simsek () · B. Kantarci The School of Electrical Engineering and Computer Science at the University of Ottawa, Ottawa, ON, Canada e-mail: [email protected]; [email protected]; [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 J. Wu, E. Wang (eds.), Mobile Crowdsourcing, Wireless Networks, https://doi.org/10.1007/978-3-031-32397-3_11

275

276

D. Cicek et al.

participants with different motivations. The motivation can vary from draining the energy of the server or other participants’ devices and clogging the system to misleading the MCS decision support mechanisms. Since this type of threat is not well investigated in the literature, there is not much information on how impactful such attacks can be. Zhang et al. demonstrate a worst-case impact scenario for a fake task submission in their work [38]. Another important threat to MCS systems is fake sensing by untrustworthy users, which is almost an unavoidable characteristic of a MCS system that leads to false payments [21]. There are many studies in the literature that focus on trustworthiness assurance, and some of them are discussed in this chapter too. Realistic adversarial attack design has a key role to prepare for potential malicious activities in real-life cybersecurity systems. Attackers would put their best effort to make the attack more impactful to maximize their outcomes. Intelligent attacks leveraging AI-based strategies can manage to stay undetected by conventional defense strategies. Hence, AI-driven defense mechanisms are also required to protect the system and its resources from such challenging attacks. Legitimacy detection is the leading defense strategy to detect and eliminate fake tasks in a MCS system that are elaborated deeply in Sect. 5 of this chapter. Due to the vast amount of data and wide geographical coverage in a MCS application, it is challenging to detect the legitimacy of the tasks or the users. Machine learning (ML) techniques are highly applied in defense strategies. To prevent or mitigate such fake task threats, there is also attack modeling strategy that models fake task injections. AI-based intelligent adversarial fake task submission design that leverages ML techniques is discussed in detail in Sect. 4 of this chapter. General concept of attack and defense modeling is illustrated in Fig. 1. Machine learning, as a powerful AI technique in computer science, has many successful applications, e.g., language processing, intrusion detection, image processing, and health care. It is also quite effective in mitigating security threats. More and more MCS applications are employing AI-enabled strategies to protect their resources against security threats due to their ability to identify patterns in complex datasets. Supervised ML methods such as DeepNN, Decision Tree, and SVM or ensemble methods such as XG Boost, ADA Boost, or Random Forest are employed in security design of MCS systems. Additionally, knowledge-based ML methods such as Self-Organizing Feature Map (SOFM), PKI, or PKI-D can also be used in the security design, and some of these techniques are covered in detail throughout the chapter. The chapter sections are organized as follows: Sect. 2 gives a background on MCS technology concerning the system architecture, use cases, and the quality of service aspects. Section 3 makes an introduction into security in MCS and lists down different threat models. Section 4 explains AI-driven attack strategies, while Sect. 5 explains AI-driven defense strategies, both elaborating on the effective machine learning models for better prediction performance. Section 6 is the conclusion with some key highlights.

AI-Driven Attack Modeling and Defense Strategies Against Fake Task. . .

277

Fig. 1 General concept for attack and defense modeling in MCS

2 Background on Mobile Crowdsensing Current smart devices (i.e., smartphones, tablets, wearables) contain several builtin sensors such as GPS, camera, microphone, accelerometer, gyroscope, etc. The advancement of sensor technologies embedded in smartphones enables the commercialization of mobile sensed data where a large number of smartphone users monitor their environment with the sensors in their devices and share them with service providers in return for a reward [15]. A key defining characteristic of mobile crowdsensing systems is whether the sensing activity is participatory or opportunistic. In a participatory model, the crowd actively participates in the data sensing activity in return for a reward, whereas in opportunistic crowdsensing, sensed data readings are done automatically in the background from the environment or computations through smart devices. MCS, due to its participatory nature where mobile residents generate valuable urban data, emerges as a key enabler of smart cities. The increasing amount of data that is

278

D. Cicek et al.

generated in our environment through ubiquitous sensors and made available for the use public authorities enhances their decision-making process through fast and informed decisions supported by real-time data. A city can only be smart if it utilizes intelligence functions that integrate and synthesize data to improve the efficiency, sustainability, equity, and quality of life in the city [1]. Mobile crowdsensing can be used to solve a variety of problems in different domains by both public and private parties. Information technologies including crowdsensing technology are increasingly used to enhance public services, improve daily lives of people, and tackle major urbanization problems such as natural disasters, traffic congestion, pollution, healthcare, or health surveillance as recently made required by the Covid19 pandemic [11].

2.1 Use Cases of MCS There is an increase in the number of studies on crowdsensing in disaster management since 2010 in parallel with several disasters happening such as Haiti earthquake 2010, US Superstorm Sandy 2012 Colorado Wildfire 2013, or Hawaii Hurricanes 2014, which have contributed growing the research interest in finding solutions to reduce human and economic loss from such disasters. It is very critical for rescue authorities to receive and process real-time data quickly in responding to a disaster to prevent potential human and economic losses. Emerging mobile crowdsensing technology can provide real-time and reliable information to support decision-making of disaster management authorities. Villela et al. [31] propose a smart and inter-operable decision support system, named RESCUER, which makes use of crowdsourced information via inclusion of dedicated rescue support staff in emergency and crisis management. Kitazato et al. [13] propose a system using mobile crowdsensing to detect real-time pedestrian flows that is crucial for disaster evacuation guidance of civilians. Pan et al. [19] use mobile crowdsensing to detect traffic anomalies according to drivers’ routing behavior. The authors use human mobility data and social media data to identify accidents, events, or disasters, etc., that create a traffic anomaly. Marjanovic et al. [17] develop a mobile-crowdsensingbased solution through low-cost wearable mobile sensor to monitor air quality and noise pollution. In this chapter, MCS technology extends the traditional procedure of measurement by fixed meteorological stations and provides good density of measurements in several locations. Capponi et al. [2] illustrate some representative application examples in health care such as HealthAware, MPCS, and DietSense that support healthy eating by food images from consumers and user activities with extraction of time and location information. Simsek et al. [27] utilize MCS-sensed data through residents’ mobile devices to develop a strategy of early detection of infected COVID-19 cases through deployment of autonomous-vehicle-based mobile assessment and testing centers.

AI-Driven Attack Modeling and Defense Strategies Against Fake Task. . .

279

2.2 System Architecture of MCS MCS system is designed in a platform that creates and assigns tasks through campaigns. There is a crowdsourcer/task initiator who assigns data-sensing tasks among platform users, a sensing platform that manages user registration, data aggregation and reporting of the results and participants (or workers) who work collaboratively with the crowdsourcer to obtain sensory data through their devices over a wide geographic area. The phase of recruiting participants and rewarding them for their contribution to the campaigns costs money to the platform. At the same time, participants also incur some cost on their end like energy consumed for data sensing. Therefore, a metric named “utility” is defined separately for the platform (platform utility) and the users (user utility) to measure the cost–reward balance for both parties [21]. MCS uses a sensing as a service (S2aaS) business model where infrastructure is made readily available to the companies, and thus companies no longer need to invest in the infrastructure to launch a sensing campaign. A wide range of technological tools and processes are employed in the system, and thus categorizing the system in 4 layers simplifies the understanding of the system components. The application layer involves the key functionalities of MCS such as user recruitment, allocation of tasks and reporting; data layer contains tools and processes for data storage, analysis, and processing; communication layer includes all sorts of communication technologies that deliver the sensed data; and lastly, the sensing layer includes all the sensors to collect the data [2].

2.3 Quality of Service in MCS The adoption of a MCS approach by the companies requires a certain level of guarantee for quality of service (QoS), while QoS comes with its challenges. One key challenge is the resource limitation. The sensors on smart devices continuously create huge amounts of raw sensor data that consume many resources such as bandwidth, energy, and storage. However, smart sensing devices have limited resources to handle immense amount of data, and this might sacrifice the quality of service for MCS applications. This resource limitation raises the question of how much of the sensed data are actually used and how much of it is redundant. Liu et al. [16] discuss that data load and hence resource cost can be significantly reduced by cutting on the redundant data and thereby enhancing quality of service. The problem of resource limitation leads to the problem of quality of data (QoD). For instance, the quality images collected in a disaster incident play an important role in planning disaster relief actions, but the useful images may not be uploaded in time due to bandwidth constraints.

280

D. Cicek et al.

MCS systems require contribution from a large number of participants to operate efficiently. However, it is not an easy task to recruit participants for MCS applications mainly due to lack of motivation or having privacy concerns [2]. To tackle this problem, a lot of research effort is put in the development of incentive mechanisms [10, 18, 35] and in investigating privacy concerns [7] or trying to tackle both incentive and privacy concerns [12]. Another non-trivial challenge in MCS applications is the data trustworthiness, meaning how accurate is the data shared by the participants and how well intended the participants are. This multi-faceted problem has been studied by Pouryazdan et al. in [21], and collaborative reputation scores have been proposed to mitigate this issue and prevent false incentive payments.

3 Security and Threat Models in MCS The multi-dimensional information collected and analyzed through mobile crowdsensing applications has led to the emergence of a new understanding of life and society. This allowed researchers and industry practitioners to go beyond traditional data collection and analysis methods with reduced effort on these processes. There is a big potential in the crowd data analytics market, but it comes with its challenges. The key security challenge of MCS researchers and practitioners is to ensure a secure, privacy-preserving, and trustworthy MCS system. Security is about protection of the sensed data and prevention of unauthorized access into MCS systems, privacy is about protection of the data owner or the participant in the sensing task, and trust is confidence in the integrity and reliability of the sensed data or the participant [8]. MCS operates on an open system where any mobile device can participate and contribute [8, 32]. Hence, malicious users can easily join and threaten the system. Malicious participants pose a big vulnerability on MCS systems that can compromise the server or device availability. Such users may try to report false sensor readings or inject fake tasks into the system based on various motivations [36]. Privacy is critical in MCS systems to protect the user from the system. Wang et al. [32] discuss and classify privacy as task privacy (for task information), identity privacy (for participants), attribute privacy (for data providers’ location, hardware details), and data privacy (for data content).

3.1 Threat Models in MCS Mobile crowdsensing (MCS) systems are prone to various threats from both internal and external adversaries. Selfish users trying to maximize their payments are considered as internal adversaries. External adversaries are malicious users who

AI-Driven Attack Modeling and Defense Strategies Against Fake Task. . .

281

are motivated to clog the system, steal sensitive information, or even crash the system. Such attacks adversely affect particularly the sensing and computational capacities of the system. MCS systems have to address these threats to achieve a secure, privacy-preserving, and trustworthy system. The types of security threats that MCS systems are exposed to are summarized here: • Eavesdropping/Privacy leakage: Data transmission in MCS systems is done mostly through wireless networks that are prone to eavesdropping attacks [32]. Hence, encryption of data is crucial to prevent privacy leakage. • Malicious task attacks: Malicious tasks are initiated with the intention to steal sensitive or identity information from the participant nodes. Narrow tasking, where some strict attribute limitations are imposed on the task, or selective tasking, where the task is shared with a limited group of participants, are both forms of malicious attacks [20]. • Fake task submission: Task initiators or participant nodes can create fake tasks aiming at draining the resources of both participating devices and MCS servers. Fake tasks usually have longer duration and are aggregated in peak hours with the motivation to keep other participant devices busy and drain their battery, storage, or computing capacity [40]. This attack type resembles the denial of service attack in networked systems. • Faked sensing attacks: Selfish participant nodes can submit false sensing reports to protect their privacy or just to deceive the incentive system and maximize their payments [33]. • Data poisoning attacks: Data poisoning is sending fake data into the MCS system with the motivation to clog the resources (devices, servers, etc.) and mislead the decision-making [22]. • Sensitive information inference: There are some semi-honest entities in MCS systems that are not directly violating privacy protocols but indirectly trying to steal sensitive information such as task details, personal attributes, or location through data mining [20]. • Collusion attack: In this type of attack, service provider or participant nodes can collude with each other to exchange information on other nodes to reveal their identity or just give false negative feedback on others to increase their own payment while decreasing others’ [32]. There are also other security threats to MCS that are not specifically designed for MCS systems but mimicked from other networked systems, such as spoofing or sybil attacks that violate identity protocols and manipulate fake identities to obtain awards [33], or attacks targeting network resources to make them unavailable such as denial of service (DoS) [36], Jamming, Clogging, Malwares, or Advanced persistent threat (APT) attacks [6].

282

D. Cicek et al.

4 AI-Driven Attack Anticipation in MCS We identify the fake task injection threat as a unique attack type in MCS platforms and present AI-driven fake task anticipation as a case study. General algorithmic workflow can be seen in Fig. 2.

Fig. 2 Mobile crowdsensing process throughout the whole campaign for fake task submission attack

AI-Driven Attack Modeling and Defense Strategies Against Fake Task. . .

283

4.1 Fake Task Injection Modeling Previous work in [37], Zhang et al. designed a model for fake task injection that targets the batteries of other users. A general overview of the adversarial model is presented in Fig. 1 where the malicious end users inject fake tasks into the MCS system. Such attacks aim at clogging the server and draining the batteries of users. We believe that the attackers would want to impact as much user population as possible. With that in mind, we built the model with some assumptions on key characteristics of fake tasks: • Fake tasks target specific clogging regions in a city. • Fake tasks are launched during the busy hours of the day. • Fake tasks tend to have longer duration times than the legitimate ones. Attack regions are defined as circular regions with radius R. To generate attacks with high impact, SOFM is utilized so that the areas with the highest mobile population are determined to ensure more participants are affected by these intelligent attack strategies.

4.2 Types of Task Movement In order to generate a realistic yet challenging task, mobility aspect has to be considered. Both legitimate and fake tasks have similar movement patterns. Each task consists of multiple sub-tasks. Task movement can be defined as movement from former sub-task to later sub-task. CrowdSenSim is a useful simulation tool to simulate the whole mobility pattern, and it is used to run all the simulations discussed in this book chapter. There are two types of movement model for both fake and legitimate task generation: Zone-free movement (ZFM) and Zone-limited movement (ZLM). In order to capture a wider sensing area, tasks are given mobility every ten minutes. In both movement models, a task movement is determined by two parameters: movement radius (.rmov ) and movement bearing (.θ ).

4.2.1

Zone-Free Task Movement (ZFM)

Zone-free movement is basically an unrestricted movement for both types of tasks as shown in Fig. 3. An attack zone, where the attacker submits the fake tasks into the MCS platform, is determined through selecting the center and the radius of the zone. Those attack zones are mostly pre-selected areas in the city. Legitimate tasks are initiated without any zone restriction; hence, they have the potential to be located in the attack zones too. In a ZFM model, fake tasks can move out of the selected

284

D. Cicek et al.

Fig. 3 Zone-free movement algorithm for task location (This figure is reproduced from the figure in [37])

Fig. 4 Mobility pattern of legitimate and fake tasks under zone-free movement

attack zones as well as legitimate tasks can move into the attack zones without any restriction as illustrated in Fig. 4.

4.2.2

Zone-Limited Task Movement (ZLM)

ZLM model is designed to create a protected zone from illegitimate tasks through introducing boundary controls to such tasks in MCS campaigns. The zone-limited task movement algorithm, as shown in Fig. 5, indicates the steps to generate a task in a zone-limited model. The legitimate tasks are set to stay out of the attack zones,

AI-Driven Attack Modeling and Defense Strategies Against Fake Task. . .

285

Fig. 5 Zone-limited movement algorithm for task location (This figure is reproduced from the figure in [37])

and hence, they are not initiated in the attack zones and their launch positions are deliberately chosen out of the attack zone. Boundary control helps illegitimate and legitimate tasks adopt two separate movement rules. The zone-limited mobility model ensures that illegitimate tasks stay in the attack zones, while legitimate tasks stay out of the attack zones. Figures 6 and 7 illustrate the steps to determine the next destination for both kinds of tasks. For the zone-limited movement model, a destination position is initially proposed as estimated position (.PEstimated ), and then it is modified to determine the corrected and the ultimate position (.PCorrected ) of the movement decision.

4.3 Self-organizing Feature Map Implementation for Attack Modeling In the attack design, from the attacker’s standpoint, the impact of the fake task injection attack is aimed to be maximized so that more participants get affected. To this end, realistic design of fake task injection attack aims to leverage positioning information to boost the damage by reaching out to a larger population of users. Fake tasks that target the best position for broad coverage affect more users than the fake task submissions whose positions are randomly selected. When fake sensing tasks recruit new devices, these recruits will be penalized with wasted battery levels;

286

D. Cicek et al.

Fig. 6 Mobility pattern of legitimate tasks under zone-limited movement

Fig. 7 Mobility pattern of fake tasks under zone-limited movement

hence, fake task submissions will affect the completion of legitimate tasks via consuming their batteries and other resources. Multi-dimensional data can be represented by two-dimensional neurons through an unsupervised learning method such as SOFM [14, 30, 40]. These neurons have special topology that is able to connect one neuron to another. SOFM learning algorithm is quite similar to winner-take-all networks, which is based on updating only the winner neuron; however, SOFM is supposed to update neighborhood neurons that are directly connected to winner neuron or the so-called the Best Matching Unit (BMU). Hence, BMU and its neighbors are always more competitive than other neurons. This process is repeated for all inputs in each epoch until SOFM stops when the maximum number of epochs is reached. In MCS platforms, where data aggregation from multiple sensory sources is essential, SOFM appears as a

AI-Driven Attack Modeling and Defense Strategies Against Fake Task. . .

287

Fig. 8 Self-organizing feature map structure and attack region selection

quite effective method to identify potential locations that fake tasks can target to achieve damage on wider participants population. Neuron weights of the network are used by SOFM as a part of the learning process to help compare against the input parameters. The weights of each neuron consist of coordinates latitude and longitude that are transformed into meters for improving sensitivity of them, and BMU has the most similar weights compared to the input vector, as shown in Fig. 8. Distance function for 2-dimensional inputs and weight vectors is shown in Eq. (1).  Dist (X, W ) =

.

 xlat − w1 × [xlat − w1 , xlon − w2 ], xlon − w2

(1)

where X is the current 2D input vector and W is the 2D node’s weight vector as shown in Fig. 8. Dist helps to understand how close w is to x leading to the information about the output layer neuron that is the best to represent the input features. Best matching unit (BMU) indicates the closest neuron’s weight to the applied input vector as given in Eq. (2).     BMUX = argmin Dist (X, W j ) for j = 1, .., m,

.

(2)

w

where BMU indicates best matching neuron y to the input x as shown in Fig. 8. W i = BMUXi ,

.

(3)

where W indicates BMU for ith input X. Detailed SOFM formulations and updating rules can be seen in [40]. General updating structure and 2D coordinates selection of user movement is depicted in Fig. 8. SOFM can achieve the maximum coverage of users after the training is completed. For the real impact of SOFM on illegitimate tasks, and determining the number of covered participants by each neuron, clogging region

288

D. Cicek et al.

radius is put as a constraint to each neuron coverage. Finally, it can be concluded that the illegitimate task position with the maximum user coverage has the highest impact on the participant smartphones in reducing their battery levels.

4.4 Region-Based SOFM Structure In a traditional learning process of SOFM, neuron positions are placed randomly. However, region-based approach offers a more effective way to initialize random positions. Maximum values of latitudes and longitudes are indicated in Fig. 9 as .Lat_max and .Lon_max, respectively. Region-based calculations also take into account the minimum values of latitudes and longitudes. As all max and min limit values are determined for latitudes and longitudes, .P Nd value is needed to calculate the total number of neurons in both latitude and longitude directions. Equation (4) is used to calculate the number of neurons in latitude direction, and a similar calculation, as in Eq. (5), is required for the number of neurons in longitude direction. In this approach, the number of neurons for latitude and longitude directions varies depending on the borders of user movement in a MCS campaign. (Lat_max − Lat_min) −1 P Nd

(4)

(Lon_max − Lon_min)  − 1. P Nd

(5)

Lat_neuron_num = 

.

Lon_neuron_num = 

.

The full process of determining the number of neurons according to the city coordinates is given in Fig. 9.

4.5

Locally Reconfigurable SOFM for More Impactful Attack Region Selection

SOFM can also be used to identify more impactful and harmful attack regions. The recent publication [3] shows that if a local search algorithm is embedded around the attack region, it is possible to find more impactful center points for SOFM neurons. In this approach, .R  indicates a bigger radius than R, which is the radius of attack region. SOFM neurons in the center of attack regions are trained locally considering individuals in .R  radius. According to individuals movement patterns, new attack centers can be obtained to ensure more individual coverage than previous attack centers. The entire process has been explained in Fig. 10. Detailed information and more explanation can be seen in [3].

AI-Driven Attack Modeling and Defense Strategies Against Fake Task. . .

289

Fig. 9 Region-based adaptive number of neurons according to pre-defined neuron distance .(P Nd )

5 AI-Driven Defense Strategies in MCS The fake task injection threat has been considered as a unique attack type in MCS platforms, and AI-driven fake task detection has been presented as a case study. Defense strategies aim to reveal the ground truth, detect legitimate, and fake tasks in the system. They enhance the trustworthiness of MCS systems and directly contribute to the motivation of legitimate participants to join the system. They proactively prevent void drainage of MCS resources and the economic loss caused by battery drainage. There have been prior studies focusing on game-theoretic incentive mechanisms to prevent malicious users from misleading the MCS system with faked sensing attacks [34]. There is less emphasis on defense mechanisms such as fake task detection or ground truth discovery. This section will discuss the state-of-the-art AI-driven defense strategies with embedded ML techniques.

290

D. Cicek et al.

Fig. 10 Modified SOFM process flow

5.1 AI-Backed Legitimacy Detection Machine learning methodology is commonly used for embedding intelligence into attack or defense mechanisms in engineering problems. Malicious activity, which targets resources of MCS system to directly affect the performance of smartphone users in legitimate tasks, is designed as fake tasks to hide itself from any detection mechanism. The sensing capacity of smartphones is utilized for collecting sensor output of smart devices for MCS campaign via task generation. Sensing resources can be affected by fake task injection attack to clog sensing resources such as computational power and battery level of smart devices. Fake task injection attack in MCS directly influences the performance of data collection process, so legitimacy detection system has an important role to eliminate

AI-Driven Attack Modeling and Defense Strategies Against Fake Task. . .

291

these harmful effects of this attack. In legitimacy detection, the prediction performance can be evaluated by true positive (TP) and true negative (TN) predictions that are true legitimate and fake tasks, respectively. False positive (FP) predictions determine the actual fake but predicted legitimate tasks, while false negative (FN) predictions determine actual legitimate but predicted fake tasks. False predictions that help keep fake tasks in the system and eliminate legitimate tasks eventually hinder the performance of the MCS campaign. For the best performance of a MCS campaign, keeping the legitimate tasks and eliminating fake tasks are the most desired cases.

5.2 Machine Learning Model Development to Increase the Performance of Legitimacy Detection Machine learning model development is key to improve the performance of fake task detection. In a MCS data collection process, the collected data are divided into two parts: training data and test data. Efficiency of a ML model in fake task detection is affected by the size of the training and test datasets. The trade-off between the size of test and training data has a direct impact on the legitimacy detection performance [26]. Initially, data acquired from MCS platform are used for training the legitimacy detection system. Upon completion of the training, the remaining data from the MCS platform can be utilized by the legitimacy detection system to eliminate any fake tasks in the MCS system. If the training process is performed with more data, then the model is able to learn more but has limited detection capability. If the size of the training data is limited, then test data will be bigger in size. Under an highly efficient AI-driven ML model, training dataset can be limited, but still its performance can be better than a bigger training dataset in eliminating more fake tasks from the system. There are three useful techniques recently applied to achieve an highly efficient AI-driven ML model. First, a supervised machine learning technique can be trained by a training dataset, and test performance can be evaluated using a confusion matrix [5, 25, 29, 39]. Feature selection can be embedded to the model to eliminate less important features and achieve a less complex problem. Hence, the same dataset will be sufficient to provide a better performance with the lower number of features. The general structure of the learning process for this technique can be seen in Fig. 11. As seen in the figure, wrapper and filter approaches are chosen as feature selection techniques. It should be considered that wrapper techniques need more computational power compared with filter-based feature selections. Second, a knowledge-based technique such as prior knowledge input (PKI) can be applied as shown in Fig. 12. In this technique, SOFM is used to generate extra knowledge that will boost the performance. Results in the recent studies [23, 24] prove that extra knowledge helps the supervised model learn easily through this

292

D. Cicek et al.

Fig. 11 Legitimacy detection: classical ML model

Fig. 12 Legitimacy detection: knowledge-based prior knowledge input (PKI) model

Fig. 13 Legitimacy detection: hybrid model, the combination of unsupervised and supervised models

AI-Driven Attack Modeling and Defense Strategies Against Fake Task. . .

293

extra knowledge about the original problem. Clustering of the training data that is obtained by SOFM provides regional knowledge to the supervised ML model so it can establish more accurate relationship between input and output through this extra cluster knowledge. Finally, the combination of supervised and unsupervised techniques can be used in hybrid approach [4, 28]. Clustering can be used to detect all legitimate tasks as well as the mixed data that consist of legitimate and fake tasks as seen in Fig. 13. After clusters are obtained by unsupervised models, supervised models are embedded to detect legitimate and fake tasks in a more balanced dataset. In the second phase, prediction performance is supposed to increase with all legitimate tasks that are already detected in the first phase. This technique has been developed to overcome imbalanced classification problems. In order to assess the prediction performance of the ML models, performance evaluation metrics are required. Prediction performance of a ML model is considered successful if more fake tasks are eliminated and less legitimate tasks are lost. To make a fair evaluation, training data should be kept in the system as it is, which means legitimate tasks are assumed as true positive (TP) and fake tasks are assumed as false positive (FP). TP means legitimate tasks are detected correctly and kept in the system. FP means fake tasks are detected as legitimate that is why the system keeps FP predicted data in the system. This is the only way not only to keep training dataset as predicted legitimate in the system but also make training data a part of evaluation metric to compare the performance of ML model and the size of the dataset. Thus, the size of the training dataset does not affect the size of the performance evaluation data. More training data increase the chance of more TP and FP predictions, while less training data can increase the chance of true negative (TN) and false negative (FN) predictions. TN prediction means fake tasks are detected as fake tasks and eliminated by legitimacy detection system. FN means legitimate tasks are falsely detected as fake tasks and are supposed to be eliminated in the system. As clearly will be figured out that if training data are limited, a more accurate ML model is required to keep more legitimate tasks and less fake tasks in the system. To make it possible, high TP and TN ratio with low FP and FN ratio should be achieved. For assessing the efficiency of the legitimacy detection mechanism, profit and cost calculation is quite helpful. Profit is calculated by subtracting the number of false predictions from the number of true predictions as shown in Eq. (6). Cost is calculated by the ratio of false predictions over the total number of tasks as shown in Eq. (7). TP indicates legitimate tasks, while TN indicates fake tasks. True predictions are required for the legitimacy detection. FNs are the legitimate tasks that are falsely eliminated, and FPs are the fake tasks falsely kept in the system. A desired legitimacy detection mechanism must achieve high-profit and low-cost performance through keeping the maximum number of legitimate tasks in the system and eliminating the maximum number of fake tasks from the system. P rof it = 100 ∗

.

(T P + T N) − (F P + F N) (T P + T N + F P + F N)

(6)

294

D. Cicek et al.

Cost = 100 ∗

.

(F P + F N) . (T P + T N + F P + F N)

(7)

Another criteria to assess the prediction performance of ML models in a MCS system are fake tasks elimination ratio (FakeT_elimination_ratio) and legitimate tasks loss ratio (LegT_loss_ratio). Legitimate task loss ratio is the number of legitimate tasks not detected as legitimate by the model divided by the total number of legitimate tasks as seen in Eq. (8). It is a viable performance indicator if this ratio is lower since it shows that the maximum number of legitimate tasks is kept in the system. LegT _loss_ratio = 100 ∗

.

(T otal_LegT − T P ) . (T otal_LegT )

(8)

Fake tasks elimination ratio is the number of correctly predicted fake tasks (TN) divided by the total number of fake tasks in the system as shown in Eq. (9). It is a good performance indicator if this ratio is higher since it shows that the maximum number of fake tasks is truly removed from the system. F akeT _elimination_ratio = 100 ∗

.

TN . (T otal_F akeT )

(9)

6 Conclusion MCS systems have vulnerabilities such as a lack of initial trustworthiness of the participants that may lead to adversarial attacks such as fake task submissions. Malicious users may be motivated to submit fake tasks into the MCS system to drain other users’ sensing resources and achieve more incentives for themselves or simply to damage the performance of the system. This chapter discussed effective AIdriven security strategies from both attack and defense angles to ensure a secure and trustworthy MCS system. We identified fake task injection as a unique attack type in MCS systems and discussed both the attack modeling and detection mechanisms around the threat of fake tasks. As another important contribution, this chapter explained how to implement machine learning models in attack and defense strategies. Supervised models are mostly used to mitigate fake task injection attacks. As an additional improvement strategy, feature selection techniques (i.e., wrapper and filter-based approach) can be embedded prior to the supervised training process. Moreover, unsupervised models can provide extra knowledge to the supervised learning process through clusteringbased similar data selection, which is an application of knowledge-based modeling such as PKI. Unsupervised models are also able to identify some legitimate-only tasks as well as the mix of legitimate and fake tasks. Hence, the imbalanced dataset can be transformed to a less imbalanced data for supervised learning process. This

AI-Driven Attack Modeling and Defense Strategies Against Fake Task. . .

295

two-step process, the collaborative (or hybrid) usage of supervised and unsupervised models, improves the performance of legitimacy detection system in order to effectively mitigate malicious activities. Performance evaluation metrics of ML models are also elaborated in this chapter. Cost and profit functions and fake task elimination and legitimate task loss ratios are defined as important evaluation criteria to determine the optimum size of the training data to maximize the machine learning model accuracy. Thus, trained machine learning model can perform better on the remaining MCS data in order to eliminate fake tasks in the MCS platform. In this chapter, we have identified fake task submission as a key threat and focused on fake task injection and detection strategies as case studies. However, as a future research direction, application of these intelligent machine learning strategies that are explained in this chapter can be applied to the attack types other than fake task injection in MCS systems.

References 1. M. Batty, K.W. Axhausen, F. Giannotti, A. Pozdnoukhov, A. Bazzani, M. Wachowicz, G. Ouzounis, Y. Portugali, Smart cities of the future. Eur. Phys. J. Spec. Topics 214(1), 481–518 (2012) 2. A. Capponi, C. Fiandrino, B. Kantarci, L. Foschini, D. Kliazovich, P. Bouvry, A survey on mobile crowdsensing systems: challenges, solutions, and opportunities. IEEE Commun. Surv. Tutorials 21(3), 2419–2465 (2019) 3. X. Chen, M. Simsek, B. Kantarci, Locally reconfigurable self organizing feature map for high impact malicious tasks submission in mobile crowdsensing. Int. Things 12, 100297 (2020) 4. Z. Chen, M. Simsek, B. Kantarci, Region-aware bagging and deep learning-based fake task detection in mobile crowdsensing platforms, in 2020 IEEE Global Communications Conference (GLOBECOM), IEEE; IEEE Commun Soc; Chunghwa Telecom; Foxconn; Huawei; Mediatek; 6G Off; Askey; Ericsson; FET Qualcomm; Taiwan Mobile; Auden Techno Corp; Keysight Technologies; Telecom Technol Ctr; NI (2020) 5. Z. Chen, Y. Zhang, M. Simsek, B. Kantarci, Deep belief network-based fake task mitigation for mobile crowdsensing under data scarcity, in ICC 2020–2020 IEEE International Conference on Communications (ICC) (IEEE, Dublin, Ireland, 2020), pp. 1–7 6. Z. Chen, J. Liu, Y. Shen, M. Simsek, B. Kantarci, H.T. Mouftah, P. Djukic, Machine learningenabled IoT security: open issues and challenges under advanced persistent threats. ACM Comput. Surv. 55, 3530812 (2022) 7. D. Christin, A. Reinhardt, S.S. Kanhere, M. Hollick, A survey on privacy in mobile participatory sensing applications. J. Syst. Softw. 84(11), 1928–1946 (2011) 8. W. Feng, Z. Yan, H. Zhang, K. Zeng, Y. Xiao, Y.T. Hou, A survey on security, privacy, and trust in mobile crowdsourcing. IEEE Int. Things J. 5(4), 2971–2992 (2018) 9. H. Habibzadeh, Z. Qin, T. Soyata, B. Kantarci, Large-scale distributed dedicated- and nondedicated smart city sensing systems. IEEE Sensors J. 17(23), 7649–7658 (2017) 10. L.G. Jaimes, I.J. Vergara-Laurens, A. Raij, A survey of incentive techniques for mobile crowd sensing. IEEE Int. Things J. 2(5), 370–380 (2015) 11. S. Kielienyu, B. Kantarci, D. Turgut, S. Khan, Bridging predictive analytics and mobile crowdsensing for future risk maps of communities against COVID-19, in Proceedings of the 18th ACM Symposium on Mobility Management and Wireless Access, MobiWac ’20 (Association for Computing Machinery, New York, 2020), pp. 37–45

296

D. Cicek et al.

12. F. Khan, A. Ur Rehman, J. Zheng, M.A. Jan, M. Alam, Mobile crowdsensing: a survey on privacy-preservation, task management, assignment models, and incentives mechanisms. Futur. Gener. Comput. Syst. 100, 456–472 (2019) 13. T. Kitazato, M. Hoshino, M. Ito, K. Sezaki, Institute of industrial science, the university of Tokyo 4-6-1 Komaba, Meguro-ku, Tokyo 153–8505, Japan, center for spatial information science, the university of Tokyo, Chiba, Japan: detection of pedestrian flow using mobile devices for evacuation guiding in disaster. J. Disaster Res. 13(2), 303–312 (2018) 14. T. Kohonen, Self-organized formation of topologically correct feature maps. Biol. Cybern. 43(1), 59–69 (1982) 15. J.S. Lee, B. Hoh, Sell your experiences: a market mechanism based incentive for participatory sensing, in 2010 IEEE International Conference on Pervasive Computing and Communications (PerCom) (IEEE, Mannheim, Germany, 2010), pp. 60–68 16. J. Liu, H. Shen, X. Zhang, A survey of mobile crowdsensing techniques: A critical component for the internet of things, in 2016 25th International Conference on Computer Communication and Networks (ICCCN) (IEEE, Waikoloa, HI, USA, 2016), pp. 1–6 17. M. Marjanovic, S. Grubesa, I.P. Zarko, Air and noise pollution monitoring in the city of Zagreb by using mobile crowdsensing, in 2017 25th International Conference on Software, Telecommunications and Computer Networks (SoftCOM) (IEEE, Piscataway, Split, 2017), pp.1–5 18. K. Ota, M. Dong, J. Gui, A. Liu, QUOIN: incentive mechanisms for crowd sensing networks. IEEE Netw. 32(2), 114–119 (2018) 19. B. Pan, Y. Zheng, D. Wilkie, C. Shahabi, Crowd sensing of traffic anomalies based on human mobility and social media, in Proceedings of the 21st ACM SIGSPATIAL International Conference on Advances in Geographic Information Systems (ACM, Orlando Florida, 2013), pp. 344–353 20. L. Pournajaf, D.A. Garcia-Ulloa, L. Xiong, V. Sunderam, Participant privacy in mobile crowd sensing task management: a survey of methods and challenges. SIGMOD Record 44(4), 12 (2015) 21. M. Pouryazdan, B. Kantarci, T. Soyata, L. Foschini, H. Song, Quantifying user reputation scores, data trustworthiness, and user incentives in mobile crowd-sensing. IEEE Access 5, 1382–1397 (2017) 22. A. Prud’Homme, B. Kantarci, Poisoning attack anticipation in mobile crowdsensing: A competitive learning-based study, in Proceedings of the 3rd ACM Workshop on Wireless Security and Machine Learning (ACM, Abu Dhabi United Arab Emirates, 2021), pp. 73–78 23. M. Simsek, B. Kantarci, A. Boukerche, Knowledge-based machine learning boosting for adversarial task detection in mobile crowdsensing, in 2020 IEEE Symposium on Computers and Communications (ISCC) (IEEE, Rennes, France, 2020), pp. 1–7 24. M. Simsek, B. Kantarci, A. Boukerche, Self organizing feature map-integrated knowledgebased deep network against fake crowdsensing tasks, in GLOBECOM 2020–2020 IEEE Global Communications Conference (IEEE, Taipei, Taiwan, 2020), pp. 1–6 25. M. Simsek, B. Kantarci, Y. Zhang, Detecting fake mobile crowdsensing tasks: ensemble methods under limited data. IEEE Veh. Technol. Mag. 15(3), 86–94 (2020) 26. M. Simsek, B. Kantarci, A. Boukerche, Utility-aware legitimacy detection of Mobile Crowdsensing tasks via knowledge-based self organizing feature map. IEEE Trans. Mob. Comput. 22(6), 3706–3723 (2023). https://doi.org/10.1109/TMC.2021.3136236 27. M. Simsek, A. Boukerche, B. Kantarci, S. Khan, AI-driven autonomous vehicles as COVID19 assessment centers: a novel crowdsensing-enabled strategy. Pervasive Mob. Comput. 75, 101426 (2021) 28. M. Simsek, B. Kantarci, A. Boukerche, Collaborative self organizing map with DeepNNs for fake task prevention in Mobile Crowdsensing, ICC 2022 - IEEE International Conference on Communications, (Seoul, Korea, Republic of, 2022), pp. 4794–4799. https://doi.org/10.1109/ ICC45855.2022.9838920

AI-Driven Attack Modeling and Defense Strategies Against Fake Task. . .

297

29. A. Sood, M. Simsek, Y. Zhang, B. Kantarci, Deep learning-based detection of fake task injection in mobile crowdsensing, in 2019 IEEE Global Conference on Signal and Information Processing (GlobalSIP) (IEEE, Ottawa, ON, Canada, 2019), pp. 1–5 30. J. Vesanto, E. Alhoniemi et al., Clustering of the self-organizing map. IEEE Trans. Neural Netw. 11(3), 586–600 (2000) 31. K. Villela, C. Nass, R. Novais, P. Simoes, A. Traina, J. Rodrigues, J. M. Menendez, J. Kurano, T. Franke, A. Poxrucker, Reliable and Smart Decision Support System for Emergency Management Based on Crowdsourcing Information, (Springer International Publishing, Cham, 2018). pp. 177–198 32. Y. Wang, Z. Yan, W. Feng, S. Liu, Privacy protection in mobile crowd sensing: a survey. World Wide Web 23(1), 421–452 (2020) 33. L. Xiao, D. Jiang, D. Xu, W. Su, N. An, D. Wang, Secure mobile crowdsensing based on deep learning. China Commun. 15(10), 1–11 (2018) 34. L. Xiao, Y. Li, G. Han, H. Dai, H.V. Poor, A secure mobile crowdsensing game with deep reinforcement learning. IEEE Trans. Inf. Forensics Secur. 13(1), 35–47 (2018) 35. X. Zhang, Z. Yang, W. Sun, Y. Liu, S. Tang, K. Xing, X. Mao, Incentives for mobile crowd sensing: a survey. IEEE Commun. Surv. Tutorials 18(1), 54–67 (2016) 36. Y. Zhang, B. Kantarci, Invited paper: AI-based security design of mobile crowdsensing systems: Review, challenges and case studies, in 2019 IEEE International Conference on Service-Oriented System Engineering (SOSE) (IEEE, San Francisco East Bay, CA, USA, 2019), pp. 17–1709 37. Y. Zhang, M. Simsek, B. Kantarci, Machine learning-based prevention of battery-oriented illegitimate task injection in mobile crowdsensing, in Proceedings of the ACM Workshop on Wireless Security and Machine Learning - WiseML 2019 (ACM Press, Miami, FL, USA, 2019), pp. 31–36 38. Y. Zhang, M. Simsek, B. Kantarci, Self organizing feature map for fake task attack modelling in mobile crowdsensing, in 2019 IEEE Global Communications Conference (GLOBECOM) (IEEE, Waikoloa, HI, USA, 2019), pp. 1–6 39. Y. Zhang, M. Simsek, B. Kantarci, Ensemble learning against adversarial AI-driven fake task submission in mobile crowdsensing, in ICC 2020–2020 IEEE International Conference on Communications (ICC) (IEEE, Dublin, Ireland, 2020), pp. 1–7 40. Y. Zhang, M. Simsek, B. Kantarci, Empowering self-organized feature maps for ai-enabled modeling of fake task submissions to mobile crowdsensing platforms. IEEE Int. Things J. 8(3), 1334–1346 (2021)

Traceable and Secure Data Sharing in Mobile Crowdsensing Jinwen Liang and Song Guo

1 Introduction With the proliferation of mobile devices with a rich set of built-in sensors, including accelerometers, cameras, smartphones, etc., mobile crowdsensing (MCS) has attracted considerable interests in edge computing, which collects sensory data passively or actively based on the allocated tasks and shares the sensory data to remote data user as requested [20]. With various sensors distributed in different areas, the emerging MCS technologies endow remote users with sensory data request and retrieval [22]. By leveraging the power of data sensing and gathering, MCS benefits a wide range of real-world applications such as traffic monitoring, environment monitoring, health data analytics, public safety, and so on [19]. Namely, MCS is a way that seeks to leverage large groups of people for incentive data collection and empower data-driven applications and services [36]. According to the report in [31], MCS is a key technology behind smart cities, which enables many smart city applications such as parking appointment. Generally, an MCS system contains a group of data owners, a group of data users, and a cloud server. As shown in Fig. 1, the workflow of mobile crowdsensing contains two stages, i.e., task allocation and data sharing. • Task allocation is a stage when specific data users provide data search requests. After receiving the search requests from data users, the cloud server allocates the crowdsensing tasks to an appropriate group of data owners so that the data collection tasks can be accomplished on time. In the task allocations stage, the cloud server aims to satisfy the requests from users and allocate tasks to specific

J. Liang () · S. Guo The Hong Kong Polytechnic University, Hong Kong, China e-mail: [email protected]; [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 J. Wu, E. Wang (eds.), Mobile Crowdsourcing, Wireless Networks, https://doi.org/10.1007/978-3-031-32397-3_12

299

300

J. Liang and S. Guo

Fig. 1 The workflow of MCS

locations for satisfying the location requests from data users and reducing the time delays. • Data sharing is a stage when specific data owners incentively accomplish the data search task and transmit the sensory stream data to the cloud server. After gathering the sensory stream data from data owners, the cloud server feeds the data user with the desire data. In the data sharing stage, the cloud server aims to aggregate the collected data from data owners and transmit the data to data users based on an access control policy. Due to limited storage and computational resources of mobile devices, a cloud server is leveraged to transmit sensory stream data and allocate MCS tasks [9]. With the powerful computational and storage capability of the cloud server, data owners reduce the computational and storage costs on resourced-limited mobile devices. Apart from the attractive benefits of cloud computing, storing, and sharing valuable data via a public cloud may raise privacy concerns because the public cloud server is always considered as not honest [18, 23]. To protect the confidential sensory data from the data owner, encrypting the data before outsourcing to the cloud server is a potential method for privacy-preserving data sharing in mobile crowdsensing [17, 27]. Once the content of data is encrypted, how to perform data sharing on encrypted data without sacrificing the usability has become a key challenge in mobile crowdsensing. In fact, the concept of identity-based encryption (IBE) [11] is introduced for enabling privacy-preserving and flexible key management for data owners to share data. To avoid data abuse, an attribute-based access control policy can manage data users’ access eligibility, which is an essential policy in MCS due to the multi-users setting [16]. Namely, the data owner can determine which data users are authorized to access the data based on the embedded access

Traceable and Secure Data Sharing in Mobile Crowdsensing

301

policies, which include the attribute or identity of authorized users. Meanwhile, an authorized user may share the private key to unauthorized users for monetary reasons or accidentally leak the private key to malicious users. Thus, traceability should be enabled for avoiding secret key leakage in mobile crowdsensing due to the multi-user setting. Furthermore, to avoid additional communication costs, interactions between data owners and data users should be avoided. Therefore, data sharing systems for MCS should be efficient, privacy-preserving, traceable, and noninteractive. To achieve privacy-preserving data sharing in MCS, a significant amount of schemes have been developed [5, 15]. IBE-based schemes support privacypreserving data sharing in MCS [11, 32]. Unfortunately, IBE-based schemes force the data owner to share the whole datasets with data users and therefore cannot enable flexible data sharing functionality. To realize flexible and privacy-preserving data sharing in MCS, attributed-based encryption (ABE) endows data owners embed access policies into the sensory data, which ensures that only authorized data users that satisfied the access policies can retrieve the sensory data [10, 16]. However, traditional ABE solutions require data owners always stay online for encrypted sensory data decryption, which is not efficient for data sharing in MCS because of large scale and volume of encrypted sensory data. To achieve flexible data sharing in MCS, several access control schemes have been developed [21, 35]. However, the scheme in [35] requires interactions between data owners and data users for access policy generation, which may incur prohibitive communication costs. Furthermore, the aforementioned schemes did not consider the traceability of secret keys, which can hardly trace unauthorized secret key sharing. The ABE-based data sharing scheme in [38] designs a tracing method in attribute sets, which can be utilized for enabling traceability. However, it is difficult for the ABE-based traitor in [38] to identify which data user leaks a secret key, and the tracing system in [38] runs parallelly to the encryption system, which may lead to additional key leakage risks in MCS. Followed by this work, some non-interactive data sharing schemes [5, 8] have been proposed by leveraging the IBE-based technique. Considering the aforementioned drawbacks, Song et al. developed a Traceable and privacy-preserving non-Interactive Data Sharing (TIDS) scheme [25], which performs traceable data sharing efficiently and securely in MCS [25]. Specifically, proxy re-encryption and attribute-based access policy are utilized in TIDS for flexible data sharing in MCS with multiple users while protecting the privacy of each participant. To achieve traceability toward malicious users, TIDS leverages traceable ciphertext policy attribute-based encryption (CPABE) with the key sanity check technique by a trusted authority. The contributions of TIDS can be summarized as twofold. • First, TIDS is a flexible and non-interactive privacy-preserving data sharing scheme, which enables multi-user sharing and low communication costs. To endow data re-encryption capability for the cloud server, a ciphertext conversion protocol in the multi-user sharing system is developed, which enables authorized users to retrieve and decrypt the re-encrypted ciphertexts.

302

J. Liang and S. Guo

• Second, TIDS achieves malicious user tracing and prevents secret key leakage by leveraging traceable CPABE. Security analysis shows that TIDS is secure against the semi-honest adversary in the Chosen Plaintext Attack (CPA) model. Performance evaluation demonstrates that comparing with the existing scheme in [5], TIDS is more efficient for flexible data sharing in the multi-user setting. The remainder of this chapter is organized as follows. Section 2 provides literature reviews for related work. Section 3 describes the TIDS scheme [25] as an example for traceable and secure data sharing scheme in MCS. Section 4 concludes the chapter and indicates several future research topics of traceable and secure data sharing in MCS.

2 Related Work In this section, we introduce related work of mobile crowdsensing, privacyenhancing techniques for mobile crowdsensing, and Blockchain-Based Traceable Mobile Crowdsensing.

2.1 Mobile Crowdsensing In recent years, mobile crowdsensing (MCS) has become an emerging sensing paradigm, which enables incentive remote data collection [28]. Data collection tasks once need a lot of manpower and material resources could be completed efficiently through MCS. For instance, MCS has been utilized in traffic monitoring scenarios for traffic congestion prediction, which thus avoids the involvement of numerous traffic police [29]. The concept of MCS was introduced by Ganti et al., which denotes a more general infrastructure than mobile phone sensing [7]. A more comprehensive differences declaration is provided by Guo et al. [13], which shows that MCS empowers ordinary citizens to incentively contribute sensory data from their mobile devices and later aggregates the data in a cloud for crowd intelligence. From the perspective of data sharing, MCS is a universal platform for urban sensing and public information sharing [29]. Through the 5G wireless networks, sensory data collected from data owners’ mobile devices can be easily transmitted to data users [12]. With cross-space data sharing functionality, MCS benefits realworld data-driven applications such as targeted advertising, mobile socializing, etc.

Traceable and Secure Data Sharing in Mobile Crowdsensing

303

2.2 Privacy-Enhancing Techniques for Mobile Crowdsensing Privacy-enhancing techniques for mobile crowdsensing have been investigated for several years. Since the chapter focuses on traceable and secure data sharing in MCS, we organized the privacy-enhancing techniques as privacy-preserving techniques and traceable techniques. Privacy-Preserving Techniques Recently, a significant amount of privacypreserving MCS schemes have been proposed for preserving the privacy of identities, locations, and the sensory data [34]. Ciphertext policy attribute-based encryption, which is promising to identify fine-grained access control policy, is always utilized to develop flexible and secure data sharing schemes in MCS [10, 16]. Unfortunately, most of the ABE-based data sharing schemes may incur prohibitive communication and computational costs because both the size of ciphertexts and the number of pairing operations during the decryption process grow linearly with respect to the complexity of the access control structure [5, 14]. Different from the ABE-based schemes, IBE-based schemes achieve secure data sharing in MCS with efficient encryption and decryption [11, 32]. However, IBE-based schemes cannot achieve flexible data sharing because these schemes require data owners to share the whole dataset with the data user. Song et al. developed TIDS, an efficient and flexible data sharing scheme with fine-grained access control, which will be described as an example in Sect. 3. Tracing Techniques Malicious data user tracing enables a content distributor such as a trusted authority to track the non-compliant behavior of a pirate decoder or a malicious user [1, 3]. A malicious data user or traitor may share the secret key to unauthorized data users. Recently, Boneh and Zhandry have studied a malicious data user tracing scheme by utilizing obfuscation-related objects [3]. Abdalla et al. developed an identity-based traitor tracing scheme. However, the identity-based traitor tracing scheme requires an exponential number of instances, which is not efficient for data sharing in MCS [1]. Zhang et al. [37] proposed a write-box ciphertext policy attribute-based encryption scheme with traceability. The scheme in [37] supports large-scale universe access structures and can be deployed in intelligent eHealth systems [39]. Boneh et al. also developed a learning with errorsbased malicious data user tracing scheme, which enables key tracing publicly. Compared with the aforementioned cryptographic-based tracing schemes, TIDS [25] achieves a competitive computational costs without interaction between data owners and data users, which endows resource-limited mobile devices (owned by data owners) with efficient malicious data user tracing.

304

J. Liang and S. Guo

3 Traceable and Privacy-Preserving Non-interactive Data Sharing (TIDS) Scheme In this section, we describe a Traceable and privacy-preserving non-Interactive Data Sharing (TIDS) scheme as an example to show how to achieve traceable and secure data sharing in MCS [25].

3.1 Problem Statement In this subsection, we state the research problem of TIDS [25]. We first introduce the system model of TIDS in MCS, which illustrates the entities involved in the system. Then, we provide a threat model for TIDS, which provides essential assumptions and potential threats in TIDS. Finally, we illustrate the design goals of TIDS.

3.1.1

System Model

As shown in Fig. 2, TIDS includes four entities, i.e., data owners (.DOs), data users (.DUs), a trusted authority (.TA), and a cloud server (.CS). • Trusted Authority (.TA): The .TA produces encryption keys and re-encryption keys for each party. Namely, the encryption keys are delivered to .DO and .DU, and the re-encryption keys are transmitted to .CS. The .DU receives two types of secret keys, the one is secret keys that associate with the attributes of .DU, and the other is a re-encryption key for .DU to convert an original ciphertext to re-

Fig. 2 System model

Traceable and Secure Data Sharing in Mobile Crowdsensing

305

encrypted ciphertexts. To achieve traceability, .TA can trace whether a malicious DU leaks or sells its secret key to unauthorized users by tracking .DU’s secret keys. • Cloud Server (.CS): .CS is a cloud server with powerful computational resources and sufficient storage capabilities for allocating crowdsensing tasks to .DOs, receiving sensory data from .DOs, and sharing sensory data to .DU. To enable privacy-preserving data sharing, original ciphertexts are re-encrypted by .CS according to .DOs’ designed access policies. • Data Owners (.DOs): .DOs are a group of data owners that incentively collect required sensory data and transmit the encrypted sensory data to .CS. Meanwhile, .DOs embed an access policy into the encrypted sensory data, which indicates whether a .DU is eligible to retrieve the encrypted data. Then, .CS converts the encrypted sensory data to another encrypted form that authorized .DU can retrieve and decrypt the desired sensory data. • Data Users (.DUs): .DUs produce crowdsensing tasks to .CS and receive the desired encrypted data from .CS. After receiving the encrypted sensory data, .DUs leverage their secret keys to decrypt the data. As a result, authorized users obtain their desired crowdsensing data. For malicious data users who may abuse their secret, .TA will trace the malicious data user via the decryption secret keys. .

3.1.2

Threat Model

The threat model assumes both .TA and .DOs are fully trusted. The .CS is assumed to be honest-but-curious adversary, which will execute the TIDS scheme correctly but is curious about the content of sensory data. Most of the DU s are assumed to be honest but some of them behave maliciously by selling their secret keys to unauthorized data users for monetary reasons. According to recent works in [24, 26], the threat model adopts the assumption of no collusion between .CS and .DUs because collusion may reduce the reputation of both .CS and .DUs [30] and economic incentive schemes can be utilized for avoiding the collusion attack [6]. Considering a malicious data user may sell the secret key to unauthorized data users, the content of shared data will be leaked. By leveraging game theory techniques and punishment methods, the problem of key-ciphertext pairs tracing can be easily addressed [33]. In the threat model, the adversary .A has oracle access to the re-encryption key generation algorithm, which is .rki→q ← ReKeyGen(ski , pkq ), where .(ski , pkq ) is an input pair generated by KeyGen. The threat model assumes that .A has the ability of selecting arbitrary plaintexts and obtaining the corresponding ciphertexts, i.e., Chosen Plaintext Attack (CPA). Furthermore, .A is curious about personal information including: • Data privacy: The content of sensory data in the database. In TIDS, the sensory data are sensitive data of data owners, which collect desired data for data users.

306

J. Liang and S. Guo

• Key privacy: The secret keys of the data owners. The secret keys of the data owners can be utilized for data decryption and therefore are sensitive data of data owners.

3.1.3

Design Goals

The design goal of TIDS is to support traceable, secure, and flexible data sharing in mobile crowdsensing [25]. The following design goals are considered in TIDS. • Data Confidentiality: Data confidentiality should be ensured. Sensitive information of sensory data should not be leaked to .CS, which means that both original and re-encrypted ciphertexts will not reveal the content of sensory data. • Traceability: Traceability should be enabled. Malicious data users who transmit their secret key to unauthorized data users should be traced. • Non-interaction: There are no interactions between .DOs and .DUs. On the one hand, .DOs are able to produce fine-grained access policies for privacy-preserving data sharing without communicating with .DUs. On the other hand, in the reencryption key generation phase, .DO should not be requested to communicate with .DU for re-encryption key generation. • Multi-user Sharing: The data sharing functionality should be enabled in MCS with a multi-user setting. Namely, .DOs can share their sensory data with desired .DUs.

3.2 Preliminaries In this subsection, several preliminaries are defined. Specifically, the building blocks include bilinear paring and access structures.

3.2.1

Bilinear Pairings

Let .G and .G1 be two multiplicative cyclic groups with the same prime order p. The generator of group .G is g. Let .e : G × G → G1 be a bilinear mapping, which has the properties as follows: • Bilinear: Let .∀ u, v ∈ G and all .a, b ∈ Zp , then .e(ua , v b ) = e(u, v)ab ∈ G1 . • Non-degenerated: .∃ g ∈ G, s.t. .e(g, g) = 1. • Computable: Given .∀ u, v ∈ G, computing .e(u, v) is efficient. According to recent cryptographic works [8, 25], the definitions of Decisional Diffie–Hellman (DDH) Problem [4] and Computational Diffie–Hellman (CDH) Problem can be given as follows.

Traceable and Secure Data Sharing in Mobile Crowdsensing

307

Definition 1 (DDH) Given group elements g, .g a , .g b , and .g c , where .a, b, c ∈ Zp and .g, g a , g b , g c ∈ G, it is difficult to determine whether .g c = g ab holds. Definition 2 (CDH) Given a tuple .(g, g a1 , g a2 ) ∈ G, where .a1 , a2 ∈ Zp are randomly selected values, it is difficult to compute the value .g a1 a2 .

3.2.2

Access Structure

In TIDS, the definition of access structure is similar to the definition in [2], which is given as follows. Definition 3 (Access Structure) Let U be the attribute universe. If .A is a monotone class of subsets of .[n], then .A ∈ {0, 1}U is an access structure for n parties. Given any .I1 , I2 from the access structure, if .I1 ∈ A and .I1 ⊆ I2 , then .I2 ∈ A. Therefore, the authorized sets are the sets in .A and the unauthorized sets are the sets not in .A.

3.3 The TIDS Scheme The TIDS scheme, which is constructed by utilizing the proxy re-encryption technique and the Linear Secret Sharing Scheme (LSSS) technique, is described in this subsection. Compared with previous work, TIDS extends previous ciphertext policy attribute-based encryption-based data sharing scheme with traceability. Specifically, first, the framework of TIDS is produced. Second, the detailed construction of TIDS is provided, which includes seven algorithms. Namely, SystSetup, KeyGen, Enc, ReKenGen, ReEnc, Dec, and Trace.

3.3.1

The TIDS Framework

To implement flexible and secure data sharing in MCS, TIDS [25] jointly leverages Linear Secret Sharing Scheme (LSSS) [2] and proxy re-encryption. By converting an original ciphertext to a transformative ciphertext, which can be decrypted by .DU that satisfies the access policy, TIDS achieves flexible and secure data sharing in MCS with multiple data users. To achieve secure data sharing, the .CS transmits an original ciphertext to .DU, who utilizes his/her secret key to decrypt the ciphertext and obtains the secret key of .DO. Then, .DU decrypts the original ciphertext for desired sensory data by utilizing the secret key of .DO. When a .DU transfers the secret key to other unauthorized users, .TA can trace the malicious behavior by utilizing the key sanity check method. The framework of TIDS is shown in Fig. 3, which involves four entities and seven phases. The procedure of TIDS is provided as follows:

308

J. Liang and S. Guo

Fig. 3 The framework of TIDS

• SystSetup: the system is initialized and public parameters P P and the master key MSK are produced by .TA. • KeyGen: .TA produces a private key for a data user .uq . Then, .TA produces public/secret key pairs .(pki , ski ), .(pkq , skq ) for the data owner .ui and the data user .uq , respectively. • Enc: The sensory data m and the secret key .ski are encrypted as .Ci and CT , respectively, by .ui based on an access policy .(M, ρ) and the public key .pki . • ReKeyGen: a re-encryption key .rki→q is produced by .TA and later be transmitted to .CS for ciphertext conversion. • ReEnc: by leveraging a novel access policy .(M , ρ ), a novel ciphertext .Cq is converted from an original ciphertext .Ci by .CS. • Dec: if .uq is a new user that satisfies .(M , ρ ), .uq decrypts .Cq as m by using .skq . If .uq is a user who satisfies .(M, ρ), .uq utilizes the private key to decrypt CT as .ski . Then, .uq utilizes .ski to decrypt the original ciphertext .Ci and obtains the corresponding sensory data m. • Trace: The user who shares the secret key to unauthorized data users will be traced by .TA, who can recover the real identity of the malicious users.

3.3.2

The Detailed Description of TIDS

(1) SystSetup(1 ) → (P P , MSK). With a security parameter  and the attribute universe U , TA runs the SystSetup for system initialization. Firstly, TA selects a bilinear mapping e : G × G → G1 randomly , where both G and G1 are cyclic groups with the same prime order p. Let g be the generator of G. Then, TA randomly chooses u, h, ω, v ∈ G, and α, a ∈ Zp . Next, TA defines a hash function H : {0, 1}∗ → K and a cryptographic symmetric encryption and symmetric decryption pair (SEnc, SDec) with two different secret keys

Traceable and Secure Data Sharing in Mobile Crowdsensing

309

k1 , k2 ∈ K, where K is a key space. According to Shamirs (k, n) threshold scheme [2], TA generates an instance I N S(k,n) and a one-way function f (x). Finally, TA sets P P and MSK as follows. P P = (g, e(g, g), u, h, ω, v, g a , e(g, g)α , H ),

(1)

MSK = (α, a, k1 , k2 ).

(2)

.

.

(2) KeyGen(MSK, S, idq ) → SKidq ,S . Let idq represent the identity of uq , and S denotes the attribute set. Given the master key MSK, TA first chooses several random values r0 , r1 , . . . , rk ∈ Zp , and then TA calculates x = SEnck1 (idq ), α y = f (x), D = SEnck2 (x||y), D = g a+D ωr0 , Y = g r0 , Y = g ar0 , Dj,1 = g rj , Dj,2 = (uS h)rj v −(a+D )r0 , where j ∈ [1, k]. The private key SKidq ,S of uq can be set as shown in Eq. (3), which is used to recover the plaintext from the original ciphertext. SKidq ,S = (D, D , Y, Y , Dj,1 , Dj,2 ).

.

(3)

Besides, TA generates a public/secret key pair (pki , ski ) = (g ski , ski ) for the data owner ui and a public/secret key pair (pkq , skq ) = (g skq , skq ) for the data user uq , respectively. The secret key ski of ui and the public key g skq of uq are used to generate a re-encryption key rki→q . (3) Enc1 (pki , m, (M, ρ), A) → Ci . The sensitive sensory data should be protected in the untrusted cloud environment. To this effect, ui encrypts the sensory data with the attributes of the specified users before outsourcing the sensory data to CS. Given an access structure A, ui generates an attribute-based access policy as (M, ρ) based on the construction of LSSS, where M is an l × n matrix and ρ is a function that transforms each row of M to an attribute. Then, ui executes the Enc1 algorithm as follows: • Step 1: Given the sensory data m, ui first selects a random value ri ∈ Zp and then encrypts the sensory data m by using ui ’s public key pki as shown in Eq. (4). Ci = Enc(pki , m) = (g ri ski , m · Z ri ),

.

(4)

where Z = e(g, g). • Step 2: ui outsources the ciphertext Ci along with his/her identity and access policy (M, ρ) to CS. Enc2 (ski , A, (M, ρ)) → CT . Given the secret key ski and the access policy (M, ρ), ui selects a random vector v = (s, y2 , . . . , yn ) that is devoted to share the random secret s, where y2 , . . . , yn ∈ Znp . Then, ui encrypts the secret key ski as follows:

310

J. Liang and S. Guo

• Step 1: For any i ∈ [1, l], DO computes λi = Mi ◦ v = Mi,1 s + Mi,2 y2 + · · · + Mi,n yn , where Mi is the corresponding vector of i-th row of M, Mi ◦ v is an inner product of Mi and v, and Mi,j is the j -th entry of Mi . Then, the vector of the shares can be represented as λ = (λ1 , λ2 , . . . , λl ) = Mv. • Step 2: ui randomly chooses a real value ti ∈ Zp for each i ∈ [1, l] and encrypts his secret key ski as CT , which is shown in Eq. (5). CT =(C = ski · e(g, g)αs , (M, ρ), C0 = g s , C1 = g as , .

{Ci,1 = ωλi v ti , Ci,2 = (uρ(i) h)−ti , Ci,3 = g ti }i∈[1,l] ).

(5)

• Step 3: Then, ui transmits CT and the identity of the specified user to CS. To achieve flexible and multi-user sharing, ui formulates a new access structure A and a new access policy (M , ρ ) and sends them to CS for ciphertext re-encryption. By doing so, CS can convert the original ciphertext Ci into another transformative ciphertext that can be decrypted by a new user who satisfies the new access policy (M , ρ ). (4) ReKeyGen(ski , g skq ) → rki→q . Given the secret key ski of ui and the public key g skq of a new user uq , TA first generates a re-encryption key as rki→q = g skq /ski for ciphertext re-encryption. Then, TA sends rki→q along with the attribute set S of the new user uq to CS. (5) ReEnc(rki→q , S, A , (M , ρ ), Ci ) → Cq . Upon receiving the original ciphertext Ci , the encrypted secret key CT , and the new access policy (M , ρ ), CS converts Ci into a new ciphertext Cq by using the re-encryption key rki→q . • Step 1: CS first checks whether the attribute set S ∈ A holds. If so, S satisfies the new access policy (M , ρ ), and then CS converts Ci into the re-encrypted ciphertext Cq as shown in Eq. (6). Cq = (e(g ri ski , g skq /ski ), m · Z ri ) .

= (Z ri skq , m · Z ri ).

(6)

• Step 2: CS returns the transformative ciphertext Cq to the new user who satisfies the new access policy (M , ρ ). (6) Dec1 (Cq , skq ) → m. If uq is a new user, uq will receive the transformative ciphertext Cq . Then, the new user uq uses the secret key skq to decrypt the ciphertext Cq and then obtains the plaintext m as shown in Eq. (7). m = Dec(skq , Cq ) =

.

m · Z ri . (Z ri skq )1/skq

(7)

Dec2 (SKidq ,S , CT ) → ski /⊥. If uq is a specified user, uq will receive the original ciphertext Ci . Subsequently, uq performs the Dec2 algorithm to decrypt the ciphertext CT and then obtains the secret key ski . By using the

Traceable and Secure Data Sharing in Mobile Crowdsensing

311

secret key ski , uq can decrypt the original ciphertext Ci and obtain the sensory data m. Based on the property of LSSS, the authorization check can be correctly executed. If S does not belong to the access structure A, then uq cannot satisfy (M, ρ), and the Dec2 algorithm  outputs ⊥. Otherwise, uq can find some constants {ωi |ωi ∈ Zp } such that i∈I ωi Mi = (1, 0, . . . , 0), where I = {i|ρ(i) ∈ S} and S ∈ A. • Step 1: If S is an authorized set, i.e., S satisfies (M, ρ), then uq first computes



B1 = e(D, C0D C1 ) = e(g, g)αs e(ω, g)(a+D )sr0 ,  = (e(Y D Y , Ci,1 )e(Dj,1 , Ci,2 )e(Dj,2 , Ci,3 ))ωi B 2 . i∈I

= e(ω, g)(a+D )r0 s . • Step 2: Next, according to the above values B1 and B2 , uq computes B = B1 /B2 = e(g, g)αs . • Step 3: Then, uq calculates and obtains the secret key of DO by calculating ski = C/B. • Step 4: Finally, uq can obtain the shared sensory data from the original ciphertext Ci by using the secret key ski . (7) Trace(I N S(k,n) , MSK, SKidq ,S ) → idq /⊥. When a secret key is abused by a malicious user, the real identity of malicious users can be recovered by TA from the secret key. Before tracing the malicious user, TA verifies whether the secret key is well formed by utilizing the key sanity check method. The private key SKidq ,S can successfully pass the key sanity check when the following conditions hold: • SKidq ,S is in the form of (D, D , Y, Y , Dj,1 , Dj,2 ), D ∈ Zp , and D, Y, Y , Dj,1 , Dj,2 ∈ G. • e(g, Y ) = e(g a , Y ). • e(g a gD , D) = e(g, g)α e(Y Y D , ω).  • e(h, j ∈[1,k] Dj,1 )e(u, j ∈[1,k] Dj,1 ) = e(g, j ∈[1,k] Dj,2 )e(Y Y D , v). If SKidq ,S does not pass the key sanity check, then the private key SKidq ,S is not well formed and the Trace algorithm outputs ⊥. Otherwise, the private key SKidq ,S is well formed, and the Trace algorithm identifies the traitor of SKidq ,S by the following operations. TA first extracts (x, y) from x||y = SDeck2 (D ) by using the secret key k2 . Then, the identity of the malicious user can be recovered by calculating idq = SDeck1 (x) with the secret key k1 . In TIDS, the Trace algorithm is utilized to trace whether a malicious data user shares the whole entries of the secret key with unauthorized data users. When a part of the secret key is shared to unauthorized data user, the unauthorized data user who possesses a part of the secret key cannot decrypt

312

J. Liang and S. Guo

the ciphertext Ci and therefore cannot obtain the sensory data m. Therefore, TA can leverage key sanity check to verify the sanity of the secret key and obtain the real identity of malicious user via the Trace algorithm.

3.4 Security Analysis In this subsection, security and traceability proofs are produced to demonstrate that TIDS achieves the design goals in Sect. 3.1.3. Before providing a formal security proof, the definition of Chosen Plaintext Attacks (CPA) is restated in Definition 4.  Definition 4 (CPA Security) For the TIDS scheme . , we associate the experiment with a CPA adversary .A. If .Cb denotes the challenge ciphertext, .A cannot distinguish which plaintext is encrypted. That is, TIDS is secure against CPA if for CP A () is negligible any polynomial-time adversary .A, the advantage function .Adv ,A  A () = 1] − 1 | ≤ negl(). in ., i.e., .|P r[ExpCP ,A 2

As shown in Fig. 4, a security game between an adversary .A and a challenger .C is developed. According to the security experiment in Fig. 4, the advantage of the adversary .A is defined as 1 CP CP A  A Adv ,A () = |P r[Exp ,A () = 1] − 2 |.

.

Then, the formal CPA security definition is provided as follows. Theorem 1 The TIDS guarantees the confidentiality of the data and the secret keys of .DO against Chosen Plaintext Attack. We prove Theorem 1 as follows.

Fig. 4 Security experiment

Traceable and Secure Data Sharing in Mobile Crowdsensing

313

Proof Data confidentiality. In TIDS, .DO utilizes the public key .pki to encrypt sensory data m. Then, .CS utilizes the re-encryption key .rki→q to re-encrypt the  A () encrypted sensory data. As shown in Fig. 4, the security experiment .ExpCP ,A processed between an adversary .A and a challenger .C can be simulated.  A (), .A has oracle access to .mi In Phase 1 and Phase 2 of the experiment .ExpCP ,A and obtains the corresponding ciphertext .Ci . For some .ri ∈ Zp , we assume that .g ski can be regarded as .g ri skq . Note that .Ci = (g ri ski , mi · T ), if .T = e(g, g)ski /skq = e(g, g)ri skq /skq = e(g, g)ri , then .Ci is a correct encrypted form of the plaintext .mi . Otherwise, .Ci is a false encrypted form of .mi . Since calculating .mi from .Ci is equal to solve the DDH problem [4], the encryption algorithm .Enc1 is semantic secure, which means that the ciphertexts produced by .Enc1 are indistinguishability under chosen plaintext attack. Therefore, the adversary cannot distinguish .DB0 from .DB1 . Similarly, since the security property of ReEnc is also based on the DDH problem, the ReEnc algorithm is secure against CPA [11]. Therefore, given a ciphertext .Cb , which is encrypted from the adversary selected plaintext .mi , .A can hardly determine whether .m0 or .m1 is .mi in the security game. Afterward, we have 1  A () = 1] − | ≤ negl(). |P r[ExpCP ,A 2

.

Therefore, the TIDS achieves data confidentiality against chosen plaintext attack. Key confidentiality. Since the secret key .ski of .DO is encrypted as .C = ski · e(g, g)αs , the confidentiality of the secret key is ensured. Generally, .CS is curious about the secret key .ski . However, recovering .ski from C is equivalent to dealing with a CDH problem. Since solving the CHD problem in the random oracle model is difficult [8], the confidentiality of secret keys is ensured. Therefore, the TIDS scheme ensures the confidentiality of both sensory data and .DO’s secret key against chosen plaintext attack.   The traceability of TIDS is defined as follows. Lemma 1 (Traceability) The TIDS scheme achieves traceability for any malicious users who intentionally leak their secret keys. The traceability of TIDS can be proved as follows. Proof To prove the traceability of TIDS, we should show that malicious data users’ real identity could be recovered. Assume that .uq shares the secret key with others. By leveraging the private key .SKidq ,S , .TA can recover and trace the malicious data user. Specifically, .TA verifies whether the secret key is well formed by utilizing the key sanity check scheme in the Trace algorithm. Note that the identity information of .DU is encrypted as follows: x = SEnck1 (idq ).

.

Theoretically speaking, .TA first extracts .(x, y) from .x||y = SDeck2 (D ) by using the secret key .k2 . Then, .TA can recover the user’s real identity by computing

314

J. Liang and S. Guo

Table 1 Comparison of computational cost Phase Key generation Encryption Re-encryption Decryption Trace

Entity TA DO CS DU TA

TIDS (k + 3)Tr + |S|Te + Tp Te + nTp (n + 1)Tp 2Tp 6Tp + 3Te

FPDS [5] (k + 6)Tr + |S|Te + Tp (3n + 4)Te (3n + 1)Tp 5Tp −

idq = SDeck1 (x) with the secret key .k1 . Hence, the Trace algorithm endows the TIDS scheme with traceability.  

.

3.5 Performance Analysis In this subsection, we provide comprehensive comparison between the TIDS [25] and the FPDS scheme [5]. First, theoretical analysis in terms of computational and storage costs is provided to show the advantages of TIDS. Second, experimental evaluations are provided to show the computational performance advantages of TIDS in terms of time costs.

3.5.1

Theoretical Analysis

The comparison of computational cost is shown in Table 1. In TIDS, we employ some cryptographic operations including exponentiation and bilinear pairing. Let .Te , .Tp , and .Tr be the time cost for each exponentiation, bilinear pairing, and random number generation, respectively. In the KeyGen algorithm on the .TA side, .TA selects .k + 3 random numbers in .G and executes 1 bilinear mapping and .|S| exponentiations, where .|S| is the size of attribute set. In the encryption phase, data owner .ui executes exponentiation and bilinear pairing for the plaintext m, i.e., the computational cost of encryption is .Te + nTp , where n is the number of attributes. Furthermore, an original ciphertext is converted to a re-encrypted ciphertext by .CS for enabling multi-user data sharing. In other words, .CS checks whether .S ∈ A holds and performs bilinear pairing for original ciphertext conversion. Hence, the computational cost of ReEnc is .(n + 1)Tp , which is linear to the factor n for flexible data sharing. In the decryption phase, .DU executes 2 bilinear maps to decrypt the re-encrypted ciphertext, and the computational cost of decryption is .2Tp . In the Trace phase, TIDS performs 6 bilinear maps and 3 exponentiations for the key sanity check, and thus the time cost should be .6Tp + 3Te . The storage overhead comparison between TIDS and the FPDS scheme in [5] is shown in Table 2. In TIDS, the public key contains 2 entries in .G and 2 entries in

Traceable and Secure Data Sharing in Mobile Crowdsensing Table 2 Comparison of storage overhead

Entity size Public key Secret key Original ciphertext Re-encrypted ciphertext

TIDS .2|G| + 2|G1 | .(|S| + 2)|G| .(n + 3)|G| .2|G1 |

315 FPDS [5] .(k + 8)|G| + |G1 | .(|S| + 2)|G| .(2n + 3)|G| + |G1 | .4|G| + |G1 |

G1 , which are constant values. Different from TIDS, the size of the public key in the FPDS scheme in [5] grows linearly with respect to the number of random values. With the benefits of short size of public keys, TIDS is more suitable for constructing MCS data sharing systems with large universal attributes. The secret key of TIDS contains .|S| + 2 entries in .G. To achieve fine-grained access control, the secret size of FPDS is the same as TIDS. Furthermore, since the size of the original ciphertext in TIDS is smaller than that of FPDS, the size of transformative ciphertext of TIDS is a constant that is smaller than that of FPDS.

.

3.5.2

Experimental Evaluations

Baselines We compare the TIDS with the FPDS scheme in [5], which achieves flexible data sharing via attribute-based encryption. Compared with FPDS, the TIDS achieves secure data sharing by utilizing attribute-based encryption and implements traceability, which was not considered in FPDS. Experimental Settings To evaluate the performance of each algorithm in TIDS, experiments are conducted on an Android phone and a laptop. The cloud server is simulated by utilizing a laptop with 8GB and RAM 1.80 GHz Intel Core i7. Both DO and DU are simulated via an Android phone with Kirin 810 processor and 6GB RAM. The cryptographic primitives and algorithms of both FPDS scheme in [5] and TIDS are implemented by utilizing the Stanford Pairing-Based Crypto (PBC) library1 in C++. The size of both G and G1 is initialized as 1024 bits. The Enron dataset2 is utilized to select keywords randomly. The number of keywords in FPDS is initialized to 5. Figure 5 shows the comparison results of the computational time costs of the key generation algorithm in TIDS and FPDS. When the number of attributes grows to 40, the TIDS requires 2.84 seconds for key generation, while the FPDS scheme in [5] requires 3.39 seconds. The comparison results show that the TIDS is more efficient than the FPDS scheme in [5] in terms of key generation. Furthermore, Fig. 5 also shows the time costs of the tracing algorithm although the FPDS scheme in [5] does not support the functionality of malicious user tracing in multi-user data sharing system. As shown in Fig. 5, the time cost of malicious user tracing

1 http://crypto.stanford.edu/pbc/. 2 http://www.cs.cmu.edu/~enron/.

316

J. Liang and S. Guo 4.0

Fig. 5 Running time of key generation and trace

Running Time (s)

3.5

Trace TIDS-KeyGen FPDS-KeyGen

3.0 2.5 2.0 1.5 1.0 0.5 0.0

10

20

30

40

Number of Attributes

8

Fig. 6 Encryption time in TIDS and FPDS

TIDS FPDS

Running Time (s)

7 6 5 4 3 2 1 0 0

5

10

15

20

25

30

35

40

45

50

Number of Attributes

keeps stable (around 0.25 seconds) with respect to the number of attributes. The comparison results show that the TIDS achieves efficient malicious user tracing functionality in multi-user data sharing scenario. Figure 6 depicts the comparison results of the computational time costs of the data encryption algorithm in TIDS and FPDS. Different from the FPDS scheme in [5], TIDS is optimized by requiring fewer exponentiations and additional hash operations for sharing sensory data encryption. As a result, the TIDS only requires around 2.5 seconds for data encryption when the number of attributes grows to 45, which is around 3 times faster than the FPDS scheme in [5]. Therefore, the TIDS is more efficient than the FPDS scheme in [5] in terms of time costs of data encryption. Figure 7 demonstrates the comparison results of the computational time costs of the data re-encryption algorithm in TIDS and FPDS. Compared with the FPDS scheme in [5], the FPDS scheme requires more exponentiations and additional hash operations for data re-encryption. As a result, the TIDS only requires 2.35 seconds for data re-encryption while the FPDS scheme in [5] requires 4.6 seconds for data re-

Traceable and Secure Data Sharing in Mobile Crowdsensing 5

Fig. 7 Re-encryption time in TIDS and FPDS

317

TIDS FPDS

Running Time (s)

4 3 2 1 0

0

5

10

15

20

25

30

35

40

45

50

40

45

50

Number of Attributes 1.4

Fig. 8 Decryption time in TIDS and FPDS Running Time (s)

1.2

TIDS FPDS

1.0 0.8 0.6 0.4 0.2 0.0

0

5

10

15

20

25

30

35

Number of Attributes

encryption when the number of attributes grows to 45. Therefore, the TIDS is more efficient than the FPDS scheme in [5] in terms of time costs of data re-encryption. Figure 8 displays the comparison results of the computational time costs of the data decryption algorithm in TIDS and FPDS. The time costs of data decryption in DU side keep stable with respect to the growth of the number of attributes. Meanwhile, the comparison results show that the decryption time cost of TIDS is more than 2 times less than that of FPDS. Therefore, the TIDS is more efficient than the FPDS scheme in [5] in terms of time costs of data decryption. The experimental evaluations show that TIDS achieves a better computational efficiency than the FPDS scheme in terms of time costs of key generation, encryption, re-encryption, and decryption. Furthermore, TIDS achieves efficient malicious user tracing, while the FPDS schemes did not consider user traceability.

318

J. Liang and S. Guo

4 Conclusion and Future Work In this chapter, we have introduced the traceable and secure data sharing issues in mobile crowdsensing. Then, we have reviewed several secure data sharing schemes in mobile crowdsensing. Afterward, we have illustrated a traceable and privacypreserving non-interactive data sharing (TIDS) scheme [25] as an example of traceable and secure data sharing solutions in mobile crowdsensing. Specifically, we have provided the fine-grained access control and secret key abuse issues in mobile crowdsensing. Then, we have described the detailed construction of TIDS. By leveraging traceable ciphertext policy attribute-based encryption and proxy re-encryption, TIDS ensures the confidentiality of sensitive sensory data while enabling flexible data sharing in mobile crowdsensing with multiple data users. Furthermore, TIDS enables malicious user tracing, which traces the real identity of malicious data user who shares secret keys to unauthorized data users. Finally, we have provided theoretical analysis including security analysis and performance analysis and experimental evaluations to show the advantages of TIDS in terms of functionality, security, and computational and storage efficiency. The emerging Web 3.0 technologies have shown the great potential of enabling peer-to-peer (or decentralized) data collection, trading, and transmission. In such a new environment, the centralized cloud server may be removed from mobile crowdsensing. Accordingly, the trustworthy issue will become more complex because all participants can hardly trust one another. As a result, developing trust, traceable, and secure data sharing techniques will become a new challenge to data sharing in mobile crowdsensing. Existing blockchain technologies ensure traceability and trustworthy in the decentralized environment. However, due to resource constraint in edge nodes, leveraging cryptographic tools to develop secure data sharing schemes for mobile crowdsensing is still an open issue. Therefore, for the future work, it is desirable to develop secure data sharing schemes for decentralized blockchainenabled mobile crowdsensing.

References 1. M. Abdalla, A.W. Dent, J. Malone-Lee, G. Neven, D.H. Phan, N.P. Smart, Identity-based traitor tracing, in Public Key Cryptography–PKC 2007: 10th International Conference on Practice and Theory in Public-Key Cryptography Beijing, April 16–20, 2007. Proceedings 10 (Springer, Berlin, 2007), pp. 361–376 2. A. Beimel, Secure schemes for secret sharing and key distribution. Israel Institute of Technology, Technion, Haifa, Israel (1996) 3. D. Boneh, M. Zhandry, Multiparty key exchange, efficient traitor tracing, and more from indistinguishability obfuscation. Algorithmica 79(4), 1233–1285 (2017) 4. B. Dan, M. Franklin, Identity-based encryption from the Weil pairing, in Advances in Cryptology—CRYPTO 2001: 21st Annual International Cryptology Conference, Santa Barbara, California, August 19–23, 2001 Proceedings (2001), pp. 213–229

Traceable and Secure Data Sharing in Mobile Crowdsensing

319

5. H. Deng, Z. Qin, L. Sha, H. Yin, A flexible privacy-preserving data sharing scheme in cloudassisted IoT. IEEE Int. Things J. 7(12), 11601–11611 (2020) 6. C. Dong, Y. Wang, A. Aldweesh, P. Mccorry, A.V. Moorsel, Betrayal, distrust, and rationality: Smart counter-collusion contracts for verifiable cloud computing, in Proceedings of the 2017 ACM SIGSAC Conference on Computer and Communications Security (2017), pp. 211–227 7. R.K. Ganti, F. Ye, H. Lei, Mobile crowdsensing: current state and future challenges. IEEE Commun. Mag. 49(11), 32–39 (2011) 8. C. Ge, W. Susilo, J. Wang, L. Fang, Identity-based conditional proxy re-encryption with fine grain policy. Comput. Standards Interf. 52(5), 1–9 (2017) 9. C. Ge, W. Susilo, Z. Liu, J. Xia, L. Fang, Secure keyword search and data sharing mechanism for cloud computing. IEEE Trans. Dependable Secure Comput. (2020). https://doi.org/10.1109/ TDSC.2020.2963978 10. V. Goyal, O. Pandey, A. Sahai, B. Waters, Attribute-based encryption for fine-grained access control of encrypted data, in Proceedings of the 13th ACM Conference on Computer and Communications Security (2006), pp. 89–98 11. M. Green, G. Ateniese, Identity-based proxy re-encryption, in Applied Cryptography and Network Security: 5th International Conference, ACNS 2007, Zhuhai, June 5–8, 2007. Proceedings 5 (2007), pp. 288–306 12. B. Guo, H. Chen, Z. Yu, X. Xie, S. Huangfu, D. Zhang, FlierMeet: a mobile crowdsensing system for cross-space public information reposting, tagging, and sharing. IEEE Trans. Mobile Comput. 14(10), 2020–2033 (2014) 13. B. Guo, Z. Yu, X. Zhou, D. Zhang, From participatory sensing to mobile crowd sensing, in Proceedings of the IEEE International Conference on Pervasive Computing and Communication Workshops (PERCOM WORKSHOPS) (IEEE, Piscataway, 2014), pp. 593–598 14. H. Guo, Z. Zhang, J. Xu, N. An, X. Lan, Accountable proxy re-encryption for secure data sharing. IEEE Trans. Dependable Secure Comput. 18(1), 145–159 (2018) 15. C. Huang, D. Liu, J. Ni, R. Lu, X. Shen, Achieving accountable and efficient data sharing in industrial internet of things. IEEE Trans. Ind. Informat. 17(2), 1416–1427 (2020) 16. K. Liang, M.H. Au, W. Susilo, D.S. Wong, G.Yang, Y. Yu, An adaptively CCA-secure ciphertext-policy attribute-based proxy re-encryption for cloud data sharing, in Information Security Practice and Experience: 10th International Conference, ISPEC 2014, Fuzhou, May 5–8, 2014. Proceedings 10 (2014), pp. 448–461 17. J. Liang, Z. Qin, J. Ni, X. Lin, X. Shen, Practical and secure SVM classification for cloud-based remote clinical decision services. IEEE Trans. Comput. 70(10), 1612–1625 (2020) 18. J. Liang, Z. Qin, S. Xiao, L. Ou, X. Lin, Efficient and secure decision tree classification for cloud-assisted online diagnosis services. IEEE Trans. Dependable Secure Comput. 18(4), 1632–1644 (2021) 19. H. Lim, G. Poh, J. Xu, V. Chittawar, PrivateLink: privacy-preserving integration and sharing of datasets. IEEE Trans. Inf. Forensics Secur. 15, 564–577 (2019) 20. J. Ni, K. Zhang, Q. Xia, X. Lin, X. Shen, Enabling strong privacy preservation and accurate task allocation for mobile crowdsensing. IEEE Trans. Mobile Comput. 19(6), 1317–1331 (2019) 21. J. Ning, Z. Cao, X. Dong, K. Liang, M. Hui, L. Wei, Auditable sigma-time outsourced attribute-based encryption for access control in cloud computing. IEEE Trans. Inf. Forensics Secur. 13(1), 94–105 (2018) 22. M.A. Rahman, M.S. Hossain, A location-based mobile crowdsensing framework supporting a massive ad hoc social network environment. IEEE Commun. Mag. 55(3), 76–85 (2017) 23. F. Song, Z. Qin, J. Zhang, D. Liu, J. Liang, X. Shen, Efficient and privacy-preserving outsourced image retrieval in public clouds, in GLOBECOM 2020-2020 IEEE Global Communications Conference (2020), pp. 1–6 24. F. Song, Z. Qin, J. Liang, X. Lin, An efficient and privacy-preserving multi-user multikeyword search scheme without key sharing, in ICC 2021-IEEE International Conference on Communications (2021), pp. 1–6

320

J. Liang and S. Guo

25. F. Song, Z. Qin, J. Liang, P. Xiong, X. Lin, Traceable and privacy-preserving non-interactive data sharing in mobile crowdsensing, in 2021 18th International Conference on Privacy, Security and Trust (PST) (2021), pp. 1–9. https://doi.org/10.1109/PST52912.2021.9647802 26. F. Song, Z. Qin, D. Liu, J. Zhang, X. Lin, X. Shen, Privacy-preserving task matching with threshold similarity search via vehicular crowdsourcing. IEEE Trans. Veh. Technol. 70(7), 7161–7175 (2021) 27. F. Song, Z. Qin, L. Xue, J. Zhang, X. Lin, X. Shen, Privacy-preserving keyword similarity search over encrypted spatial data in cloud computing. IEEE Int. Things J. (2021). Early Access https://doi.org/10.1109/JIOT.2021.3110300 28. J. Tang, S. Fu, X. Liu, Y. Luo, M. Xu, Achieving privacy-preserving and lightweight truth discovery in mobile crowdsensing. IEEE Trans. Knowl. Data Eng. (2021). https://doi.org/10. 1109/TKDE.2021.3054409 29. L. Wang, D. Zhang, Y. Wang, C. Chen, X. Han, A. M’hamed, Sparse mobile crowdsensing: challenges and opportunities. IEEE Commun. Mag. 54(7), 161–167 (2016) 30. X. Wang, J. Ma, X. Liu, Y. Miao, Search in my way: Practical outsourced image retrieval framework supporting unshared key, in IEEE INFOCOM 2019-IEEE Conference on Computer Communications (2019), pp. 2485–2493 31. C. Wiltz, What is crowdsensing? . . . and where can i park? (2018) https://www.designnews. com/electronics-test/what-crowdsensing-and-where-can-i-park 32. P. Xu, T. Jiao, Q. Wu, W. Wang, H. Jin, Conditional identity-based broadcast proxy reencryption and its application to cloud email. IEEE Trans. Comput. 65(1), 66–79 (2015) 33. G. Xu, H.Li, H. Ren, X. Lin, X. Shen, Dna DNA similarity search with access control over encrypted cloud data. IEEE Trans. Cloud Comput. (2020). Early Access https://doi.org/10. 1109/TCC.2020.2968893 34. L. Xue, J. Ni, C. Huang, X. Lin, X. Shen, Forward secure and fine-grained data sharing for mobile crowdsensing, in 2019 17th International Conference on Privacy, Security and Trust (PST) (2019), pp. 1–9. https://doi.org/10.1109/PST47121.2019.8949066 35. Y. Yang H. Lu, J. Weng, Y. Zhang, J. Weng, K. Sakurai, Fine-grained conditional proxy reencryption and application, in Proceedings of the Provable Security, pp. 206–222 (2014) 36. Y. Zhan, P. Li, K. Wang, S. Guo, Y. Xia, Big data analytics by crowdlearning: architecture and mechanism design. IEEE Netw. 34(3), 143–147 (2020) 37. Y. Zhang, J. Li, D. Zheng, X. Chen, H. Li, Accountable large-universe attribute-based encryption supporting any monotone access structures, in Proceedings of the ACISP (Springer, Berlin, 2016), pp. 509–524 38. L. Zhen, Z. Cao, D.S. Wong, Traceable cpCP-abeABE: How to trace decryption devices found in the wild. IEEE Trans. Inf. Forensics Secur. 10(1), 55–68 (2015) 39. Y. Zheng, R. Lu, Y. Guan, J. Shao, H. Zhu, Achieving efficient and privacy-preserving exact set similarity search over encrypted data. IEEE Trans. Dependable Secure Comput. (2020). Early Access https://doi.org/10.1109/TDSC.2020.3004442

User Privacy Protection in MCS: Threats, Solutions, and Open Issues Zhibo Wang, Xiaoyi Pang, Peng Sun, and Jiahui Hu

1 Introduction Fueled by the pervasiveness of sensor-rich mobile devices (e.g., smartphones, smartwatches), along with the unprecedented advances in mobile and communication technologies (e.g., 5G cellular networks, WiFi), mobile crowdsourcing (MCS) has become an increasingly popular and effective paradigm for distributed data collection and processing [1–3]. MCS is a combination of mobile crowd and outsourcing, which refers to the process of outsourcing tasks to an unspecified crowd of mobile users via an open call. Compared to traditional wireless sensor networks, MCS has significant advantages such as widespread spatio-temporal coverage, low (deployment and maintenance) cost, scalability, flexibility and so on. Most importantly, MCS can fully unleash the potential of crowd intelligence to solve complicated problems (e.g., image labeling, audio translation) that are usually difficult for a designated agent. Given these great benefits, MCS has been extensively applied in numerous applications, such as urban sensing, smart transportation and smart healthcare, to name a few [4–6]. However, due to user participation and data crowdsourcing, serious privacy concerns may arise for mobile users in MCS. For example, the sensed data collected from mobile users may contain their private or sensitive information (e.g., location,

Z. Wang () · J. Hu School of Cyber Science and Technology, Zhejiang University, Hangzhou, China e-mail: [email protected]; [email protected] X. Pang School of Cyber Science and Engineering, Wuhan University, Wuhan, China e-mail: [email protected] P. Sun College of Computer Science and Electronic Engineering, Hunan University, Changsha, China e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 J. Wu, E. Wang (eds.), Mobile Crowdsourcing, Wireless Networks , https://doi.org/10.1007/978-3-031-32397-3_13

321

322

Z. Wang et al.

identity). With the ever-increasing privacy sensitivity of mobile users and stringent privacy legislation like the General Data Protection Regulation (GDPR) launched by the European Union, the privacy concerns of mobile users need to be well addressed. Otherwise, they will be reluctant to participate in MCS campaigns or their sensed data will not be allowed to be collected. Noticing the importance of user privacy protection in MCS, researchers have devoted large research efforts toward investigating the privacy threats and proposing effective privacy protection methods. This book chapter provides a review and summary of works on user privacy protection in MCS. Specifically, we first identify the privacy threats in MCS and the corresponding requirements for privacy protection. Then, we summarize and discuss some of the existing privacy protection methods. Finally, we point out some promising open issues for future research. We expect this book chapter can inspire further investigations and innovations into privacy protection solutions for MCS.

2 User Privacy Threats and Requirements In MCS systems, mobile users are taken as sensors and are required to complete some sensing or computing tasks. When completing these tasks, users often face a lot of privacy threats since their uploaded sensing data or computing results may contain some sensitive information (e.g., environmental information surrounding users, users’ physical and social activities information). In the process of data flow, attackers have many opportunities to obtain data and infer user privacy. In this section, we discuss the privacy threats for users in mobile crowdsourcing. We first give the threat model and review some privacy attacks, then summarize the user privacy threats from the perspective of task and data, respectively.

2.1 Threat Model Figure 1 shows the typical architecture and data flow of a MCS system, which contains the following entities: requesters (end users), mobile users, MCS platform (server) and third-party service provider [7]. In a MCS system, data flows between these entities, and the untrustworthiness of any of them may cause the privacy leakage of users. Besides, the vulnerable wireless communication channel also introduces a great risk of user privacy leakage [8]. In the following, we describe the potential adversaries or attack points of adversaries and present the commonly used threat models of the MCS system. The potential adversaries or attack points in MCS systems are as follows. • Untrusted MCS platform: In the whole life cycle of MCS, mobile users interact with the MCS platform (server) frequently since they are requested to upload their sensing data or computation results for MCS tasks to the MCS platform.

User Privacy Protection in MCS: Threats, Solutions, and Open Issues

Requesters (End Users)

Services Providers & Analysts

Platform（Server） Requests

323

Data/Results Publishing

Feeds

Data Collection

Task Allocation

Incentive

Computing results uploading

Mobile Users

Fig. 1 The typical architecture and data flow of a MCS system

The MCS platform has access to users’ original data, thus can obtain a lot of information about users. Once the MCS platform is hacked and becomes untrusted, it may cause damage to user privacy. To be specific, an untrusted MCS platform will follow the protocol execution faithfully but is curious regarding user individual sensitive information. It may infer user privacy from the submitted data or even sell users’ data or private information in exchange for profits. Furthermore, if the untrusted server stores the user’s historical submitted data and tracks the changes over time, more information about the user can be inferred, for example, the trajectory of the user, causing a greater privacy threat to users. • Malicious mobile users: MCS systems usually recruit mobile users to participate in MCS tasks. It gives chances for malicious mobile users to join in the MCS tasks and obtain information about other mobile users. On the one hand, malicious mobile users may infer other users’ uploaded data according to their own data and the statistical results published by the MCS platform. On the other hand, some MCS tasks need to be performed by users with specific characteristics (e.g., the user at a particular location, the users with specific occupations), allowing malicious users be aware of sensitive attributes of other users. • Malicious requesters: A MCS service requester (end user) can send data or computing requests as well as incentives to mobile users through the MCS platform. For example, the requester may ask mobile users to report their checkin information so as to count the number of people in each region [9, 10]. With MCS, requesters can obtain all the information they want without leaving their office. Note that the required data is often spatial-temporal or contains some individual information about mobile users. Malicious requesters can also

324

Z. Wang et al.

be attackers since they can deliberately post tasks that require specific users to do and then identify users with specific attributes and infer their sensitive information. • Vulnerable communication channel: In MCS systems, MCS platforms allocate MCS tasks to mobile users via wireless communication channel, and mobile users upload their sensing data or computing results via wireless communication channels too. If the wireless communication channel is vulnerable, the adversaries can easily monitor or eavesdrop the transmitted data, leading to the disclosure of user privacy. Therefore, the vulnerable wireless communication channel is an important attack entry in MCS. There are some commonly used threat models in MCS systems. For example, the MCS platform is assumed untrusted and does not collude with users, and mobile users are trusted and do not collude with each other [10–16]; the adversary can be the MCS platform or a mobile user, and all entities do not collude with each other [17, 18]; the MCS platform, mobile users and requesters are all untrusted but they do not collude with each other [19].

2.2 Privacy Attacks To better understand the privacy threats for users in MCS systems, we introduce some potential privacy attacks in MCS. Inference Attack Inference attack is a data mining technology whose purpose is to illegally obtain knowledge about topics or databases [20]. An Inference attack occurs when an attacker is able to infer more robust information from trivial information about a database without directly accessing it [21]. In MCS systems, the MCS platform can infer a user’s sensitive information based on the collected sensing data. For example, if the MCS platform collects data from accelerometers in a user’s mobile devices, then based on the recorded motion patterns, the platform can infer the user’s sensitive information such as driving behavior, level of intoxication, age, gender, touchscreen inputs, geographic location and so on [22]. Besides, the attacker can infer whether or not a particular data point was used in calculations based on the computational results of the collected data, which is known as the membership inference attack and often happens in machine learning as a service [23, 24]. In MCS platform, the collected data is usually sensitive, then if the user’s data is involved in the calculation, the user’s sensitive information would be exposed. For example, the MCS platform can collect valuable cancer patient data. These data can be used to compute a predictive model to predict cancer-related health outcomes when given a patient’s data as input. The membership inference attack can infer whether an individual’s data is in the training data of the predictive model based on the model’s output and try to infer if the individual is a cancer patient.

User Privacy Protection in MCS: Threats, Solutions, and Open Issues

325

Linkage Attack The goal of the linkage attack is to piece together known information to explore some unknown information. Specifically, the adversary can collect auxiliary information about a certain user from multiple data sources and then combine them together to generate user profile and then infer sensitive individual information. There is a famous example of the linkage attack. Netflix published anonymous data about movie rankings for 500,000 customers in 2007, and some researchers de-anonymized the data by linking these data to auxiliary data from IMDb. Eavesdropping (Man-in-the-Middle) Attack Eavesdropping attack attempts to steal user data transmitted over the network. It typically occurs when a device connects to a network where the traffic is not secured or encrypted to send data to other devices. In MCS systems with vulnerable wireless communication channels, the attacker can access user submitted data through an eavesdropping attack, which may contain sensitive information. These confidential information can be intercepted, stolen, and sold. Thus the user privacy and data confidentiality would be breached.

2.3 Privacy Threats In MCS systems, users are required to complete specific sensing or computing tasks. The task-specific user data flows among entities in the insecure MCS environment, which makes the user face a lot of privacy threats, from both task and data.

2.3.1

Privacy Threats from Task

In MCS systems, requesters post their demands in MCS platforms to find appropriate mobile users to do specific things for them. For example, a requester may want someone to go to a store and check the price of an item for him, or let someone tell traffic conditions in a certain place at a certain time. He may also try to find some doctors to answer his question. MCS tasks are the carriers of requesters’ demands, which have certain attribute requirements for the task participants, leading to privacy threats for users participating in tasks. The corresponding MCS tasks in the above examples have strict requirements for the user’s location, time and place, and occupation, respectively. Thus, when users complete these tasks, their sensitive information such as locations and occupations will be leaked. What’s more, the content of the MCS task can also cause privacy threats for the requester (end user). For instance, a task that a requester wants some doctors to answer his question implies that he may suffer from diseases.

326

2.3.2

Z. Wang et al.

Privacy Threats from Data

In the MCS system, users complete tasks and upload task-specific user data to the MCS platform. The MCS platform stores the data and then feeds it back to the requester and publishes it to third-party service providers for deep utilization. During the process, there are lots of privacy breaches for users. Since users’ data that contains user sensitive information can be accessed by multiple parties (i.e., the MCS platform, the requester and the third-party service provider), the user privacy can be easily leaked to all these parties. On the one hand, the submitted data itself may be closely related to the user privacy, such as heart rates and fingerprints. On the other hand, the submitted data may be spatial-temporal or contains environmental and social information surrounding the user, making it possible to infer extra sensitive information about the user, such as location, hobby, and occupation.

2.4 Privacy Leakage in the Whole Data Flow Process As we discussed above, user privacy might be leaked in every phase of the whole data flow process (including task allocation, user incentive, data collection and data publishing) in MCS. Since users’ data flows between all entities in the MCS system through the wireless network, the untrustworthiness of any one of these entities and the vulnerabilities of wireless communication channel may cause privacy leakage of users. In the following, we discuss the privacy leakage in every phase of the data flow process when the MCS platform is untrusted.

2.4.1

Privacy Leakage in Task Allocation

In task allocation, the platform usually allocates tasks to mobile users based on their locations with the objective of maximizing the social welfare of the platform. Tasks are usually allocated to the closest mobile users as the smaller distance means less payment to the user. Therefore, each user needs to upload its location as well as its identity to the platform before task allocation. The platform may further sell these data to third parties to make profits. In this case, although appropriate task allocation can be realized, the user identity and location information are exposed.

2.4.2

Privacy Leakage in User Incentive

With the purpose of stimulating users’ participation, an incentive is closely associated with task allocation that helps the platform to decide how to allocate tasks to appropriate mobile users. The reverse auction is commonly used in incentive mechanisms where mobile users can submit bids to the platform to compete for the tasks they are interested in. The platform will determine the winner for each

User Privacy Protection in MCS: Threats, Solutions, and Open Issues

327

task and the payment to each winner by taking the submitted bids and some other factors (e.g., reputation) into consideration. The value of the bid to a task of a user is mainly determined by the distance between the user and its interested task, and with multiple bids submitted by a user, it is possible to infer the true location of the user. When the platform is untrusted, the location information might be disclosed to the public. The privacy threat of identity also exists in the incentive process.

2.4.3

Privacy Leakage in Data Collection

In data collection, mobile users usually need to move to some specified location at some specified time to accomplish the spatio-temporal crowdsourcing tasks, and then upload the task-specific data to the platform. In this process, the identities and locations of users, sensing/computing data of users will all be obtained by the platform. Besides, the submitted data and the aggregated statistics over the submitted data can be further mined to obtain more private information of users, e.g., the individual location information can be inferred from the aggregated statistics over the check-in data.

2.4.4

Privacy Leakage in Data Processing and Publishing

After receiving users’ uploaded data, the MCS platform may feed these aggregated data back to requesters directly, or process the data first (e.g., truth discovery) to meet the needs of requesters and then feed back the processing results to requesters. Also, the aggregated data can be further published to third parties for data mining purposes. Thus, users’ privacy could be compromised to those parties that have access to the aggregated data in the data processing and publishing phase. Those parties can infer individual information of users based on the aggregated data and its statistics, such as individual location, occupation, and so on. In general, there are mainly three privacy threats in the whole data flow process in MCS, identity privacy, location privacy, and data privacy. The privacy threats in MCS are summarized in Table 1. The identity and location privacy threats are involved in all phases, and the data privacy threat is involved in the data collection and data publishing phases. Thus, it is necessary to integrate privacy protection into the whole data flow process of MCS to meet user privacy requirements.

Table 1 Privacy threats in the whole data flow process of MCS systems Data flow phase Task allocation Incentive Data collection Data publishing

Identity privacy    

Location privacy    

Data privacy

 

328

Z. Wang et al.

2.5 Requirements for User Privacy It is worth noting that while enjoying the convenience brought by MCS, people become more and more concerned about their privacy and have stronger and stronger privacy protection requirements. In recent years, many countries have promulgated relevant laws and regulations to strictly regulate the protection of personal data and privacy, e.g., the EU General Data Protection Regulation (GDPR), the Cybersecurity Law of the People’s Republic of China and California Consumer Privacy Act (CCPA). There is a strong preference that personal information should be protected when the user participates in MCS, especially the user identity, location, bid information and data. Figure 2 shows the user privacy protection requirements in the whole life cycle of a crowdsourcing task from task allocation to data publishing. In each phase, users have some privacy requirements. User Identity Privacy In the age of the Internet, a user’s identity information is associated with a lot of information. Based on the identity of a person, much sensitive information about the person can be found. In MCS systems, user identity can be easily exposed to the untrusted MCS platforms and be further exposed to other potential adversaries. Thus, the user has protection requirement for identity privacy in MCS systems. One countermeasure is to hide the users’ identities when taking the task and reporting results, which can be achieved by anonymous communication technology, dynamic pseudonym generation technology and so on. User Location Privacy Users’ location information is closely related to their personal security. With the location information of a user, the adversary can carry out physical assaults on the user or infer other sensitive information of the user. For example, if the user located at a hospital, the condition of his health can be inferred. If the attacker could receive continuous updates of user location in real time, he can obtain the moving trajectory of the user and infer more sensitive

Fig. 2 User privacy requirements in the whole data flow process of MCS systems

User Privacy Protection in MCS: Threats, Solutions, and Open Issues

329

information such as the user’s lifestyle, personal habits, professional duties and so on. In MCS systems, the untrusted MCS platforms can continuously acquire the user location update and may expose the user location and trajectory information to other potential adversaries. Hence, for privacy, user location should be protected against the MCS platforms in MCS systems. Anonymity and obfuscation can be solutions for location privacy protection. The former tries to anonymize the location and mobility traces, and the latter hides the real location information by submitting the disturbed location. User Bid Privacy Bid plays an important role in user incentive in MCS systems. According to the bids submitted by mobile users for tasks they are interested in, the MCS platform selects some users as the winners to perform tasks. Generally, the bid value stands for the true cost to perform a specific task, which is highly related to the user device model, the distance between the user and the task, etc. Thus, the adversaries with user bid information can infer user individual information (e.g., location, device usage preference). There is also a requirement that user bid privacy should be protected in MCS systems. Cryptography techniques and perturbation technologies can both provide effective solutions. User Data Privacy In the MCS system, users join in MCS tasks and submit taskspecific data to the MCS platform. For sensing tasks, the users collect corresponding sensing data. For computation tasks, the users use local resources and data to complete the tasks and upload the calculation results. Both the sensing data and calculation results may contain sensitive information of users, such as identities, location information, biometric information, device model and so on. Therefore, there is a strong requirement to guarantee user data privacy in MCS systems to avoid the disclosure of user personal information. The most popular method is to encrypt or sanitize the user’s task-specific data before submitting while limiting the impact on data quality. Personalized Privacy Protection It is worth noting that in practice, some users are more concerned about privacy leaks than others. That is, different users may demonstrate different sensitivity levels to the same degree of privacy leakage and have different privacy preferences [25]. In MCS systems, the “one size fits all” user privacy-preserving approaches are not that applicable to real-world scenarios. For instance, perturbation-based data preserving methods add the same amount of noise to all users’ data, which may lead to the situation that some users are overprotected while others are insufficiently protected [16]. It causes high data quality loss. Hence, when providing privacy guarantees for users in MCS systems, it is necessary to take each user’s personalized privacy requirement into consideration to achieve a better trade-off between user privacy and data quality.

330

Z. Wang et al.

3 Privacy Protection Technologies In this section, we introduce some mainstream privacy protection technologies applied in MCS systems. Specifically, we first introduce anonymization-based techniques in Sect. 3.1, then review and discuss perturbation-based methods in Sect. 3.2, and finally focus on encryption-based schemes in Sect. 3.3.

3.1 Anonymization-Based Technologies Anonymization-based technologies have been extensively studied and widely used for user privacy protection in MCS systems [1–3]. Initial research works on anonymization protected users’ private and sensitive information by masking or eliminating identifiers (e.g., names, addresses), as well as quasi-identifiers (e.g., gender, postcode) [26]. Afterward, the notion of k-anonymity [27] was introduced, which was built on the idea that by combining sets of data with similar attributes, it is rather difficult for adversaries to identify information about anyone who contributes to that data. Thus, people usually refer to k-anonymity as the power of “blending into the crowd”. Specifically, users’ data is generalized into a larger group, and information regarding the group may correspond to anyone of the involved users. Therefore, with k-anonymity, the identifying data (both identifiers and quasi-identifiers) and the sensitive data (e.g., medical records, passwords, social security number, etc.) of a specific user cannot be connected to one another, and thus privacy is protected. Besides, researchers have further developed other refinements of k-anonymity, such as .(α, k)-anonymity [28], l-diversity [29], t-closeness [30], and m-invariance [31]. These definitions introduce extra restrictions on the released data values. Generally, there are two different ways to implement k-anonymity, i.e., generalization and suppression, which are introduced in the sequel.

3.1.1

Generalization

The first approach to provide k-anonymity is generalization, which is built up on the definition and utilization of generalization relationships among domains and values that attributes can assume. The core idea of generalization is to eliminate identifying information that can be inferred from data by reducing an attribute’s specificity or enlarging the range intervals. Generalization refers to substituting some particular values into a more general one. For example, users aged 44, 50, 52, and 60 can be all flagged as older than 40.

User Privacy Protection in MCS: Threats, Solutions, and Open Issues

3.1.2

331

Suppression

Suppression is a complementary approach to provide k-anonymity. It refers to the process of removing some information from a data set so that it is not released. For example, if we want to make a data set composed of postcodes less informative, we may delete the last two digits of the postcodes and reduce their accuracy (e.g., changing 37966 to 379**). Nevertheless, it is worth noting that only when the data points are irrelevant to the goal of data collection can suppression be used. However, many existing works have demonstrated that given some background knowledge, adversaries can still re-identify individual users from anonymized data, thus compromising privacy [32, 33]. This was verified by the recent Netflix data breach event, where individual users were identified from the released anonymized dataset of movie ratings. Thus far, how to achieve secure release of anonymized data for further analysis is still an open yet challenging problem.

3.2 Perturbation-Based Technologies Unlike anonymization-based technologies which mix multiple users’ data together, perturbation-based methods offer privacy protection by modifying the original data from individual users. Perturbation-based privacy protection schemes can be generally grouped into the following two categories [34].

3.2.1

Randomized Response

Randomized response, developed in the 1960s, is a data collection strategy designed for surveys collecting statistics on sensitive topics [35–37]. It can help survey respondents to retain confidentiality and privacy by establishing a probabilistic connection between survey questions and the corresponding answers. For example, when a user in the MCS system is asked to provide an answer to a question on a sensitive topic, he or she can flip a coin before answering it in a secret manner. If the coin comes up heads, he or she simply answers “Yes”; otherwise (if the coin comes up tails), he or she chooses to tell the truth. Using this procedure, the user obtains plausible deniability for any “Yes” answers he or she provides, thus the user can hold their own private data and release them to curators in a differentially private manner, and the individual data is protected to a certain level. In general, this randomized response technique, when applied to one-time data collection, can well protect users’ privacy, irrespective of any adversaries’ computing capabilities or prior background knowledge. Specifically, randomized response mechanisms that satisfy local differential privacy have been widely used to protect individual privacy in MCS systems [36, 38–41]. For instance, Google Chrome uses RAPPOR [36] to constantly collect users’ responses to sensitive questions such as the default

332

Z. Wang et al.

homepage of the browser and the default search engine in a randomized responsebased differentially private manner. However, it is worth noting that the privacy guarantee offered by randomized response will be weakened if the survey (i.e., data collection process) is repeated many times for the same user. Randomized response has been widely applied in MCS systems, especially for those tasks with binary or categorical data.

3.2.2

Differential Privacy

Differential privacy (DP) is a rigorous privacy model. A randomized algorithm .A is differentially private if and only if the inclusion of a single instance in the input dataset causes statistically indistinguishable changes to the output of .A, regardless of the adversary’s computing capability and background knowledge. Formally, DP is defined as follows: Definition 1 (.(, δ)-DP [42]) A randomized algorithm .A : D → R with domain D and range .R is considered to satisfy .(, δ)-DP if for any two adjacent datasets  .D, D ∈ D that differ on at most one data sample and any subsets .S ⊆ R, it satisfies that .

.

    Pr[A(D) ∈ S] ≤ e Pr A D  ∈ S + δ.

(1)

where . is the privacy budget which controls the privacy level, and smaller values of  imply better privacy protection. The parameter .δ accounts for the probability that plain .-DP is broken.

.

The Gaussian mechanism (denoted by .A) [43] is widely adopted to achieve (, δ)-DP. Specifically, given any function f that maps a dataset .D ∈ D to a vector d .o ∈ R , .(, δ)-DP is realized through injecting Gaussian noises into each of the d coordinates of the output vector .o. Formally, we have .

A(D) = f (D) + N(0, σ 2 Id ),

.

(2)

where .N denotes the Gaussian distribution, .Id represents the d-dimensional identity matrix, and the noise magnitude  .σ is proportional   to the sensitivity of f , which is defined as .f = maxD,D  f (D) − f D  2 (.D, D  ∈ D are two adjacent datasets that differ on one data sample). DP is a lightweight technology for privacy protection and has become widespread in MCS systems. For example, [11, 44, 45] use DP to protect user privacy in task allocation in MCS, and [46, 47] provide .-differential privacy for “one-time” release of statistical data. In order to further provide strong privacy guarantee for real-time data, a new privacy model called “w-event DP” [48] has been proposed, which can provide provable privacy guarantee for any event sequence occurring at any successive w time stamps. A stream prefix of an infinite time-series .S =

User Privacy Protection in MCS: Threats, Solutions, and Open Issues

333

(D1 , D2 , . . .) at time stamp t is defined as .St = (D1 , D2 , . . . , Dt ). Then the “wevent DP” can be defined as: Definition 2 (w-Event DP [48]) A mechanism .A satisfies w-event .-differential  privacy, if for all sets .S ⊆ Range(A) and all w-neighboring stream prefixes .St , .St and all t, it holds that, 

Pr[A(St ) ∈ S] ≤ e · Pr[A(St ) ∈ S]

(3)

A mechanism satisfying w-event DP will protect the sensitive information that may be disclosed from any event sequence occurring at any successive w time stamps. Therefore, researchers adopt w-event DP to protect real-time and timeseries data privacy in MCS systems [9, 10]. Although perturbation-based privacy protection methods can provide a rigorous privacy guarantee for users, it remains to be further investigated how to achieve a desirable balance between privacy and data usability since introducing perturbation into data inevitably causes information loss.

3.3 Encryption-Based Technologies The basic idea of encryption is to secure the data reporting process via cryptographic methods [1, 49]. Different from anonymization-based and perturbationbased schemes, encryption does not need to modify the actual data. Instead, it is more like enforcing a lock on the data, thus incurring no information loss. Among various encryption-based privacy protection technologies, homomorphic encryption [50], as a special form of encryption, is becoming increasingly popular. Homomorphic encryption refers to the process of conversing data into ciphertext that can be computed and analyzed as if it was in the form of plaintext. Given this property, homomorphic encryption enables users to perform sophisticated mathematical operations on encrypted data without the need for decrypting it, and the computed encrypted result can exactly match that obtained through operations performed on the corresponding plaintext. We introduce two kinds of homomorphic encryption (i.e., fully homomorphic encryption and partially homomorphic encryption) as follows.

3.3.1

Fully Homomorphic Encryption

Fully homomorphic encryption, which was first introduced in [51], is a special homomorphic encryption scheme that allows arbitrary analytical functions to be applied directly to encrypted data without decrypting it while acquiring the same encrypted results compared to the case when those functions are applied to the plaintext. The benefits of fully homomorphic encryption are two-fold:

334

Z. Wang et al.

• End-to-end security. Data with fully homomorphic encryption is protected and private end-to-end and is only decrypted when acquiring the final computation result. In this way, the probability that sensitive information is revealed is minimized. Also, there is no need for any trusted third parties. Instead, data remains secure and private in untrusted application scenarios. • No compromise on data usability. Unlike anonymization-based schemes, fully homomorphic encryption involves no mask or removal of any features of the data. Besides, compared to perturbation-based methods which introduce random noises into the raw data, fully homomorphic encryption promises the same analysis results after decrypting the encrypted result with no utility loss in terms of data accuracy. These great benefits make fully homomorphic encryption a promising privacy protection technology in MCS systems. However, fully homomorphic encryption usually incurs high computation overhead and needs large storage space, which limits its use in computationally heavy applications [52]. To address this issue, researchers have proposed employing fully homomorphic encryption in conjunction with other privacy computing techniques such as secure multi-party computation (SMC) [53].

3.3.2

Partially Homomorphic Encryption

Unlike fully homomorphic encryption, partially homomorphic encryption allows only select mathematical functions (i.e., either addition or multiplication, but not both) to be performed an unlimited number of times on the ciphertext [54]. For instance, we can consider an algorithm as additively homomorphic encryption if we can obtain the same result when adding two ciphertexts together just as encrypting the summation of the corresponding two plaintexts. Several partially homomorphic cryptosystems have been proposed in the literature. Among them, [55] is one of the most famous additively homomorphic encryption schemes, while [56] and [57] are typical examples of multiplicative homomorphic cryptosystems. Note that these partially homomorphic encryption methods can be well applied in MCS systems that need to collect users’ sensitive data and perform some simple operations (e.g., summation, averaging) in a secure and privacy-preserving manner. Partially homomorphic encryption also bears significant benefits in terms of endto-end security and no compromise on data usability. Furthermore, in comparison with fully homomorphic encryption which requires a lattice-based cryptosystem that is significantly more complex, partially homomorphic encryption is much more practical.

User Privacy Protection in MCS: Threats, Solutions, and Open Issues

335

4 Privacy Protection for Mobile Users In this section, we introduce some commonly used methods for user privacy protection in every phase of the whole data flow process of MCS (including task allocation, user incentive, data collection and data publishing). Among them, privacy-preserving task allocation mechanisms often try to protect user identity and location privacy. User bid privacy is the focus of privacy-preserving incentive mechanisms. Privacy-preserving data collection and publishing mechanisms emphasis on the protection of user data privacy.

4.1 User Privacy Protection in Task Allocation During the phase of task allocation, each user needs to upload its location and identity to the platform, which cannot be trusted and may be benefited from selling the information to third parties. Hence, protecting identities and locations is essential to users in task allocation. Protecting User Location Privacy Task allocation relies on the distances between tasks and mobile users to assign tasks appropriately and requires users to upload their location information. However, the location information may be disclosed during the task allocation process. For the location privacy, spatial cloaking is the most straightforward technique that blurs a user’s exact location into a spatial region in order to preserve the location privacy [58, 59]. In particular, k-anonymity is a popular method for spatial cloaking, where it is infeasible to distinguish one from other .k − 1 nodes. Therefore, some k-anonymity based location privacypreserving methods [60, 61] were proposed to protect users’ locations in task allocation. However, the privacy guarantees provided by spatial cloaking can be easily downgraded if adversaries hold certain prior knowledge, which is known as the inference attack. For example, the real location l of mobile user A is mapped to a region R, where contains only one school located at .ls . If A is known to be a student and always performs tasks in region R, the attacker can infer that the real location l is .ls with a high probability. To avoid the above issue, some authors [11, 44, 45] introduced differential privacy (DP) into task allocation in MCS, providing theoretically guaranteed location privacy protection regardless of adversaries’ prior knowledge. In these privacy-preserving task allocation mechanisms, mobile users obfuscate their real locations based on DP methods and then upload the obfuscated locations to the platform. The platform implements task allocation based on received obfuscated locations and attempts to minimize users’ travel distance. Specifically, [44] assumed that there is a trusted third party (cellular service providers) between the platform and mobile users to play the role of location obfuscation, while in practice the cellular service providers have no motivation to participate in user location privacy protection. In [45], the mobile users obfuscate

336

Z. Wang et al. Server

1. Task Request

Data Requester

2. Task Publish

Obfuscated Distance Laplace Mechanism

Actual Distance

Fig. 3 The framework of personalized privacy-preserving task allocation in mobile crowdsensing with obfuscated distance

locations according to a predefined probability function P (encoding the probability of mapping arbitrary location l to .l ∗ ). The platform firstly generates the geoobfuscation function P according to the task locations so that the negative effects of location obfuscation on user travel distance can be reduced. Then, mobile users can download the function P to obfuscate their locations and upload obfuscated locations to the platform. Finally, the platform assign tasks to proper users based on their obfuscated locations and the function P , attempting to minimize the travel distance. However, the proposed geo-obfuscation function is related to the task locations and needs to be updated constantly based on the actual situation, which may discourage users from joining in tasks since they are required to repeatedly download the function before obfuscating their actual locations. Most importantly, both of them employ the same level of privacy protection for all users, which cannot satisfy the different privacy demands of workers. Hence, a personalized privacy-preserving task allocation framework was proposed in [11], where each user locally obfuscates the distances between itself to tasks based on its personal privacy protection level, and uploads the obfuscated distances as well as its personal privacy protection level to the platform. Figure 3 shows the personalized privacy-preserving task allocation framework in mobile crowdsensing systems with distance obfuscation. The data requesters create spatial-temporal tasks on the server which then publishes tasks to workers. Workers can obtain the accurate locations of tasks so that they can determine whether to apply for tasks. Instead of applying for the interested tasks using the actual locations or actual distances, the workers can use Laplace mechanism to obfuscate the distance with their personalized privacy budget .. Note that . can be different for different workers which depends on the privacy protection requirement of each worker. The

User Privacy Protection in MCS: Threats, Solutions, and Open Issues

337

server selects the winner for each task (winner selection process) and gives the winners payments once their uploaded data are certified (payment determination process). However, without knowing the true distances, it is difficult to find the optimal task allocation to minimize the total travel distance. To solve this problem, the author proposed a suboptimal solution that finds the winner of a task as the worker who has the largest probability of being closest to the task. Let start from the special case that only two workers .wi and .wj applying for the task. Worker .wi uploads .i and .d˜ik , and worker .wj uploads .j and .d˜j k . The server aims to obtain the probability of .dik smaller than .dj k , denoted by .P (dik ≤ dj k ). The worker .wi gets the sanitized distance .d˜ik by adding the Laplace noise on .dik , hence the real distance of .wi is dik = d˜ik − ηi ,

.

ηi ∼ Laplace(0, 1/i )

(4)

ηj ∼ Laplace(0, 1/j )

(5)

Similarly the real distance of .wj is .

dj k = d˜j k − ηj ,

where .ηi and .ηj are variables that follow the Laplace distribution. Note that the smaller of .i , the larger of the added noise, and the stronger the privacy protection level. Then we have P (dik ≤ dj k ) = P (d˜ik − ηi ≤ d˜j k − ηj ) .

= P (d˜ik − d˜j k ≤ ηi − ηj )

(6)

where .d˜ik − d˜j k is known by the server. The above equation can be seen as a probability problem about two-dimensional continuous variables .(ηi , ηj ) in the plane set D, denoted by .P ((ηi − ηj ) ∈ D). The plane D can be denoted by D = {(ηi , ηj ) : ηi − ηj ≥ d˜ik − dj˜ k }

.

(7)

The double integral operation can be used for solving this problem, we have P (d˜ik − d˜j k ≤ ηi − ηj ) = P ((ηi − ηj ) ∈ D)  . f (ηi , ηj )dηi dηj =

(8)

D

where .f (ηi , ηj ) denotes the joint probability density function of .(ηi , ηj ). Since the variables .ηi , .ηj are independent from each other, we have

338

Z. Wang et al.



 f (ηi , ηj )dηi dηj =

f (ηi )f (ηj )dηi dηj

D

D



.

=

∞



−∞



ηi −(d˜ik −d˜j k )

−∞

(9)

f (ηi )f (ηj )dηj dηi

Note that .

f (ηi )f (ηj ) =

i j −(i |ηi |+j |ηj |) e 4

(10)

Equation (9) can be seen as a function, where the inputs of the function are (d˜ik , d˜j k , i , j ), and the output is the probability of .dik smaller than .dj k . We call this function as the Probability Compare Function (PCF). Combining Equations (6) and (9), we have

.

P (dik ≤ dj k ) = P CF (d˜ik , dj˜ k , i , j )  . = f (ηi , ηj )dηi dηj

(11)

D

If .P (dik ≤ dj k ) > 1/2, we can say that worker .wi will be closer than worker .wj with a larger probability. Based on this principle, we can compare any two workers who apply the task and finally find the worker who has the largest probability of being closest to the task. Protecting User Identity Privacy In MCS systems, the platform needs to distinguish every user from others so that tasks can be allocated to specific users. To achieve the objective, the platform requires each mobile user to have an identity and upload it for task allocation, which refers to any subset of his attributes that can be uniquely detected. Since the identity can be used to trace users’ activities, the protection of identity privacy is important in task allocation. The pseudonym is a popular scheme to prevent attacks targeted on identity privacy. However, attackers can easily link the activities of a user with its pseudonym. This issue can be solved by changing pseudonyms frequently, but it may introduce excessive communication overhead. Therefore, the pseudonym mechanisms are not sufficient to protect identity privacy, and they cannot ensure the communication efficiency and good unlinkability at the same time. To solve this problem, the l-anonymity is introduced to hide the real identity of each mobile user in a group of users. In particular, the author [62] proposed a privacy-preserving scheme AnonySense for MCS, aiming to protect both identity privacy and location privacy. The scheme adopted the spatial-temporal cloaking technique to guarantee location privacy and l-anonymity report aggregation technique to prevent the platform from associating the report with the identity of the user. Another effective and popular identity privacy protection tool is group signature [63], which enables signers to sign messages on behalf of a group. As a result,

User Privacy Protection in MCS: Threats, Solutions, and Open Issues

339

the attacker cannot determine the true signer of the message so that anonymity and unlinkability of mobile users can be ensured.

4.2 User Privacy Protection in Incentive Incentive design is one of the most important problems in MCS, which can stimulate users to participate in MCS tasks. For completing MCS tasks, a user may need to move to specific locations at specific times to collect sensing data, which costs not only the user’s time but also physical resources. Without an appropriate incentive, users may be unwilling to participate in MCS tasks and submit high-quality sensing data. Reverse auction-based incentive mechanisms have been commonly used in MCS, where users submit bids to the platform to compete for tasks. According to the bids submitted by mobile users, the platform selects some users as the winners to perform tasks. Generally, the bid value is associated with the information about the mobile user’s possible cost for task execution, its work plan or its ability, which obviously belongs to users’ personal privacy. Considering that, protecting user bid privacy is important and many researchers have focused on this topic. Protecting User Bid Privacy Typically, the bid value stands for the true cost to perform a specific task regardless of the users’ strategy behavior, which may reveal sensitive information of users [64], such as the location information or the model of mobile devices. Some mechanisms employed cryptography techniques [65, 66] to protect bid privacy in auctions, but these encryption-based methods usually suffer from high computation costs and are vulnerable to inference attacks where the true bid value can be revealed from the winner sets or payments published by the platform. Based on Trusted Third Party (TTP), Wang et al. [67] presented a reverse auction-based incentive mechanism and assumed one node is added to the winner set only when his/her marginal efficiency is the highest one. Since the real bids of users did not upload to the platform, the bid privacy is protected. However, it is vulnerable to a single point of failure due to the over-reliance on the TTP and also suffers from high communication overhead. To overcome these shortcomings, DP was introduced to bid-preserving mechanisms [68, 69] in recent years given its lightweight designs and theoretical privacy protection guarantees. The users’ real bid values can be obfuscated with DP before being submitted to the platform so that to protect bid privacy. However, all these bid-preserving mechanisms based on DP rely on a trusted platform where users still submit their true bids to it. Once the platform is curious or even attacked, the true bids might be disclosed and all the existing bid-preserving mechanisms would fail in bid protection. To solve this issue, as shown in Fig. 4, Wang et al. [70] design a privacypreserving incentive mechanism that protects the true bid against the untrusted platform while minimizing the social cost of winner selection. In particular, this

340

Z. Wang et al.

Fig. 4 The workflow of the privacy-preserving incentive mechanism under the untrusted platform

mechanism proposed a differentially private bid obfuscation function that allows each user to obfuscate its bids locally without disclosing the true bids to the untrusted platform. Then, the winner selection problem with the obfuscated taskbid pairs is formulated as an integer linear programming problem and optimized to minimize the total cost. To protect bid privacy against the untrusted platform, each user obfuscates its true bid b to an obfuscated one .b∗ locally and submit the obfuscated task-bid pairs to the platform. An intuitive way is that each user obfuscates bids arbitrarily at his wish, which, however, may violate the properties of truthful incentive mechanism (e.g., individual rationality and truthfulness). Therefore, a global bid obfuscation function is needed to help each user make suitable obfuscation while .-differential privacy is satisfied. The designed differentially private obfuscation function based on exponential mechanism to ensure bid privacy protection by obfuscating the true bid b to .b∗ , where .b, b∗ ∈ B. In order to form the bid set B without revealing users’ private information, as shown in step (2) and step (3) of Fig. 4, a level of multiple agents are deployed between the users and the untrusted platform. The exponential mechanism for the bid obfuscation function maps the true bid ∗ .b ∈ B to obfuscated bid .b ∈ B while providing .-differential privacy, which can be defined as follows. 

q(b, b∗ ) ∗ ∀b, b∗ ∈ B (12) .P (b |b) ∝ exp  · 2q

User Privacy Protection in MCS: Threats, Solutions, and Open Issues

341

where the outcome of .P (b∗ |b) is the probability of mapping b to .b∗ , and .q(b, b∗ ) is the score function measuring the closeness of .b∗ to the original b. Generally speaking, the closer between .b∗ and b, the higher of .q(b, b∗ ). The score function .q(b, b∗ ) is key factor to map the input b and the outcome .b∗ ∈ B to a real-valued score, which measures the degree of closeness of .b∗ to the true bid b. The principle is that the closer for .b∗ to b, the larger the probability .P (b∗ |b). Therefore, a monotonically decreasing function can be adopted as the score function with respect to .|b − b∗ |. Following this principle, a power score function .f (x) = −x 1/2 is designed. Hence, the score function can be q(b, b∗ ) = −|b − b∗ |1/2

.

(13)

Thus, the global sensitivity is .|bmax − bmin |1/2 = b1/2 where .bmin and .bmax denote the minimum and the maximum element of B, and .b = bmax − bmin . Let . b denote an arbitrary element in B, the obfuscation function based on the power score function will be

 |b − b∗ |1/2 P (b |b) ∝ exp  · − 2 · b1/2 

∗ |1/2 . exp  · − |b−b 1/2 2·b 

= |b− b|1/2 exp  · − 1/2 b∈B ∗

(14)

2·b

Once receiving the obfuscated task-bid pairs from the users, the platform begins to select appropriate users as winners. The objective in Eq. (15) is to select winners for all tasks while minimizing the total social cost. Winner Selection Problem The objective is to select a set of winners from users so that all the tasks can be allocated while minimizing the social cost. That is, min



j

ci

ui ∈Stj

s.t.

Stp



Stq = ∅ ∀tp , tq ∈ T , and p = q

|Stj | = λj

(15)

∀tj ∈ T

where the objective is to minimize the total social cost of winners which is the true total cost for users completing the selected tasks. The first constraint indicates that each user can only be selected to perform one task, and the second constraint ensures that every task can be completely allocated.

342

Z. Wang et al.

4.3 User Privacy Protection in Data Collection and Publishing As mentioned above, identity and location privacy threats are involved in all phases, and the data privacy threat mainly exists in the data collection and publishing phase. Since we have discussed the protection methods for location privacy and identity privacy in task allocation, data privacy is the main issue to be focused in data collection and publishing. The problem of data privacy leakage originates from the reality that data collected in MCS contain abundant sensitive information, such as health, identity, job, belief, etc. Consequently, it is necessary to provide data privacy preservation methods in MCS for the security and privacy consideration of users. Protecting User Data Privacy in Data Collection Compared to the single data, many MCS systems consider the collection of time-series data, which can provide more useful information to both groups and individuals. For example, MCS applications, including traffic monitoring, urban noise mapping, and patient condition monitoring, collect time-series data to provide service. However, the adjunct to the benefits is a serious privacy breach. For example, a malicious service provider can utilize the medical information of a user to send targeted ads or even unjustly determine whether to accept the renewal of his insurance and labor contracts. Usually, time-series data is collected for two goals: one is for the statistical analysis and the other is for the time-series analysis. Hence, the privacy-preserving data collection approaches cannot suppress the data utility too much. To achieve the first goal, Castelluccia et al. [71] designed an inexpensive symmetric-key homomorphic encryption scheme that allows an aggregator to compute aggregates on encrypted data. To achieve the second goal, Papadimitriou et al. [72] firstly consider projecting time-series data into the frequency domain for perturbation, such that the structure of the time-series data can still be preserved. Fan et al. [73] studied time-series sanitization with metric-based privacy and preserved the unique patterns in time-series. However, it is worth noting that all these methods are not for real-time data collection and cannot provide strict guarantee for the accuracy of statistical estimation. To ensure the data utility, the patterns of time-series data, characterized by the correlations of data points, are crucial for personalized services. Hence, Wang et al. [13] focus on real-time data collection against the honest-but-curious platform and aim to protect each user’s privacy while preserving the pattern utility. To this end, the authors proposed a data collection approach PatternLDP, which only samples remarkable points in time series and adaptively perturbs them according to their impacts on local patterns. As shown in Fig. 5, the PatternLDP mainly consists of three parts: pattern-aware sampling, importance characterization and importanceaware randomization. The pattern-aware sampling component is modeled as an optimization problem whose objective is to minimize the number of remarkable points while the representation error does not exceed the error tolerance. To solve the problem efficiently, a criterion called feasible space is adopted to determine the sampling points. Specifically, in real-time sampling, the objective is to find the farthest next

User Privacy Protection in MCS: Threats, Solutions, and Open Issues

Predicted value

Remarkable Point

343

Remarkable Points

Original value

User

Budget Allocation

Perturbed Data

Importance Pattern-aware Sampling

Importance Characterization

Server

Boundary Limitation

PID Error Original Time-series

Importance-aware Randomization

Real-time Data Collection

Fig. 5 The overview of the PatternLDP [13]

remarkable points .s[j ] while the representation error does not exceed the error tolerance ., which can be described as follows: max

|i − j |

s.t.

error(s[k]) ≤ 

∀k ∈ [i, j ]

(16)

where i is the timestamp of last remarkable point, .error(s[k]) is the representation error for data point .s[k] and . is the user-specified error tolerance. Since patterns are represented by the linear connection of remarkable points, the Vertical Distance  between two adjacent (VD) between the actual data point .v[k] and the fit line .ij sampling points can be used as the certain measure for evaluating the representation error: ) error(s[k]) = V D(s[k], ij

.

(17)

Then the slope limitation to indirectly describe the representation error of each point, and the constraint in Eq. 16 can be transformed as: low(s[i], s[k]) ≤ l(s[i], s[j ]) ≤ up(s[i], s[k])

.

(18)

where .low(s[i], s[k]) denote the slope of the straight line connected by points .s[i] and .s[k], l(s[i],s[j]) is the slope of the straight line connected by points .s[i] and .(j, v[j ] − ), and .up(s[i], s[k]) represents the slope of the straight line connected by points .s[i] and .(j, v[j ] + ). Consequently, for the coming data point .s[j ], the limits of the representation error for its slope are: llow = max low(s[i], s[k]) i