Intelligent Data Engineering and Analytics: Proceedings of the 10th International Conference on Frontiers in Intelligent Computing: Theory and Applications (FICTA 2022) 981197523X, 9789811975233

Table of contents :
Organisation
Preface
Contents
About the Editors
1 Implementing Holding Time Based Data Forwarding in Underwater Opportunistic Routing Protocol using Unetstack3
1.1 Introduction
1.2 Related Work
1.3 Hold Time Computation
1.4 Implementation of Hold Time and Packet Forwarding
1.4.1 Hold Time
1.4.2 Data Forwarding
1.4.3 Overhearing
1.5 Experimental Setup
1.6 Results and Analysis
1.7 Conclusion
References
2 Weighted Low-Rank and Sparse Matrix Decomposition Models for Separating Background and Foreground in Dynamic MRI
2.1 Introduction
2.2 Related Work
2.3 Implementation Process of Work
2.4 Algorithms
2.4.1 RPCA
2.4.2 WNNM
2.4.3 WSNM
2.5 Experimental Results and Analysis
2.6 Conclusion
References
3 Remote Sensing Image Fusion Based on PCA and Wavelets
3.1 Introduction
3.2 Proposed Image Fusion Approach in Wavelet Domain
3.2.1 Principal Component Analysis (PCA)
3.2.2 Morphological Hat Transformation
3.2.3 Wavelets
3.3 Experimental Results
3.3.1 Dataset
3.3.2 Simulation Results
3.4 Conclusion
References
4 AgriBlockchain: Agriculture Supply Chain Using Blockchain
4.1 Introduction
4.2 Related Works
4.3 Architecture of Proposed Model
4.3.1 Farmer Contract
4.3.2 Customer Contract
4.3.3 Products Contract
4.3.4 Order Contract
4.3.5 Distributor Contract
4.3.6 Payment Contract
4.4 Security Analysis
4.4.1 Entity Transactions
4.4.2 Attacks Analysis
4.5 Result and Discussions
4.5.1 Experimental Set-up
4.5.2 Performance Analysis
4.6 Conclusion
References
5 Hybrid Energy Systems Integrated Power System Imbalance Cost Calculation Using Moth Swam Algorithm
5.1 Introduction
5.2 Hybrid Energy Systems and Locational Marginal Price
5.3 Problem Formulation
5.4 Moth Swarm Algorithm
5.5 Results and Discussion
5.6 Conclusion
References
6 MPA Optimized Model Predictive Controller for Optimal Control of an AVR System
6.1 Introduction
6.1.1 Background and Literature Survey
6.1.2 Motivation and Contributions
6.2 System Under Consideration
6.3 MPA Based Model Predictive Controller
6.4 Results and Discussion
6.4.1 Basic Test Case
6.4.2 Stability Analysis
6.4.3 Robustness Analysis
6.5 Conclusion
Appendix
References
7 Lightweight Privacy Preserving Framework at Edge Layer in IoT
7.1 Introduction
7.2 Related Work
7.3 Design of Proposed System
7.4 Methodology
7.5 Conclusion
References
8 Concept Drift Aware Analysis of Learning Engagement During COVID-19 Pandemic Using Adaptive Windowing
8.1 Introduction
8.2 Related Work
8.3 Methodology
8.3.1 Experiments with Real Data Set
8.4 Results
8.5 Conclusions and Future Directions
References
9 Optimal Power Flow Considering Uncertainty of Renewable Energy Sources Using Meta-Heuristic Algorithm
9.1 Introduction
9.2 Problem Formulation
9.2.1 Constraints
9.3 CE_CMAES
9.4 Test System and Results Discussion
9.4.1 Statistical Analysis
9.5 Conclusions
References
10 A Triple Band High Gain Antenna Using Metamaterial
10.1 Introduction
10.2 Structural Evolution
10.3 Simulated Results and Discussion
10.4 Conclusions
References
11 Physical Layer Security in MIMO Relay-Based Cognitive Radio Network
11.1 Introduction
11.2 System Model
11.2.1 Direct Transmission (DT)
11.2.2 Amplify and Forward (AF) Relaying
11.2.3 Channel Representation
11.3 Relay Selection Scheme and Problem Formulation
11.4 Simulation Results
11.5 Conclusion
References
12 A Novel Approach for Bug Triaging Using TOPSIS
12.1 Introduction
12.2 Motivation
12.3 Related Work
12.4 Methodology
12.5 Experiment and Results
12.6 Threats to Validity
12.7 Conclusion and Future Scope
References
13 Energy-Efficient Resource Allocation in Cognitive Radio Networks
13.1 Introduction
13.2 System Model
13.3 Resource Allocation
13.3.1 Sub-carrier Allocation
13.3.2 Power Allocation
13.4 Results and Discussions
13.5 Conclusions
References
14 Medical Internet of Things and Data Analytics for Post-COVID Care: An Analysis
14.1 Introduction
14.2 Literature Survey
14.2.1 Internet of Things
14.2.2 Data Analytics for Remote Patient Monitoring
14.3 Proposed Framework
14.4 Post-COVID Data Analysis
14.5 Discussion
14.6 Conclusions
References
15 Mammography Image Classification and Detection by Bi-LSTM with Residual Network Using XG-Boost Approach
15.1 Introduction
15.1.1 Classification and Detection of Breast Masses
15.1.2 Pre-processing of Mammogram
15.1.3 Digital Mammogram Segmentation
15.1.4 Mammogram-Based Feature Extraction
15.1.5 Classification of Mammogram
15.2 Related Works
15.3 The Proposed Method
15.3.1 Proposed Methodology: Flowchart
15.4 Results and Discussion
15.5 Conclusion
References
16 ECG Biometric Recognition by Convolutional Neural Networks with Transfer Learning Using Random Forest Approach
16.1 Introduction
16.1.1 ECG Based Biometric Systems
16.1.2 Biometric Identification System Based on Deep-ECG
16.2 Related Works
16.3 The Proposed Method
16.3.1 Proposed Methodology: Flowchart
16.3.2 Transfer Learning Algorithm
16.4 Results and Discussion
16.5 Conclusion
References
17 Design and Analysis of a Metamaterial-Based Butterworth Microstrip Filter
17.1 Introduction
17.2 Design and Analysis
17.2.1 Geometrical Pattern
17.2.2 Design Procedure: Conventional Low Pass Filter
17.2.3 Simulations and Analysis
17.3 Fabrication and Measurement
17.4 Conclusion
References
18 Internet of Things-Enabled Irrigation System in Precision Agriculture
18.1 Introduction
18.2 Literature Survey
18.3 Case Study: IoT-Enabled Irrigation System in Precision Agriculture
18.4 Other Applications of Precision Agriculture
18.4.1 Soil Sampling and Mapping
18.4.2 Fertilizer
18.4.3 Crop Disease and Pest Management
18.4.4 Yield Monitoring, Forecasting, and Harvesting
18.5 Challenges and Existing Solutions
18.5.1 Weak Internet Connectivity in Agriculture Fields
18.5.2 High Hardware Costs
18.5.3 Disrupted Connectivity to Cloud
18.5.4 Lack of Infrastructure
18.5.5 Lack of Security
18.6 Conclusions and Future Work
References
19 A Hybrid Feature Selection Framework for Breast Cancer Prediction Using Mutual Information and AdaBoost-RFE
19.1 Introduction
19.2 Literature Survey
19.3 Methods
19.3.1 Mutual Information
19.3.2 Recursive Feature Elimination
19.3.3 Proposed Hybrid MI-AdaBoost(w)-RFE Feature Selection Method
19.3.4 Data Description and Preprocessing
19.4 Experimental Results and Discussion
19.4.1 Evaluation Matrix
19.4.2 Model Verification
19.4.3 Shortcomings
19.5 Conclusion and Future Work
References
20 2D-CTM and DNA-Based Computing for Medical Image Encryption
20.1 Introduction
20.2 Literature Survey
20.3 Methodology
20.3.1 Key Generation
20.3.2 Encryption
20.3.3 Decryption
20.4 Performance Analysis
20.4.1 Experimental Setup and Dataset Details
20.4.2 Results and Discussion
20.5 Conclusion
References
21 A Review of Financial Fraud Detection in E-Commerce Using Machine Learning
21.1 Introduction
21.2 Fraud Detection Techniques
21.2.1 Rule-Based Fraud Detection
21.2.2 Machine Learning for Fraud Detection
21.3 Literature Survey
21.3.1 Supervised Fraud Detection
21.3.2 Unsupervised Fraud Detection
21.3.3 Semi-supervised Fraud Detection
21.4 Community-Based Fraud Detection on E-Commerce
21.5 Evaluation Measures
21.6 Challenges in Fraud Detection
21.7 Conclusion
References
22 ReEDNet-An Encoder–Decoder Framework for Single Image Dehazing
22.1 Introduction
22.2 The Proposed Network
22.2.1 The Proposed ReEDNet
22.3 Analysis of Network
22.4 Experiments and Results and Analysis
22.4.1 Implementation Details
22.4.2 Dataset
22.4.3 Evaluation Metric
22.4.4 Quantitative Analysis
22.4.5 Qualitative Analysis
22.5 Conclusion
References
23 Detection of Flood Events from Satellite Images Using Deep Learning
23.1 Introduction
23.2 Related Work
23.3 Methodology
23.3.1 Image Acquisition
23.3.2 Preprocessing of the Images
23.3.3 Creating Semantic Segmentation Model
23.3.4 Training Model
23.3.5 Testing Model
23.4 Results and Analysis
23.5 Conclusion
References
24 Development and Implementation of an Efficient Deep Residual Network for ECG Classification
24.1 Introduction
24.2 Methodology
24.2.1 Neural Network Training and Architecture
24.2.2 Quantized Neural Networks
24.2.3 Deployment
24.3 Dataset Description and Analysis
24.4 Numerical Experiments
24.5 Conclusion and Future Scope
References
25 Study of Class Incremental Learning Strategies for Intrusion Detection System
25.1 Introduction
25.2 Literature Review
25.3 Experimental Methodology
25.3.1 Data Preprocessing
25.3.2 Continual Learning Strategies
25.4 Result and Analysis
25.4.1 Results
25.5 Conclusion
References
26 Classification of High-Priority Tweets for Effective Rescue Operations During Natural Disaster Combining Twitter’s Textual and Non-textual Features
26.1 Introduction
26.2 Related Works
26.3 Methodology
26.3.1 Preprocessing
26.3.2 Annotation
26.3.3 GloVe Vectors—Random Forest Classifier Model
26.3.4 Combined Features Random Forest Model (CFRF)
26.4 Results and Discussion
26.4.1 Experiment
26.4.2 Dataset
26.4.3 Results
26.5 Conclusion
References
27 An Energy Efficient Offloading Technique for UAV-Assisted MEC Using Nature Inspired Algorithm
27.1 Introduction
27.2 Models and Notations
27.2.1 Notations
27.2.2 System Models
27.2.3 Energy Usage Model
27.3 Problem Formulation
27.4 Proposed PSO-Based EEFOUM Algorithm
27.4.1 Initialization of Particle
27.4.2 Fitness
27.4.3 Velocity and Position
27.5 Implementation Results
27.6 Conclusion
References
28 Trajectory Planning and Data Collection of UAVs Over Disaster a Affected Areas
28.1 Introduction
28.2 Literature Survey
28.3 System Model
28.4 Problem Formulation
28.5 Proposed Work
28.5.1 Initialization and Chromosome
28.5.2 Fitness
28.5.3 Mutation and Crossover
28.6 Simulation Results
28.7 Conclusion
References
29 Hand Gesture Recognition on Skeletal Data Using Multi-head Neural Network
29.1 Introduction
29.2 Literature Review
29.3 Proposed Architecture and Working
29.4 Dataset
29.5 Experimental Setup
29.6 Results
29.7 Conclusion
References
30 Ship Detection from Satellite Images with Advanced Deep Learning Model (Single Shot Detector (SSD))
30.1 Introduction
30.2 Related Work
30.3 System Model
30.3.1 Workflow of Ship Detection
30.3.2 TensorFlow
30.3.3 Convolution Neural Networks (CNNs)
30.3.4 Single Shot Detector
30.4 Results and Analysis
30.5 Conclusion and Future Work
References
31 Abstractive Text Summarization for English Language Using NLP and Machine Learning Approaches
31.1 Introduction
31.2 Dataset Description
31.3 Literature Review
31.4 Methodology
31.4.1 Pre-processing Data
31.4.2 BiLSTM-Based Approach/1st Approach
31.4.3 RoBERTa-Based Approach/2nd Approach
31.4.4 DistilRoBERTa-Based Approach/3rd Approach
31.5 Results and Discussion
31.5.1 BiLSTM-Based Approach
31.5.2 RoBERTa-Based Approach
31.5.3 DistilRoBERTa-Based Approach
31.6 Conclusion and Future Scope
References
32 Comparative Modeling of CPU Execution Times of AES and ECC-AES Hybrid Algorithms Over Mobile Cloud Environment
32.1 Introduction
32.2 Related Work
32.3 Description of Our Model
32.3.1 Scheme 1
32.3.2 Scheme 2
32.4 Experimental Analysis of Data
32.4.1 Scheme 1
32.4.2 Scheme 2
32.5 Comparison of Models
32.5.1 Comparison Between Scheme 1 and 2 Based on Hybrid Algorithm.
32.5.2 Comparison Between Scheme 1 and 2 Based on AES Algorithm
32.5.3 Comparison Based on Scheme 1 Between Hybrid and AES Algorithms
32.6 Conclusions
References
33 A Multi-view Representation Learning Approach for Seizure Detection Over Multi-channel EEG Signals
33.1 Introduction
33.2 Related Work
33.3 Description of Dataset
33.4 Proposed Approach
33.4.1 Considered Experimental Data of CHB-MIT
33.4.2 Preprocessing of EEG Data
33.4.3 EEG Data Segmentation
33.4.4 Deep Learning Model
33.4.5 Model Evaluation
33.5 Experimental Setup
33.6 Performance Evaluation Criteria
33.6.1 Accuracy
33.6.2 Recall (Sensitivity)
33.6.3 Specificity
33.6.4 F1-Score
33.7 Results and Analysis
33.8 Conclusion and Future Scope
References
34 Machine Learning Approach to Analyze Breast Cancer
34.1 Introduction
34.2 Background
34.3 Model Design and Methodology
34.3.1 Dataset Description
34.3.2 Diagnosis Model
34.4 Implementation Aspects
34.4.1 Results
34.4.2 Discussion
34.5 Conclusion
References
35 A Hybrid Adaptive Image Retrieval Approach by Using Clustering and Neural Network Techniques
35.1 Introduction
35.2 Prerequisite
35.2.1 CBIR Architecture
35.2.2 Image Similarity
35.2.3 Relevance Feedback
35.2.4 Texture Representations
35.2.5 Back Propagation-Based Neural Network (BPNN)
35.3 A Hybrid Adaptive Approach for Content-Based Image Retrieval Method
35.4 Results
35.5 Conclusion and Future Scope
References
36 2D Convolutional LSTM-Based Approach for Human Action Recognition on Various Sensor Data
36.1 Introduction
36.2 Literature Survey
36.3 Methodology
36.3.1 Architecture Overview
36.3.2 Pre-processing
36.3.3 Exploratory Data Analysis (EDA)
36.3.4 Segmentation
36.3.5 Feature Extraction and Classification
36.4 Results and Discussions
36.4.1 Dataset
36.4.2 Evaluation Metrics
36.4.3 Comparative Experimental Results
36.5 Conclusion
References
37 FCPSO: Evaluation of Feature Clustering Using Particle Swarm Optimization for Health Data
37.1 Introduction
37.2 Related Works
37.3 An Overview of Particle Swarm Optimization
37.4 System Model and Problem Formulation
37.5 Proposed Framework
37.5.1 Preprocessing
37.5.2 Initialization of Population and Generation of Cluster
37.5.3 Fitness Computation
37.5.4 Update Position and Velocity
37.6 Data Observation and Simulation Results
37.6.1 Overview of Datasets
37.6.2 Simulation Setup
37.6.3 Simulation Results
37.7 Conclusion
References
38 Geometric Representation of Obstacles Depth in a Partially Unknown Environment for Achieving Optimized Navigation by Mobile Robots
38.1 Introduction
38.2 Related Work
38.3 Methodology
38.3.1 Cubot as Customized Platform
38.3.2 Fused Sensor Data as Input
38.3.3 Graphical Approach of Path Planning
38.4 Experimental Analysis
38.4.1 Sensor Calibration
38.4.2 Sensor Fusion for on Path Obstacle Detection
38.4.3 Graph Theoretic Path Planning Using Fused Sensor Data
38.4.4 Data Derivation
38.5 Conclusion
References
39 Applied Picture Fuzzy Sets to Smart Autonomous Driving Vehicle for Multiple Decision Making in Forest Transportation
39.1 Introduction
39.2 Terms of Picture Fuzzy Sets
39.3 The Proposed Method
39.3.1 Distance from the Vehicle to the Location of the Signal Lights or the Nearest Obstacle
39.3.2 The Angle Created by the Current Direction of the Vehicle with the vehicle’s Next Direction
39.4 Case Study in the Proposed Method
39.4.1 The Status of the Signal Lights in Direction of the Vehicle
39.5 Conclusions
References
40 Handwritten Mathematical Character Recognition Using Machine Learning and Deep Learning
40.1 Introduction
40.2 Literature Survey
40.3 Proposed Work
40.3.1 Data Set
40.3.2 Classification Algorithms and Implementation
40.4 Results and Discussion
40.5 Conclusion and Future Scope
References
41 A Comparative Analysis of Sentiment Analysis and Emotion Detection Problems on Texts
41.1 Introduction
41.2 Related Works
41.3 Methodology
41.4 Experiments
41.4.1 Dataset Annotation Experiments
41.4.2 Machine Learning Experiments
41.5 Result and Analysis
41.5.1 Relation Between Sentiment and Emotion
41.5.2 Machine Learning Experiment Analysis
41.6 Conclusion and Future Work
References
42 Transfer Learning-Based Advanced Deep Learning Architecture for the Identification of HIV-1 Integration Sites Using Imbalanced Dataset
42.1 Introduction
42.2 Literature Background
42.2.1 HIV Integration Site Detection
42.2.2 VGG19
42.3 System and Methods
42.3.1 Experimental Setup and Dataset Specifications
42.3.2 Transfer Learning
42.3.3 Encoding
42.3.4 Convolution Neural Network
42.3.5 Proposed Model for Predicting the HIV IS
42.4 Results and Discussion
42.4.1 Results
42.4.2 Analysis of the Results
42.5 Conclusion and Future Scope
References
43 Optimal Evacuation Planning Using Integer Programming
43.1 Introduction
43.2 Model Framework
43.3 Problem Formulation
43.4 Drawbacks/Challenges
43.5 Conclusion
References
44 Automatic Speech Recognition Analysis Over Wireless Networks
44.1 Introduction
44.2 Background
44.2.1 Voice Over IP System
44.2.2 VoIP Codecs
44.2.3 Automatic Speech Recognition
44.3 Methodology
44.4 Experiment Results
44.5 Conclusions
References
45 Entropy-Based Clustering for Subspace Pattern Discovery in Ordinal Survey Data
45.1 Introduction
45.1.1 Types of Survey Data
45.1.2 Challenges in Survey
45.1.3 Entropy and Clustering
45.2 Background and Related Work
45.3 The Proposed Methodology
45.4 Experimental Results and Discussion
45.4.1 Dataset
45.4.2 Results
45.4.3 Discussion
45.5 Conclusion
References
46 AlexNet Model for Sign Language Recognition
46.1 Introduction
46.2 Proposed Design Methodology
46.2.1 CNN
46.2.2 AlexNet Architecture for SLR
46.3 Experimental Result
46.3.1 Dataset
46.3.2 Simulation Results and Analysis of Segmentation Approach
46.3.3 Performance Evaluation
46.4 Conclusion
References
47 An Improved Query Similarity Model for Online Health Community Forum Using Cross-Attention Mechanism on Siamese Network
47.1 Introduction
47.2 Related Work
47.3 Methodology
47.3.1 Dataset
47.3.2 Query Similarity Model Using Siamese Network with Cross-Attention Mechanism
47.4 Experiment and Results
47.5 Conclusion and Future Direction
References
48 COVID Detection Using Chest X-ray Images Using Ensembled Deep Learning
48.1 Introduction
48.2 Related Work
48.3 Research Approach
48.3.1 Data Collection Phase
48.3.2 Data Preprocessing Phase
48.3.3 Model Used Phase
48.3.4 Ensemble Model for Features Extraction Phase
48.3.5 Implementation and Training Phase
48.4 Result and Analysis
48.4.1 Performance Evaluation of Model
48.5 Conclusion and Future Scope
References
49 Robot Motion Path Planning Using Artificial Bee Colony Algorithm
49.1 Introduction
49.2 Artificial Bee Colony Algorithm
49.2.1 Initialization Phase
49.2.2 Employed Phase
49.2.3 Onlooker Phase
49.2.4 Scout Phase
49.3 Navigation Architecture
49.4 ABC Algorithm Robot Motion Path Planning
49.4.1 Initialization Phase
49.4.2 Scout Phase
49.4.3 Employed Phase
49.4.4 Onlooker Phase
49.5 Simulation Observation and Results
49.5.1 Unbranched Path or Linear Path
49.5.2 Branched Path or Nonlinear Path
49.6 Conclusion
References
50 Predicting the Tomato Plant Disease Using Deep Learning Techniques
50.1 Introduction
50.2 Literature Work
50.3 Research Methodology
50.4 Results
50.5 Conclusion
References
51 Tracking of Lost Objects Using GPS and GSM
51.1 Introduction
51.2 The Overall System
51.2.1 Architecture
51.2.2 Components
51.2.3 Prototype
51.3 Algorithm
51.4 Results
51.5 Conclusion
References
52 Decentralization of Car Insurance System Using Machine Learning and Distributed Ledger Technology
52.1 Introduction
52.2 Literature Survey
52.3 Proposed Methodology
52.3.1 Overall CIAS Architecture Design
52.3.2 Machine Learning Model to Detect Damages
52.3.3 Implementation of Blockchain Technology
52.4 Future Work
52.5 Conclusion
References
53 Heart Disease Detection from Gene Expression Data Using Optimization Driven Deep Q-Network
53.1 Introduction
53.2 Literature Survey
53.2.1 Challenges
53.3 Proposed Methodology
53.4 Heart Disease Detection Using Deep Q-Network
53.4.1 Training Procedure of Proposed PDHO Algorithm
53.5 Results and Discussion
53.6 Conclusion
References
54 Learning-Based Scheme for Efficient Content Caching in Vehicular Networks
54.1 Introduction
54.2 Related Work
54.3 System Model
54.3.1 Caching of Content
54.3.2 Delivery of Content
54.3.3 Problem Formulation
54.4 Learning Automata-Based Content Caching
54.4.1 Proposed Algorithm
54.5 Performance Evaluation
54.5.1 Simulation Environment Set-up
54.5.2 Overall Performance
54.6 Conclusion
References
55 Adaptive Resource Allocation in WiMAX Networks for Improved Quality of Service (QoS)
55.1 Introduction
55.1.1 Related Work
55.2 2-Level Scheduling Algorithm
55.2.1 Level-1 Scheduling
55.2.2 Level-2 Scheduling
55.3 Simulation Results
55.4 Conclusions
References
Author Index

Citation preview

Smart Innovation, Systems and Technologies 327

Vikrant Bhateja Xin-She Yang Jerry Chun-Wei Lin Ranjita Das   Editors

Intelligent Data Engineering and Analytics Proceedings of the 10th International Conference on Frontiers in Intelligent Computing: Theory and Applications (FICTA 2022)

123

Smart Innovation, Systems and Technologies Volume 327

Series Editors Robert J. Howlett, Bournemouth University and KES International, Shoreham-by-Sea, UK Lakhmi C. Jain, KES International, Shoreham-by-Sea, UK

The Smart Innovation, Systems and Technologies book series encompasses the topics of knowledge, intelligence, innovation and sustainability. The aim of the series is to make available a platform for the publication of books on all aspects of single and multi-disciplinary research on these themes in order to make the latest results available in a readily-accessible form. Volumes on interdisciplinary research combining two or more of these areas is particularly sought. The series covers systems and paradigms that employ knowledge and intelligence in a broad sense. Its scope is systems having embedded knowledge and intelligence, which may be applied to the solution of world problems in industry, the environment and the community. It also focusses on the knowledge-transfer methodologies and innovation strategies employed to make this happen effectively. The combination of intelligent systems tools and a broad range of applications introduces a need for a synergy of disciplines from science, technology, business and the humanities. The series will include conference proceedings, edited collections, monographs, handbooks, reference books, and other relevant types of book in areas of science and technology where smart systems and technologies can offer innovative solutions. High quality content is an essential feature for all book proposals accepted for the series. It is expected that editors of all accepted volumes will ensure that contributions are subjected to an appropriate level of reviewing process and adhere to KES quality principles. Indexed by SCOPUS, EI Compendex, INSPEC, WTI Frankfurt eG, zbMATH, Japanese Science and Technology Agency (JST), SCImago, DBLP. All books published in the series are submitted for consideration in Web of Science.

Vikrant Bhateja · Xin-She Yang · Jerry Chun-Wei Lin · Ranjita Das Editors

Intelligent Data Engineering and Analytics Proceedings of the 10th International Conference on Frontiers in Intelligent Computing: Theory and Applications (FICTA 2022)

Editors Vikrant Bhateja Department of Electronics Engineering Faculty of Engineering and Technology Veer Bahadur Singh Purvanchal University Jaunpur, Uttar Pradesh, India Jerry Chun-Wei Lin Western Norway University of Applied Sciences Bergen, Norway

Xin-She Yang School of Science and Technology Middlesex University London London, UK Ranjita Das Department of Computer Science and Engineering National Institute of Technology Agartala Agartala, West Tripura, India

ISSN 2190-3018 ISSN 2190-3026 (electronic) Smart Innovation, Systems and Technologies ISBN 978-981-19-7523-3 ISBN 978-981-19-7524-0 (eBook) https://doi.org/10.1007/978-981-19-7524-0 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 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 Singapore Pte Ltd. The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721, Singapore

Organisation

Chief Patron Director, NIT Mizoram

Patrons Prof. Saibal Chatterjee, Dean (Academics), NIT Mizoram Dr. Alok Shukla, Dean (Dean RC), NIT Mizoram Dr. P. Ajmal Koya, Dean (Faculty Welfare), NIT Mizoram Dr. K. Gyanendra Singh, Dean (Students’ Welfare), NIT Mizoram

General Chairs Dr. Jinshan Tang, College of Computing, Michigan Technological University, Michigan, USA Dr. Xin-She Yang, Middlesex University London, UK

Publication Chairs Dr. Jerry Chun-Wei Lin, Western Norway University of Applied Sciences, Bergen, Norway Dr. Peter Peer, Faculty of Computer and Information Science, University of Ljubljana, Slovenia

v

vi

Organisation

Prof. Vikrant Bhateja, Veer Bahadur Singh Purvanchal University, Jaunpur, Uttar Pradesh, India

Organising Chairs Dr. Ranjita Das, Assistant Professor, Department of CSE, NIT Mizoram Dr. Sandeep Kumar Dash, Department of CSE, NIT Mizoram

Publicity Chairs Dr. Sandeep Kumar Dash, Department of CSE, NIT Mizoram Mr. Lenin Laitonjam, Department of CSE, NIT Mizoram Ms. B. Sneha Reddy, Department of CSE, NIT Mizoram Dr. Amit Kumar Roy, Department of CSE, NIT Mizoram Dr. Ranjita Das, Assistant Professor, Department of CSE, NIT Mizoram

Advisory Committee Aime’ Lay-Ekuakille, University of Salento, Lecce, Itlay Annappa Basava, Department of CSE, NIT Karnataka Amira Ashour, Tanta University, Egypt Aynur Unal, Standford University, USA Bansidhar Majhi, IIIT Kancheepuram, Tamil Nadu, India Dariusz Jacek Jakobczak, Koszalin University of Technology, Koszalin, Poland Dilip Kumar Sharma, IEEE UP Section Ganpati Panda, IIT Bhubaneswar, Odisha, India Jagdish Chand Bansal, South Asian University, New Delhi, India João Manuel R. S. Tavares, Universidade do Porto (FEUP), Porto, Portugal Jyotsana Kumar Mandal, University of Kalyani, West Bengal, India K. C. Santosh, University of South Dakota, USA Le Hoang Son, Vietnam National University, Hanoi, Vietnam Milan Tuba, Singidunum University, Belgrade, Serbia Naeem Hanoon, Multimedia University, Cyberjaya, Malaysia Nilanjan Dey, TIET, Kolkata, India Noor Zaman, Universiti Tecknologi, PETRONAS, Malaysia Pradip Kumar Das, Professor, Department of CSE, IIT Guwahati Roman Senkerik, Tomas Bata University in Zlin, Czech Republic Sandeep Singh Sengar, Cardiff Metropolitan University, UK Sriparna Saha, Associate Professor, Department of CSE, IIT Patna

Organisation

vii

Sukumar Nandi, Department of CSE, IIT Guwahati Swagatam Das, Indian Statistical Institute, Kolkata, India Siba K. Udgata, University of Hyderabad, Telangana, India Tai Kang, Nanyang Technological University, Singapore Ujjawl Maulic, Department of CSE, Jadavpur University Valentina Balas, Aurel Vlaicu University of Arad, Romania Yu-Dong Zhang, University of Leicester, UK

Technical Program Committee Chairs Dr. Steven L. Fernandes, Creighton University, USA Dr. Mufti Mahmud, Nottingham Trent University, Nottingham, UK

Technical Program Committee A. K. Chaturvedi, Department of Electrical Engineering, IIT Kanpur, India Abdul Rajak A. R., Department of Electronics and Communication Engineering, Birla Institute of Technology and Science Dr. Nitika Vats Doohan, Indore, India Ahmad Al-Khasawneh, The Hashemite University, Jordan Alexander Christea, University of Warwick, London, UK Amioy Kumar, Biometrics Research Lab, Department of Electrical Engineering, IIT Delhi, India Anand Paul, The School of Computer Science and Engineering, South Korea Anish Saha, NIT Silchar Apurva A. Desai, Veer Narmad South Gujarat University, Surat, India Avdesh Sharma, Jodhpur, India Bharat Singh Deora, JRNRV University, India Bhavesh Joshi, Advent College, Udaipur, India Brent Waters, University of Texas, Austin, Texas, USA Chhaya Dalela, Associate Professor, JSSATE, Noida, Uttar Pradesh, India Dan Boneh, Computer Science Department, Stanford University, California, USA Dipankar Das, Jadavpur University Feng Jiang, Harbin Institute of Technology, China Gengshen Zhong, Jinan, Shandong, China Harshal Arolkar, Immd. Past Chairman, CSI Ahmedabad Chapter, India H. R. Vishwakarma, Professor, VIT, Vellore, India Jayanti Dansana, KIIT University, Bhubaneswar, Odisha, India Jean Michel Bruel, Departement Informatique IUT de Blagnac, Blagnac, France Jeril Kuriakose, Manipal University, Jaipur, India Jitender Kumar Chhabra, NIT, Kurukshetra, Haryana, India

viii

Organisation

Junali Jasmine Jena, KIIT DU, Bhubaneswar, India Jyoti Prakash Singh, NIT Patna K. C. Roy, Principal, Kautaliya, Jaipur, India Kalpana Jain, CTAE, Udaipur, India Komal Bhatia, YMCA University, Faridabad, Haryana, India Krishnamachar Prasad, Department of Electrical and Electronic Engineering, Auckland, New Zealand Lipika Mohanty, KIIT DU, Bhubaneswar, India Lorne Olfman, Claremont, California, USA Martin Everett, University of Manchester, England Meenakhi Rout, KIIT DU, Bhubaneswar, India Meenakshi Tripathi, MNIT, Jaipur, India Mrinal Kanti Debbarma, NIT Agartala M. Ramakrishna, ANITS, Vizag, India Mukesh Shrimali, Pacific University, Udaipur, India Murali Bhaskaran, Dhirajlal Gandhi College of Technology, Salem, Tamil Nadu, India Ngai-Man Cheung, Assistant Professor, University of Technology and Design, Singapore Neelamadhav Padhi, GIET University, Odisha, India Nilay Mathur, Director, NIIT Udaipur, India Philip Yang, Price Water House Coopers, Beijing, China Pradeep Chouksey, Principal, TIT College, Bhopal, Madhya Pradesh, India Prasun Sinha, Ohio State University Columbus, Columbus, OH, USA R. K. Bayal, Rajasthan Technical University, Kota, Rajasthan, India Rajendra Kumar Bharti, Assistant Professor, Kumaon Engineering College, Dwarahat, Uttarakhand, India S. R. Biradar, Department of Information Science and Engineering, SDM College of Engineering and Technology, Dharwad, Karnataka, India Sami Mnasri, IRIT Laboratory Toulouse, France Savita Gandhi, Professor, Gujarat University, Ahmedabad, India Soura Dasgupta, Department of TCE, SRM University, Chennai, India Sushil Kumar, School of Computer and Systems Sciences, Jawaharlal Nehru University, New Delhi, India Ting-Peng Liang, National Chengchi University Taipei, Taiwan V. Rajnikanth, EIE Department, St. Joseph’s College of Engineering, Chennai, India Veena Anand, NIT Raipur Xiaoyi Yu, National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, China Yun-Bae Kim, SungKyunKwan University, South Korea

Preface

This book is a collection of high-quality peer-reviewed research papers presented at the 10th International Conference on Frontiers in Intelligent Computing: Theory and Applications (FICTA-2022) held at National Institute of Technology, Mizoram, Aizawl, India, during June 18–19, 2022, (Decennial Edition of FICTA Conference). The idea of this conference series was conceived by few eminent professors and researchers from premier institutions of India. The first three editions of this conference: FICTA-2012, 2013, and 2014 were organized by Bhubaneswar Engineering College (BEC), Bhubaneswar, Odisha, India. The fourth edition FICTA-2015 was held at NIT, Durgapur, West Bengal, India. The fifth and sixth editions FICTA-2016 and FICTA-2017 were consecutively organized by KIIT University, Bhubaneswar, Odisha, India. FICTA-2018 was hosted by Duy Tan University, Da Nang City, Vietnam. The eighth edition FICTA 2020 was held at NIT, Karnataka, Surathkal, India. The ninth edition FICTA-2021 was held at NIT, Mizoram, Aizawl, India. All past editions of the FICTA conference proceedings are published by Springer. Presently, FICTA-2022 is the tenth edition of this conference series which aims to bring together researchers, scientists, engineers, and practitioners to exchange and share their theories, methodologies, new ideas, experiences, applications in all areas of intelligent computing theories and applications to various engineering disciplines like Computer Science, Electronics, Electrical, Mechanical, Bio-Medical Engineering, etc. FICTA-2022 had received a good number of submissions from the different areas relating to computational intelligence, intelligent data engineering, data analytics, decision sciences, and associated applications in the arena of intelligent computing. These papers have undergone a rigorous peer-review process with the help of our technical program committee members (from the country as well as abroad). The review process has been very crucial with minimum two reviews each; in many cases, 3–5 reviews along with due checks on similarity and content overlap as well. This conference witnessed more than 500 submissions including the main track as well as special sessions. The conference featured many special sessions in various cutting-edge technologies of specialized focus which were organized and chaired by eminent professors. The total toll of papers included submissions received across ix

x

Preface

country along with many overseas countries. Out of this pool, only 110 papers were given acceptance and segregated as two different volumes for publication under the proceedings. This volume consists of 55 papers from diverse areas of Evolution in Computational Intelligence. The conference featured many distinguished keynote addresses in different spheres of intelligent computing by eminent speakers like: Dr. Xin-She Yang (Reader at Middlesex University London, UK) and Dr. Sumit K. Jha (Professor of Computer Science at the University of Texas-San Antonio (UTSA) US.) Dr. Xin-She Yang’s keynote lecture on “Nature-Inspired Algorithms: Insights and Open Problems” gives an idea on nature-inspired algorithms such as the particle swarm optimization, bat algorithm, and firefly algorithm which have been widely used to solve problems in optimization, data mining, and computational intelligence. Also, Dr. Sumit K. Jha’s talk on the Trust in Artificial Intelligence received ample applause from the vast audience of delegates, budding researchers, faculty, and students. We thank the advisory chairs and steering committees for rendering mentor support to the conference. An extreme note of gratitude is given to Dr. Sandeep Kumar Dash (Head, Department of CSE, NIT Mizoram, Aizawl, India) and Dr. Ranjita Das (Department of CSE, NIT Mizoram, Aizawl, India) for providing valuable guidelines and being an inspiration in the entire process of organizing this conference. We would also like to thank Department of Computer Science and Engineering, NIT Mizoram, Aizawl, India, who came forward and provided their support to organize the tenth edition of this conference series. We take this opportunity to thank authors of all submitted papers for their hard work, adherence to the deadlines, and patience with the review process. The quality of a refereed volume depends mainly on the expertise and dedication of the reviewers. We are indebted to the technical program committee members who not only produced excellent reviews but also did these in short time frames. We would also like to thank the participants of this conference, who have participated in the conference above all hardships. Jaunpur, Uttar Pradesh, India London, UK Bergen, Norway Agartala, West Tripura, India

Prof. Vikrant Bhateja Dr. Xin-She Yang Dr. Jerry Chun-Wei Lin Dr. Ranjita Das

Contents

1

2

Implementing Holding Time Based Data Forwarding in Underwater Opportunistic Routing Protocol using Unetstack3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . K. Bhujange, B. R. Chandavarkar, and P. Nazareth Weighted Low-Rank and Sparse Matrix Decomposition Models for Separating Background and Foreground in Dynamic MRI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . K. V. Sridhar, Ramireddy Mounica, and Madishetty Saisantosh Raviteja

1

13

3

Remote Sensing Image Fusion Based on PCA and Wavelets . . . . . . . Reetika Mishra, Vikrant Bhateja, Rupa Banerjee, Aime’ Lay-Ekuakille, and Roman Senkerik

25

4

AgriBlockchain: Agriculture Supply Chain Using Blockchain . . . . . Rishikesh, Shubham Kant Ajay, and Ditipriya Sinha

35

5

Hybrid Energy Systems Integrated Power System Imbalance Cost Calculation Using Moth Swam Algorithm . . . . . . . . . . . . . . . . . . Nitesh Kumar Singh, Chaitali Koley, and Sadhan Gope

49

MPA Optimized Model Predictive Controller for Optimal Control of an AVR System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Veena Sharma, Vineet Kumar, R. Naresh, and V. Kumar

61

6

7

8

Lightweight Privacy Preserving Framework at Edge Layer in IoT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Kavita Agrawal, Suresh Chittineni, P. V. G. D. Prasad Reddy, K. Subhadra, and Elizabeth D. Diaz Concept Drift Aware Analysis of Learning Engagement During COVID-19 Pandemic Using Adaptive Windowing . . . . . . . . . Sachin Gupta and Bhoomi Gupta

71

81

xi

xii

9

Contents

Optimal Power Flow Considering Uncertainty of Renewable Energy Sources Using Meta-Heuristic Algorithm . . . . . . . . . . . . . . . . . Jigar Sarda, Hirva S. Patel, Kwang Y. Lee, and Kartik Pandya

93

10 A Triple Band High Gain Antenna Using Metamaterial . . . . . . . . . . . 105 Souvik Halder, Abhijyoti Ghosh, Sudipta Chattopadhyay, and L. Lolit Kumar Singh 11 Physical Layer Security in MIMO Relay-Based Cognitive Radio Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 V. Upadhyay and M. R. Bharti 12 A Novel Approach for Bug Triaging Using TOPSIS . . . . . . . . . . . . . . . 125 Pavan Rathoriya, Rama Ranjan Panda, and Naresh Kumar Nagwani 13 Energy-Efficient Resource Allocation in Cognitive Radio Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137 Kartik Verma and Manoranjan Rai Bharti 14 Medical Internet of Things and Data Analytics for Post-COVID Care: An Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 Salka Rahman, Shabir Ahmad Sofi, Suraiya Parveen, and Saniya Zahoor 15 Mammography Image Classification and Detection by Bi-LSTM with Residual Network Using XG-Boost Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 Aman Chhabra and Manoranjan Rai Bharti 16 ECG Biometric Recognition by Convolutional Neural Networks with Transfer Learning Using Random Forest Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177 Ankur and Manoranjan Rai Bharti 17 Design and Analysis of a Metamaterial-Based Butterworth Microstrip Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 Samiran Pramanik, Kamlendra Kumar Rathour, Edhot Chakma, and Chaitali Koley 18 Internet of Things-Enabled Irrigation System in Precision Agriculture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201 Shabir A. Sofi and Saniya Zahoor 19 A Hybrid Feature Selection Framework for Breast Cancer Prediction Using Mutual Information and AdaBoost-RFE . . . . . . . . 213 Himanshu Dhoke and Aakanksha Sharaff 20 2D-CTM and DNA-Based Computing for Medical Image Encryption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225 Mobashshirur Rahman and Piyush Kumar

Contents

xiii

21 A Review of Financial Fraud Detection in E-Commerce Using Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237 Abhay Narayan, S. D. Madhu Kumar, and Anu Mary Chacko 22 ReEDNet-An Encoder–Decoder Framework for Single Image Dehazing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249 Konark Keshaw, Abhishek Pandey, Gopa Bhaumik, and M C Govil 23 Detection of Flood Events from Satellite Images Using Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259 Anushree Rambhad, Dhirendra Pratap Singh, and Jaytrilok Choudhary 24 Development and Implementation of an Efficient Deep Residual Network for ECG Classification . . . . . . . . . . . . . . . . . . . . . . . . 269 Rishabh Arya, Ujjawal Agrawal, Ananya Singh, Eshaan Gupta, and Priya Ranjan Muduli 25 Study of Class Incremental Learning Strategies for Intrusion Detection System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281 Parvati Bhurani, Satyendra Singh Chouhan, and Namita Mittal 26 Classification of High-Priority Tweets for Effective Rescue Operations During Natural Disaster Combining Twitter’s Textual and Non-textual Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293 E. Arathi and S. Sasikala 27 An Energy Efficient Offloading Technique for UAV-Assisted MEC Using Nature Inspired Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 309 Santanu Ghosh, Pratyay Kuila, and Tarun Biswas 28 Trajectory Planning and Data Collection of UAVs Over Disaster a Affected Areas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319 Abisek Dahal, Santanu Ghosh, Pratyay Kuila, and Tarun Biswas 29 Hand Gesture Recognition on Skeletal Data Using Multi-head Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327 Amrita Rai and Rajneesh Rani 30 Ship Detection from Satellite Images with Advanced Deep Learning Model (Single Shot Detector (SSD)) . . . . . . . . . . . . . . . . . . . . 337 Johnson Kolluri and Ranjita Das 31 Abstractive Text Summarization for English Language Using NLP and Machine Learning Approaches . . . . . . . . . . . . . . . . . . . . . . . . 351 Kartik Singhal and Anupam Agrawal

xiv

Contents

32 Comparative Modeling of CPU Execution Times of AES and ECC-AES Hybrid Algorithms Over Mobile Cloud Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363 Ambili Thomas and V. Lakshmi Narasimhan 33 A Multi-view Representation Learning Approach for Seizure Detection Over Multi-channel EEG Signals . . . . . . . . . . . . . . . . . . . . . . 375 Shubham Chandra Joshi, Gopal Chandra Jana, and Anupam Agrawal 34 Machine Learning Approach to Analyze Breast Cancer . . . . . . . . . . . 387 Satya Ranjan Dash, Saurav Roy, Jnyana Ranjan Mohanty, Dulani Meedeniya, and Manoj Ranjan Mishra 35 A Hybrid Adaptive Image Retrieval Approach by Using Clustering and Neural Network Techniques . . . . . . . . . . . . . . . . . . . . . . 395 Pankaj Pratap Singh and Shitala Prasad 36 2D Convolutional LSTM-Based Approach for Human Action Recognition on Various Sensor Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405 Ajeet Pandey, Piyush Kumar, and Shitala Prasad 37 FCPSO: Evaluation of Feature Clustering Using Particle Swarm Optimization for Health Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 419 Pintu Kumar Ram and Pratyay Kuila 38 Geometric Representation of Obstacles Depth in a Partially Unknown Environment for Achieving Optimized Navigation by Mobile Robots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429 Rapti Chaudhuri, Suman Deb, and Hillol Das 39 Applied Picture Fuzzy Sets to Smart Autonomous Driving Vehicle for Multiple Decision Making in Forest Transportation . . . . 441 Hai Van Pham and Hai Nam Nguyen 40 Handwritten Mathematical Character Recognition Using Machine Learning and Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . 455 Priti Yadav, Bijendra Kumar Ray, and Gargi Khanna 41 A Comparative Analysis of Sentiment Analysis and Emotion Detection Problems on Texts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 465 Aparajita Sinha and Kunal Chakma 42 Transfer Learning-Based Advanced Deep Learning Architecture for the Identification of HIV-1 Integration Sites Using Imbalanced Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 477 Minakshi Boruah and Ranjita Das 43 Optimal Evacuation Planning Using Integer Programming . . . . . . . . 491 Aastha Goel

Contents

xv

44 Automatic Speech Recognition Analysis Over Wireless Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499 Mohamed Hamidi, Ouissam Zealouk, and Hassan Satori 45 Entropy-Based Clustering for Subspace Pattern Discovery in Ordinal Survey Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509 Bhupendera Kumar and Rajeev Kumar 46 AlexNet Model for Sign Language Recognition . . . . . . . . . . . . . . . . . . . 521 Shreya Singh, Vikrant Bhateja, Shivangi Srivastav, Pratiksha, Jerry Chun-Wei Lin, and Carlos M. Travieso-Gonzalez 47 An Improved Query Similarity Model for Online Health Community Forum Using Cross-Attention Mechanism on Siamese Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 531 B. Athira and Sumam Mary Idicula 48 COVID Detection Using Chest X-ray Images Using Ensembled Deep Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 543 Rohit Beniwal, Arun Vaishy, Aryan, and Gaurav Kumar Dhama 49 Robot Motion Path Planning Using Artificial Bee Colony Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 555 Manish and Sushil Kumar 50 Predicting the Tomato Plant Disease Using Deep Learning Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567 Rishikesh Bhupendra Trivedi, Daksh Mittal, Anuj Sahani, Clely Voyena Fernandes, Somya Goyal, Jyotir Moy Chatterjee, and Vanshika Mehta 51 Tracking of Lost Objects Using GPS and GSM . . . . . . . . . . . . . . . . . . . 577 Mukul Nagar, Himanish Bhattacharya, Ayush Gupta, Somya Goyal, and Istiaque Ahmed 52 Decentralization of Car Insurance System Using Machine Learning and Distributed Ledger Technology . . . . . . . . . . . . . . . . . . . . 587 Saroj Kumar Nanda, Sandeep Kumar Panda, Madhabananda Das, and Suresh Chandra Satapathy 53 Heart Disease Detection from Gene Expression Data Using Optimization Driven Deep Q-Network . . . . . . . . . . . . . . . . . . . . . . . . . . 601 Chetan Nimba Aher and Ajay Kumar Jena 54 Learning-Based Scheme for Efficient Content Caching in Vehicular Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 613 Satish Vemireddy, Sai Virinchi, R Vineeth Kumar, and Nihanth Kumar

xvi

Contents

55 Adaptive Resource Allocation in WiMAX Networks for Improved Quality of Service (QoS) . . . . . . . . . . . . . . . . . . . . . . . . . . 625 V. Hindumathi, Santhosh Kumar Veeramalla, and T. Vasudeva Reddy Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 635

About the Editors

Dr. Vikrant Bhateja is Associate Professor in Department of Electronics Engineering, Faculty of Engineering and Technology, Veer Bahadur Singh Purvanchal University, Jaunpur, Uttar Pradesh, India. He holds a doctorate in ECE (Bio-Medical Imaging) with a total academic teaching experience of 19+ years with around 190 publications in reputed international conferences, journals and online chapter contributions, out of which 35 papers are published in SCIE indexed high impact factored journals. Among the international conference publications, four papers have received “Best Paper Award”. Among the SCIE publications, one paper published in Review of Scientific Instruments (RSI) Journal (under American International Publishers) has been selected as “Editor Choice Paper of the Issue” in 2016. He has been instrumental in chairing/co-chairing around 30 international conferences in India and abroad as Publication/TPC Chair and edited 50 book volumes from Springer Nature as Corresponding/Co-editor/Author on date. He has delivered nearly 20 keynotes, invited talks in international conferences, ATAL, TEQIP and other AICTE sponsored FDPs and STTPs. He has been Editor-in-Chief of IGI Global--International Journal of Natural Computing and Research (IJNCR) an ACM & DBLP indexed journal from 2017 to 22. He has guest edited Special Issues in reputed SCIE indexed journals under Springer Nature and Elsevier. He is Senior Member of IEEE and Life Member of CSI. Xin-She Yang obtained his D.Phil. in Applied Mathematics from the University of Oxford. He then worked at Cambridge University and National Physical Laboratory (UK) as Senior Research Scientist. Now he is Reader at Middlesex University London, Fellow of the Institute of Mathematics and its Application (IMA), and Book Series co-Editor of the Springer Tracts in Nature-Inspired Computing. He was also IEEE Computational Intelligence Society Task Force Chair for Business Intelligence and Knowledge Management (2015 to 2020). He has published more than 25 books and more than 400 peer-reviewed research publications with over 73000 citations, and he has been on the prestigious list of highly cited researchers (Web of Sciences) for seven consecutive years (2016–2022).

xvii

xviii

About the Editors

Jerry Chun-Wei Lin received his Ph.D. from the Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan, Taiwan in 2010. He is currently Full Professor in the Department of Computer Science, Electrical Engineering and Mathematical Sciences, Western Norway University of Applied Sciences, Bergen, Norway. He has published more than 500 research articles in top-tier journals and conferences, 12 edited books, as well as 33 patents (held and 36 filed, 3 US patents). His research interests include data analytics, soft computing, artificial intelligence/machine learning, optimization, IoT, and privacypreserving and security technologies and bio-informatics. He is Editor-in-Chief of the International Journal of Data Science and Pattern Recognition and Associate Editor for several top-tier journals, e.g., IEEE TNNLS, IEEE TDSC, IEEE TCYB, among others. He has been recognized as most cited Chinese Researcher respectively in 2018, 2019, 2020, and 2021 by Scopus/Elsevier. He is Fellow of IET (FIET), ACM Distinguished Scientist, and IEEE Senior Member. Dr. Ranjita Das is currently working as Assistant Professor in the Department of Computer Science and Engineering, National Institute of Technology Agartala. She was previously working in NIT Mizoram as Assistant Professor. She has total 11 years of teaching experience. She did her Ph.D. from NIT Mizoram, M.Tech. from Tezpur University, and B.Tech. from NIT Agartala. Dr. Das’s research covers wide areas related to pattern recognition, computational biology, information retrieval, and image processing. She has published several research papers in reputed journals and conferences. She has been Program Chair and Organizing Chair of many international conferences and international workshops. She also serves as Reviewer and Technical Program Committee Member for a number of IEEE and Springer journals and conferences. She won the best paper awards in FICTA-2021, IC4E2020, ICACCP-2019 and IEEE-INDICON 2015. She has been involved in sponsored research projects in the broad areas of computational biology and machine learning technologies, funded by the DBT and DST-SERB, etc. She was Board of Governor Member of NIT Mizoram.

Chapter 1

Implementing Holding Time Based Data Forwarding in Underwater Opportunistic Routing Protocol using Unetstack3 K. Bhujange, B. R. Chandavarkar, and P. Nazareth Abstract Underwater Acoustic Sensor Networks (UASNs) play a vital role in many applications such as underwater resource exploration, environment monitoring, and surveillance. These applications demand high reliability. However, achieving reliable communication in the UASNs is one of the challenging issues due to varying link quality. To improve the reliable transmission in the UASNs, Opportunistic Routing (OR) is a promising routing technique. One of the challenges OR protocols face is cooperation between nodes. To overcome this, many techniques have been found. Holding time is one of the mainly used techniques to coordinate between nodes. Various holding time algorithms already exist. The difference in the algorithms is the use of different parameters for calculating hold time, and depth difference is the most used parameter. The major goal of this paper is to create a framework in UnetStack3 for holding time-based protocols that can be used for numerous different protocols to compute hold time and overhear neighboring packets by doing minor modifications in the calculations.

1.1 Introduction Water covers the majority of the Earth’s surface, and exploring the deeper portions of the water bodies is difficult. Because of the considerable attenuation and time fluctuations, as well as the very narrow bandwidth, underwater networks use acoustic waves rather than electromagnetic waves. Underwater Acoustic Sensor Networks (UASNs) enable us to explore underwater regions while also allowing communication across oceans and providing oceanographic data. UASNs may also be used for government surveillance, in submarines to communicate with one another and to obtain precise information on aquatic fauna [1]. Network dynamics, energy conservation, fault tolerance, scalability, environmental considerations, data aggregation, and other issues must be addressed by UASNs [2, 3]. Further, each transmission’s K. Bhujange · B. R. Chandavarkar · P. Nazareth (B) Wireless Information Networking Group (WiNG), Department of Computer Science and Engineering, National Institute of Technology Karnataka, Surathkal, Mangalore, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_1

1

2

K. Bhujange et al.

reliability is not guaranteed; there is a considerable likelihood that the simple unicast transmission will fail. To address these issues, various routing protocols have been developed. One such protocol is the Opportunistic Routing (OR) protocol, in which data is transmitted to the most suited candidate or a cluster of candidate nodes [4]. The optimal path selection aids in energy conservation and leads to a longer network lifetime. When using numerous nodes, there is a greater risk of collision, which leads to increased network overload. The nodes must coordinate so that only one node transmits during the allotted time and the other nodes wait for it to complete the transmission [5]. The framework works on the principles of holding time-based data forwarding algorithm that can coordinate nodes so that only the high-priority node will transmit the data while other nodes in the cluster should be in a holding state [6]. Also during the hold time, the node should overhear any transmission done by the neighboring nodes [7], which can be used to drop the packet in case the same packet has already been forwarded. The algorithm is implemented using UnetStack3 [8] an agent-based underwater network simulator. An agent is used to analyze the packets received, calculate the holding time, and then forward the packet. The agent is also able to overhear the packet forwards from the neighboring nodes. The paper is organized as follows: Sect. 1.2 covering the literature regrading holding time and Sect. 1.3 is about hold time computation. Design and implementation of the framework in Sects. 1.5 and 1.4 contains the setup used for the implementation, and Sect. 1.6 is the analysis of the results and conclusion in Sect. 1.7.

1.2 Related Work Coordination [9] is carried out using various methods, and the most common and popular is the use of holding time. The holding time coordinates the node communication by allowing only one node to forward the packet when other nodes are in a hold state. The holding time calculation is different for each different routing protocol, the distance between the nodes being one of the most commonly used factors. The holding time is different for each node based on the priority of the node in the forwarding nodes collection [10]. RTS-CTS is one of the fundamental technologies used for collision avoidance and traffic control [11]. This method can be employed in underwater networks. The request RTS is sent by node and will wait for CTS from the receiver. The method’s efficiency is determined by various elements such as attenuation, ambient noise, spreading loss, and communication parameters. Yan et al. [12] discuss one of the first approaches to using holding time in depthbased protocol for UASNs. Nodes at different depths use different holding times. The idea of holding time was introduced to avoid conflicting packets and also helps in packet delivery ratio [12]. The depth difference between the current node and the previous-hop node was used to calculate the holding time initially. The primary

1 Implementing Holding Time Based Data Forwarding in Underwater …

3

conditions of holding time were for a node to be able to overhear any neighboring successful transmission. As energy is not a constraint in calculation, low-energy nodes can be problem for transmission. The forwarding time in Stateless Opportunistic Routing Protocol [13] is calculated based on the depth difference between the previous hop and the next hop from the forwarding node. The holding time should decrease as the depth difference from the current node increases, and it should be long enough for other nodes to notice a successful transmission.

1.3 Hold Time Computation In this paper, we implement the hold time equation used in the Hydrocast [14] routing protocol to verify the framework. When a node receives a packet, the hold time for the node is determined using two-hop connectivity and the distance between the nodes. The holding time equations are defined based on Fig. 1.1. There are three nodes defined, namely i, j, and c, with i and j being the potential forwarding nodes and c being the sender node. The hold time is linearly related to distance between the sender node and current node. The equation for holding time is given as follows: f (dxP ) = α(R − dxP )

(1.1)

where R is the transmission range and dxP is the progress toward the surface for given node x. Ensuring low hold time for high-priority nodes is one of the main objectives in Eq. (1.1). The α is responsible for the nodes to be able to overhear the transmissions from the neighboring nodes. The α is given as α>

Fig. 1.1 Nodes and notation for hold time computation [14]

tci − tcj + ti j + tack diP − diP

(1.2)

4

K. Bhujange et al.

where tack is acknowledgment transmission delay and the remaining part of Eq. 1.2 is for the propagation delay over distance dic − d jc + di j , where i, j, and c are the nodes. As the alpha value does not have any limit, a maximum allowable hop delay γ is defined. For experimental setup, the tack value is assumed to be 0.1 s and the propagation delays are set to 0.2 s.

1.4 Implementation of Hold Time and Packet Forwarding The framework is implemented in UnetStack3 [8], a modeling tool for underwater network communication. We employed the physical layer [15] to send data for our algorithm. If we transmit the packet directly using the routing table, it is difficult to keep the packet at a certain node; hence, the physical layer is used. The datagram request is used to transmit packets of data. The packet is adjusted so that it may be used by the algorithm. The packet can be of different sizes n based on user requirements. Generally, the packet is 32 or 64 bytes. The basic structure of the packet is shown in Fig. 1.2. The first byte of the packet is used for the packet’s unique identifier (UID), the next three bytes are used for the sorted cluster array, and the remaining bytes are used for the data to be delivered. The data is broadcast at each node as given below: phy Q critical for each paired treatment. It reveals that the CE-CMAES method differs significantly from the other five strategies in each trial. As a result, it is apparent that CE-CMAES delivers the greatest methods and competitive results when compared to all examined procedures. Furthermore, Wilcoxon test is conducted with a significance level of 0.05 between the two algorithms. The null hypothesis is rejected in all situations since the p-value is less than H0, and CE-CMAES is statistically different from that of other algorithms, which is denoted as ‘Yes’ in Table 9.3.

9.5 Conclusions This paper offered a new optimization strategy for solving the OPF problem that takes renewable energy sources into account (RESs). The goal of this project is to lower the fuel costs of a modified IEEE 57 bus system using RESs. To solve the OPF effectively, the proposed technique, CE-CMAES, employs a Cross Entropy and covariance matrix adaption evolutionary algorithm method. The suggested method’s

9 Optimal Power Flow Considering Uncertainty of Renewable Energy …

103

results are compared to those of other optimization methods including CEEPSO, Levy DEEPSO, and VNS. In comparison with the other approaches, the CE-CMAES obtained the lowest fitness. Statistical analyses using Wilcoxon signed rank, Tukey HSD, and ANOVA show that the mean of CE-CMAES is distinct from the CEEPSO, Levy DEEPSO, and VNS, indicating that it is valid. When compared to other metaheuristic methods, the proposed method delivers superior solutions and hence has the ability to solve the OPF problem.

References 1. Bai, W., Abedi, M., Lee, K.: Distributed generation system control strategies with PV and fuel cell in microgrid operation. Control. Eng. Pract. 53, 184–193 (2016) 2. Gope, S.: Dynamic optimal power flow with the presence of wind farm. Lambert Academic Publishing (2012) 3. Chaib, A., Bouchekara, H., Mehasni, R., Abido, M.: Optimal power flow with emission and non-smooth cost functions using backtracking search optimization algorithm. Int. J. Electr. Power Energy Syst. 81, 64–77 (2016) 4. Lim, S., Montakhab, M., Nouri, H.: Economic dispatch of power system using particle swarm optimization with constriction factor. Int. J. Innovations Energy Syst. Power 4(2), 29–34 (2009) 5. Liu, Y., Benoit, L., Ye, I., Liu, X., Cheng, Y., Morel, G., Fu, C.: Economic performance evaluation method for hydroelectric generating units. Energy Conserv. Manage. 44(6), 797–808 (2003) 6. Mahaboob, S., Suresh, V., Sivanagaraju, S.: Optimal power flow solution in the presence of renewable energy sources. Iran. J. Sci. Technol. Trans. Electr. Eng. 45(11), 61–79 (2020) 7. Moradi, H., Esfahanian, M., Abtahi, A., Zilouchian, A.: Optimization and energy management of a standalone hybrid microgrid in the presence of battery storage system. Energy 147, 226–238 (2018) 8. Etta, G., Miguel, H., Salman, M., Goncalo, C., Robin G.: A stochastic optimal power flow for scheduling flexible resources in microgrids operation. Appl. Energy 201–208 (2018) 9. Reddy, S., Vuddanti, S., Jung, C.: Review of stochastic optimization methods for smart grid. Front Energy 2, 197–209 (2016) 10. Surender, S., Bijwe, P. and Abhyankar, A.: Real-Time economic dispatch considering renewable power generation variability and uncertainty over scheduling period. IEEE Syst. J. 9(4), 1440– 1451 (2015) 11. Cabus, P.: River flow prediction through rainfall-runoff modelling with a probability-distributed model (PDM) in Flanders, Belgium. Agric. Water Manage. 95(7), 859–868 (2008) 12. Sarda, J., Pandya, K.: Optimal active–reactive power dispatch considering stochastic behavior of wind, solar and small-hydro generation. In: Applications of Artificial Intelligence Techniques in Engineering. Advances in Intelligent Systems and Computing, pp. 255–263 (2019) 13. Wu, Y., Debs, S., Marsten, R.: A direct nonlinear predictor corrector primal dual interior point algorithm for optima power flows. IEEE Trans. Power Syst. 9(May), 876–883 (1994) 14. Lu, C., Hsu, Y.: Reactive power/voltage control in a distribution substation using dynamic programming. IEE Proc.-Gener. Transm. Distrib. 142(6), 639–645 (1995) 15. Lee, K., Yang, F.: Optimal reactive power planning using evolutionary algorithms: a comparative study for evolutionary programming, evolutionary strategies, genetic algorithms and linear programming. IEEE Trans. Power Syst. 13(1), 101–108 (1998) 16. Fortenbacher, P., Demiray, T.: Linear/quadratic programming-based optimal power flow using linear power flow and absolute loss approximations. Int. J. Electr. Power Energy Syst. 107, 680–689 (2019)

104

J. Sarda et al.

17. Sarda, J., Pandya, K., Lee, K.Y.: Hybrid cross entropy—cuckoo search algorithm for solving optimal power flow with renewable generators and controllable loads. Optimal Control Appl. Method 1–25 (2021) 18. Sarda, J., Pandya, K., Lee, K.Y.: Dynamic optimal power flow with cross entropy covariance matrix adaption evolutionary strategy for systems with electric vehicles and renewable generators. Int. J. Energy Res. 45, 10869–10881 (2021) 19. Rueda, J., Lee, K.Y., Erlich, I.: Competition & panel: evaluating the performance of modern heuristic optimizers on smart grid operation problems. https://site.ieee.org/psace-mho/2017smart-grid-operation-problems-competition-panel/ (2017) 20. Rubinstein, R., Krose, D.: The Cross-Entropy Method. A Unified Approach to Combinatorial Optimization, Monte- Carlo Simulation, and Machine Learning. Springer, New York (2004) 21. Hansen, N.: The CMA evolution strategy: a comparing review. In: Towards a New Evolutionary Computation, Advances on Estimation of Distribution Algorithms, vol. 192, pp.1769–1776. Springer (2006) 22. Carvalho, L., Miranda, V., Silva, A., Marcelino, C., Wanner, E., Sumaili, J.: Cross entropy method and evolutionary particle swarm optimization. In: IEEE 2017 PES General Meeting. https://site.ieee.org/psace-mho/2017-smart-grid-operation-problems-competition-panel/ 23. Abaci, K., Yamacli, V.: Differential search algorithm for solving multi-objective optimal power flow problem. Int. J. Electr. Power Energy Syst. 79, 1–10 (2016)

Chapter 10

A Triple Band High Gain Antenna Using Metamaterial Souvik Halder , Abhijyoti Ghosh , Sudipta Chattopadhyay , and L. Lolit Kumar Singh

Abstract In this paper, a metamaterial-based triple band high gain antenna is proposed. A simple patch antenna is modified into a triple band antenna using metasurface and superstrate. Metasurface and metamaterial superstrate used two different split ring resonator (SRR)-based unit cells. The metasurface is made of dual-layer symmetry single ring resonator pair (D-SSRRP) unit cells, and the superstrate consists of BSRR type unit cells. This novel structure makes a simple patch to radiate in multiband without altering the geometry of the main radiating patch or ground. This structure also increases the gain by 4.5 dBi. The resonating bands are 5.82, 5.97, and 6.35 GHz. The proposed antenna has a volume is about 0.07λ3 .

10.1 Introduction In this era of miniaturization, communications systems are also getting smaller, compact, and more affordable day by day, which makes the microstrip antenna (MA) most preferable because of its low-profile structure and low fabrication cost. The main disadvantages of MA are low gain, narrow bandwidth, etc., in short, low radiation efficiency [1]. On the other hand, present communication devices use multiple technologies, and most of them operate different frequency bands, so multiband antennas are preferable. The field of research on microstrip antenna design has become keener toward designing multiband antenna with improved radiation efficiency. There are several techniques to increase the gain of a patch antenna. The most popular technique is using metamaterial structure within the antenna environment [2–13] like metasurface or metamaterial superstrate, antenna load, or in the antenna plane. Metamaterial superstrate and metasurface are very popular for gain enhancement, as Gao et al. [3] used two different types of D-SSRRP type metasurface with MNG property for gain enhancement of a dual band antenna; it enhances the gain by 2dBi. Samantaray et al. [4] also used metasurface in the ground plane with partial ground on S. Halder (B) · A. Ghosh · S. Chattopadhyay · L. Lolit Kumar Singh Department of Electronics and Communication Engineering, Mizoram University, Aizawl 796004, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_10

105

106

S. Halder et al.

a dual band slotted antenna, which enhances the gain by 3.5 dBi gain and bandwidth by 150%. In [5], zero index metamaterial (ZIM) as superstrate improved gain of single band antenna by 5 dBi. In [6], 3.7dBi, gain improvement was realized by using used SRR-based superstrate on a single band antenna array. In [7], a modified split ring resonator (MDSRR) superstrate improves the gain of a single band antenna by 5 dBi. A metamaterial superstrate had used to enhance the gain of a dual band slotted antenna, where the gain improved by 4.5 dBi [8]. In [9], gain improved by 4.45 dBi by using three no of superstrate. Metamaterials are used to achieve unusual properties of RF devices by introducing the effect of negative permittivity and permeability called the double negative (DNG) property [10]. Material with a negative value of either permittivity or permeability is not available in nature. It can be realized artificially by engineering, the inclusion of an array of metallic structures in a dielectric or magnetic material. Then, those will exhibit unusual properties within a particular frequency range. Split ring resonator (SRR) is the oldest and most widely used metamaterial structure [2, 11]. The SRR [10, 11] acts as a parallel LC resonating circuit, and it provides low radiation loss, high-quality factors, and strong magnetic coupling. The SRRs are primarily two types edge couple and broadside coupled [11]. The broadside coupled SRR has some advantages over edge coupled SRR; those are isotropy in the plane and much smaller electrical size. This paper presents a prototype of a high gain multiband antenna. Two different SRR-based metamaterials structures improve the gain and make a simple patch antenna resonant in three different frequency bands. One is D-SSRRP [5] used in the antenna plane as metasurface. The D-SSRRP unit cell consists of two no of SRR pair, each unit of SRR is edge coupled, and each pair of SRR is broadside coupled. The metamaterial superstrate consists of simple broadside coupled SRRs (BSRR).

10.2 Structural Evolution Initially, the design was started with a simple microstrip antenna, antenna-1, designed for the resonating frequency of 5.9 GHz using a 1.6 mm thick FR4 substrate with coaxial feed. The dimension of the radiating patch was 10.7 × 14.4 mm2 , as shown in Fig. 10.1. The antenna provided 300 MHz bandwidth, from 5.76 to 6.06 GHz, with 3.37 dBi gain. In antenna-2, a metamaterial superstrate is introduced in antenna-1, shown in Fig. 10.2, to improve the antenna efficiency. The metamaterial type was chosen carefully. Here, broadside-SRR (BSRR) type metamaterial has been used for the metamaterial superstrate. The BSRR consists of two same-size square-shaped split ring resonators, placed on top and bottom of the substrate where split rings are placed opposite each other, as shown in Fig. 10.3a. The unit cell of BSRR MTM has been designed to get negative permittivity (ε) and permeability (μ) below 6 GHz. The optimized length of each side of BSSR is 7.4 mm, designed on a 1.6 mm thick FR4 substrate and the length

10 A Triple Band High Gain Antenna Using Metamaterial

107

SW1

SW1

Ground plane

L

SL1

SL1

W

Feeding Point

Patch

(a)

(b)

Fig. 10.1 Antenna-1, conventional MA resonating at 5.9 GHz a top, b bottom view. L = 10.7 mm, W = 14.4 mm, SL1 = 20.3 mm, SW2 = 24 mm

SW2

SW2

W SL2

SL2

L

GL

GW

(a)

(b)

SW2

d1

gap SL2

(c)

(d)

Fig. 10.2 Antenna-2, antenna using BBSR metamaterial superstrate resonating at 5.9 GHz a top, b bottom view of main antenna, c top view of metamaterial superstrate, d side view. L = 10.7 mm, W = 14.4 mm, GL = 20.3 mm, GW = 24 mm, SL2 = SW2 = 28 mm, d 1 = 1.6 mm, gap = 3.4 mm

108

S. Halder et al.

Fig. 10.3 BSRR unit cell, a 3D view and it’s b frequency versus permittivity (ε) and permeability (μ) plot (design parameters: m = 7.4 mm, c1 = 1 mm)

of split is 1 mm; dimensions are shown in Fig. 10.3b. The BSRR unit cell operates in DNG mode from 1 to 7.5 GHz. The frequency versus real part of permittivity and permeability is shown in Fig. 10.3c. The relative permittivity (ε) and permeability (μ) have been derived from the S11 and S21 parameter of the unit cell using the NRW method. The metamaterial superstrate consisting of a 3 × 3 array of proposed BSRR unit cells is placed with a gap of 1.6 mm from each other, as shown in Fig. 10.2c. The area of the antenna substrate is increased to 28 × 28 mm2 to match the size of the superstrate. The best result obtains when the gap between two substrates is 3.4 mm. It improves the antenna gain by 2.19 dBi and bandwidth by 60 MHz. This design provides 5.56 dBi gain without affecting the resonant frequency (5.9 GHz) and provides 360 MHz bandwidth from 5.76 to 6.12 GHz. In this model, the S11 value at resonating frequency is − 19.12 dB, lower than the conventional one, but it is still in the excellent range (Fig. 10.4). In the third model, antenna-3, a dual-layer symmetry single ring resonator pair (D-SSRRP) is introduced in the design of antenna-2. The unit cell of D-SSRRP consists of two identical structures placed on both sides of the substrate. Each side consists of two identical SRRs placed side by side where splits are located in the center of the SRR facing each other like a mirror image, shown in Fig. 10.5a, b. The D-SSRRP is also designed for resonant frequency 6 GHz. The D-SSRRP was designed on an FR4 substrate. Post optimization, the sides of SRR become 4 mm long, and the split length is 1 mm long. The D-SSRRP unit cell operates in DNG mode from 1 to 7.9 GHz, as shown in Fig. 10.5c. The relative permittivity (ε) and permeability (μ) have been derived using the same procedure used in BSRR. In antenna-3, the array of 5 × 2 unit cells is placed on both radiating sides and an array of 2 × 2 unit cells on the non-radiating side of the patch antenna. After placing the metamaterial structures around the ground and patch, without affecting any, the total area of substrate became 38 × 43mm2 . The gap between the primary

10 A Triple Band High Gain Antenna Using Metamaterial

109

5 0

S11 (dB)

-5 -10

-10dB indicator

-15

Conv BSRR

-20 -25 4

5

6

7

8

Frequency (GHz)

Fig. 10.4 Comparison of frequency versus return loss between conventional and with BSRR model

Fig. 10.5 D-SSRRP unit cell a 3D view and it’s b frequency versus permittivity (ε) and permeability (μ) plot (design parameters: a = b = 4 mm, c = 1 mm, d = 0.2 mm)

antenna substrate and the metamaterial superstrate is kept at 3.4 mm, as shown in Fig. 10.6d. The schematic diagram of this model is shown in Fig. 10.6. This simulation resonates in three different frequencies those are 5.82, 5.97, and 6.35 GHz, as shown in Fig. 10.7. The maximum gain is observed 7.96 dBi at 5.82 GHz.

10.3 Simulated Results and Discussion In the previous section, three simulation schematics have been described. Antenna-1 is the conventional one, and the rest are used MTM for performance enhancement.

110

S. Halder et al.

SW3 W SL3

GW

L

GL Patch Ground Plane

(a)

(b)

SW3

g1 g2

SL3

gap d2

(c)

(d)

Fig. 10.6 Schematics of proposed antenna-3, a BSRR-based MTM superstrate, b top and c bottom view, d side view of D-SSRRP-loaded MTM antenna. SL3 = 37 mm, SW3 = 43 mm, W = 14.4 mm, L = 10.7 mm, GL = 20.3 mm, GW = 24 mm, g1 = 5.8 mm, g2 = 9.8 mm, d 1 = 0.6 mm, gap = 3.4 mm 0

S11 (dB)

-5

-10

-10dB indicator

-15

Antenna-3 Antenna-2 Antenna-1

-20

-25 5

6

7

Frequency (GHz)

Fig. 10.7 Comparison of S 11 parameters between antenna-1, antenna-2, and antenna-3

10 A Triple Band High Gain Antenna Using Metamaterial Fig. 10.8 Reflection coefficient profile of proposed antenna-3

111

0

S11 (dB)

-5

-10dB indicator -10

-15

Band-2 Band-3

Band-1 5.0

5.5

6.0

6.5

7.0

Frequency (GHz)

They enhance the gain against the conventional antenna without modifying the main radiating patch and ground plane. In antenna-2, where MTM superstrate is used, observed gain enhancement of 2.19 dBi. The final prototype, i.e., antenna-3, uses both metamaterial structures, BSRR and D-SSRRP. The maximum 7.96 dBi gain was achieved. The total gain improvement of 4.59 dBi is achieved compared to the conventional antenna. The bandwidth of antenna-2 increases by 60 MHz in comparison with the conventional one. The proposed antenna, i.e., antenna-3, resonates in three different frequencies those are 5.82 GHz, 5.97 GHz, and 6.35 GHz bandwidths are 220 MHz, 40 MHz, and 140 MHz, respectively, shown in Figs. 10.8 and 10.9. The performance comparisons between the three antenna models are listed in Table 10.1. The maximum gain obtains in antenna-3 is 7.96dBi in band-1. Gain observed in band-2 and band-3 is 6.60 dBi and 6.90 dBi, respectively. The half power beam width of the conventional antenna (antenna-1) is 88.47°, and CP-XP isolation is 23.24 dB. In the proposed model, half power beam widths are 59.14°, 55.34°, and 55.5° for band-1, band-2, and band-3, respectively. The maximum CPXP isolation is 23.15 dB observed for the band-1. In Table 10.2, the comparison of antenna volume and gain between the proposed antenna and the existing literature is enlisted. The antenna volume is lower than the other models used metamaterial superstrate in recent literature, which is 0.07λ3 .

10.4 Conclusions In this paper, an SRR-based multiband high gain antenna is proposed. The proposed metamaterial superstrate enhances the gain by 2.19 dBi. After using D-SSRRP in the antenna plane and BSRR superstrate, the antenna gain is increased by 4.59 dBi

S. Halder et al. Frequency- 5.85GHz HCP ECP

Gain (dBi)

10 8 6 4 2 0 -2 -4 -6 -8 -10 -12 -14 -16 -18 -20 -150 -120 -90 -60 -30

0

30

60

90

120 150

14 12 10 8 6 4 2 0 -2 -4 -6 -8 -10 -12 -14 -16 -18 -20 -150 -120 -90 -60 -30

Frequency- 5.97GHz HCP ECP

0

30

Theta (deg)

Theta (deg)

(a)

(b)

Gain (dBi)

Gain (dBi)

112

10 8 6 4 2 0 -2 -4 -6 -8 -10 -12 -14 -16 -18 -20 -150 -120 -90 -60 -30

60

90

120 150

Frequency- 6.35GHz HCP ECP

0

30

60

90

120 150

Theta (deg)

(c) Fig. 10.9 Gain at resonating frequencies obtain for proposed model

Table 10.1 Comparison of performance between three simulated antenna models

Antenna model

Resonance frequency (GHz)

Bandwidth (MHz)

Gain (dBi)

Antenna-1

5.9

300

3.37

Antenna-2

5.9

360

6.56

Antenna-3

5.82, 5.97, 6.35

220, 40, 140

7.96, 6.60, 6.90

compared to the conventional one. The triple band property makes the antenna valuable for the system operating in multiple frequencies. There is a slight shift observed for primary frequency because of the loading effect of the metamaterial. The overall antenna volume is increased compared to the conventional antenna. The proposed antenna volume is 0.07 λ3 and it is still smaller than the other superstrate antenna

10 A Triple Band High Gain Antenna Using Metamaterial

113

Table 10.2 Performance comparison between the proposed antenna and existing literature Refs. No.

Antenna volume (λ3 )

No. of band

Gain enhancement (dBi)

Gain enhancement method

[5]

2.09

1

5.0

ZIM superstrate

[6]

0.2

1

3.7

Metamaterial superstrate

[7]

2.129

1

2.1

Metamaterial superstrate

[8]

0.23

2

4.5

Metamaterial superstrate

[9]

1.15

3

4.45

Metamaterial superstrate

This work

0.07

3

4.59

Metasurface + metamaterial superstrate

discussed in the literature. The antenna can be used for un-license band RF communication as it covers the entire 5.7 and 5.8 GHz un-license band radio (UBR) spectrum available in India.

References 1. Garg, R., Bhartia, P., Bhal, I., Ittipiboon, A.: Microstrip Antenna Design Handbook. Artech House Antenna and Propagation Library (2001) 2. Ziolkowski, R.W., Kipple, A.D.: Application of double negative materials to increase the power radiated by electrically small antennas. IEEE Trans. Antennas Propag. 51(10), 2626–2640 (2003) 3. Gao, X.J., Cai, T., Zhu, L.: Enhancement of gain and directivity for microstrip antenna using negative permeability metamaterial. AEU Int. J. Electron. Commun 70(7), 880–885 (2016). https://doi.org/10.1016/j.aeue.2016.03.019 4. Samantaray, D., Bhattacharyya, S., Srinivas, K.V.: A modified fractal-shaped slotted patch antenna with defected ground using metasurface for dual band applications. Int. J. RF Microw. Comput. Aided Eng. 29, e21932 (2019).https://doi.org/10.1002/mmce.21932 5. Rajanna, P.K.T., Rudramuni, K., Kandasamy, K.: A High-gain circularly polarized antenna using zero-index metamaterial. IEEE Antennas Wirel. Propag. Lett. 18(6), 1129–1133 (2019) 6. Arora, C., Pattnaik, S.S., Baral, R.N.: SRR Superstrate for Gain and Bandwidth Enhancement of Microstrip Patch Antenna Array. Prog. Electromagn. Res. B 76, 73–85 (2017) 7. Pattar, D., Dongaokar, P., Nisha, S.L., Amith, S.: Design and implementation of metamaterial based patch antenna. In: IEEE International Conference for Innovation in Technology (INOCON) (2020) 8. Samantaray, D., Bhattacharyya, S.: A gain-enhanced slotted patch antenna using metasurface as superstrate configuration. IEEE Trans. Antennas Propag. 1–1 (2020) 9. Sumathi, K., Lavadiya, S., Yin, P., Parmar, J., Patel, S.K.: High gain multiband and frequency reconfigurable metamaterial superstrate microstrip patch antenna for C/X/Ku-band wireless network applications. Wireless Netw. 27(3), 2131–2146 (2021)

114

S. Halder et al.

10. Nutan, R.A., Raghavan, S.: Split ring resonator and its evolved structures over the past decade. In: IEEE International Conference ON Emerging Trends in Computing, Communication and Nanotechnology (ICECCN-2013), pp. 625–629 (2013) 11. Marques, R., Mesa, F., Martel, J., Medina, F.: Comparative analysis of edge- and broadsidecoupled split ring resonators for metamaterial design—theory and experiments. IEEE Trans. Antennas Propag. 51(10), 2572–2581 (2003) 12. Zhu, H.L., Liu, X.H., Cheung, S.W., Yuk, T.I.: Frequency-reconfigurable antenna using metasurface. IEEE Trans. Antennas Propag. 62(1), 80–85 (2014) 13. Bakhtiari, A., Sadeghzadeh, R.A., Moghadasi, M.N.: Gain enhanced miniaturized microstrip patch antenna using metamaterial superstrates. IETE J. Res. 1–6. (2018)

Chapter 11

Physical Layer Security in MIMO Relay-Based Cognitive Radio Network V. Upadhyay and M. R. Bharti

Abstract Cognitive radio networks are one of the predominant technologies to solve the spectrum scarcity problem. In this paper we tackle the problem of secure and safe data transmission between secondary user transmitter and secondary user receiver through relays in the presence of eavesdroppers in cognitive radio-based MIMO system. MIMO stands for Multiple-Input Multiple-Output. MIMO is a wireless technology that increases the data capacity with the help of multiple transmitting and receiving antennas. In this work, the best amplify and forward relay to maximize the secrecy capacity in a MIMO system is selected. The simulation and analytical results show the improved performance through secrecy capacity and secrecy outage probability.

11.1 Introduction With advancing technology, there has been an unparalleled rise in usage of the wireless devices. With increasing wireless devices, it has been seen that the demand for frequency spectrum has increased as well. Now the cognitive radio (CR) network comes into play, which solves the problem of frequency spectrum scarcity by improving the spectrum utilization of wireless networks [1]. In a CR network, secondary users (SUs) are allowed to dynamically transmit together with primary users (PUs) on the same spectrum band that has been licensed to the PUs by the government agency such as Federal Communication Commission (FCC) while keeping the total interference power at the PU receiver below the interference temperature limit [1, 2]. In CR networks, the physical layer performs numerous functions like coding of message signals, transmission, and modulation. It was further improved to support spectrum sensing to dynamically access the spectrum [3]. Hence, the security of the CR network is one of the important aspects to ensure secure communication. V. Upadhyay (B) · M. R. Bharti National Institute of Technology, Hamirpur 177005, HP, India e-mail: [email protected] M. R. Bharti e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_11

115

116

V. Upadhyay and M. R. Bharti

The traditional method of security involves cryptographic approaches, which can be further divided into different types of protocols to prohibit the eavesdroppers from decoding the message signal sent from the transmitter, but the distribution of cryptographic keys makes it a difficult process [1]. With the arrival of the mobile ad hoc Network the nodes in the network become more vulnerable to get attacked. Some of the major challenges were in the power limitation and memory terminals which in the absence of centralized hardware deepens the security related issue [1]. The data reaches destination in the form of packets by traveling from different nodes distributed in the network. Nodes in the network can be trusted or untrusted. The trusted nodes follow the proper protocols of sensing functions, while the untrusted nodes operate maliciously, violating the network security requirements. The aim of the set of malicious users (nodes) is to prevent licensed SUs from using the spectrum band, thus disrupting the proper operation of the network and exploiting the vulnerabilities of the physical layer [3]. The other significant problem in the wireless communication comes in the form of eavesdropper attack which access or listen to the information message sent by the SU in the physical or network layer. The eavesdropper can be of both active and passive types. In the active type, eavesdroppers react and reply after overhearing the information, thus trying to establish a connection with the PU. For the passive type eavesdropper, the channel information of the eavesdropper is unknown, and it silently spies on the wireless transmission for the sake of confidential information. The passive eavesdroppers, can only read or listen using passive receivers, but when active, it can inject malicious signal into the traffic. The eavesdropper can switch between active and passive, thus making it difficult to detect easily [3, 4]. The detection of the eavesdropper can be done according to a number of parameters like secrecy outage probability (SOP) and secrecy capacity. In this paper, we have used both of these parameters for enhancing the physical layer security with the help of amplify and forward (AF) relay in the network [3]. In the CR network, the interference is generated at the PUs, to counter this issue MIMO technique is used. In MIMO technique, there are multiple antennas at the transmitter and receiver. In MIMO, we are trying to exploit the spatial multiplexing technique. In spatial multiplexing the large information is split up and sent through different antennas at same time using same frequency spectrum. MIMO relaying techniques help to enhance reliability, improve throughput, improve coverage and most importantly, they significantly reduce the transmit power level of SU transmitters [5, 6]. MIMO technique also helps to improve the energy efficiency and spectral efficiency of the system [7]. The above facts regarding the MIMO relay network motivate us to analyze the secrecy rate by including it in our network model, in which the same frequency spectrum is used by PUs and SUs. Also in the CR network, relay selection is one of the important issues for improving the secrecy parameter. The relay may use either the decode-and-forward (DF) relaying strategy or an amplify-and-forward (AF) relaying strategy [8]. The DF relay decodes and retransmits the incoming message signal from transmitter. The DF relay, hence, can vary the communication rate between source to relay and relay to destination [9]. The AF relay simply amplifies the received message signal at the relay

11 Physical Layer Security in MIMO Relay …

117

and transmits it to the destination. This makes the implementation of the AF relay in the CR network simpler [10]. The CR network, which we have considered in our paper, consists of a PU, a SU transmitter and a SU receiver that can communicate through a set of relays. In this paper, the optimal relay selection in MIMO relay network has been taken into consideration with the objective of maximizing secrecy capacity. The rest of the sections have been organized in the paper as follows. In Sect. 11.2, the system model has been explained. In Sect. 11.3, the optimal relay selection scheme and problem formulation have been given, while in Sect. 11.4, the simulation results have been presented. Finally in the last Section, conclusion of the work has been mentioned.

11.2 System Model In this work, we have considered a CR network model as presented in Fig. 11.1, consisting of a PU, an eavesdropper, a set of N relays R1 , R2 , R3 , ..., R N and SU transmitter and receiver node. The communication is done through a single hop relay in a MIMO-based approach by the SU. The eavesdropper in the network tries to maliciously hear the SU source information. Using the above described model, various parameters have been analyzed for direct transmission (DT), conventional selection (CS) and optimal selection (OS) cases. Further, the number of antennas in the MIMO system has also been varied to analyze our parameters variation, i.e., secrecy capacity and secrecy outage probability, toward a practical MIMO system. The channels between SU transmitter to relay, relay to SU receiver, and SU transmitter to SU receiver are assumed to undergo flat fading.

11.2.1 Direct Transmission (DT) In direct transmission, the MIMO-based SU transmitter antenna sends data directly to the SU receiver antenna without transmitting it through a single hop relay in the network. The received signal at the SU receiver by direct end-to-end transmission is given as  yd = Ps h SD x + Z d . (11.1) Here, x is the transmitted symbol from the SU transmitter, and Ps is the average transmit power per symbol from the SU transmitter. Z d is the complex Gaussian noise at the SU receiver and h SD is the channel gain between the transmitter and receiver.

118

V. Upadhyay and M. R. Bharti

Fig. 11.1 MIMO relay based CR network.

Similarly, the received message signal at the eavesdropper is  ye = Ps h SE x + Z e ,

(11.2)

where, h SE is the channel gain between the SU transmitter and eavesdropper and Z e denotes the complex Gaussian noise at the eavesdropper. The received message signal at the PU is  y p = Ps h SP x + Z p . (11.3) where, h SP is the channel gain between the SU transmitter and PU and Z p denotes the complex Gaussian noise at the PU.

11.2.2 Amplify and Forward (AF) Relaying In this paper, a set of N, AF relays has been used for two stages of message transmission. In the first stage, the message from one of the MIMO transmitter antennas of the SU is sent to all the trusted relays in the CR network. We consider that each relay node also consists of N number of receiving antennas. The role of relay nodes is to amplify and forward the transmitted message to one of the antennas of the SU receiver. In the second stage, from the set of trusted relay nodes, the relay node, which gives maximum secrecy capacity within the interference power constraints at the PU is allowed to transmit the amplified message signal.

11 Physical Layer Security in MIMO Relay …

119

The messages received at one of the trusted relay nodes during the first stage of transmission is  yr = Ps h SR x + Z r . (11.4) Here, yr is a N × N matrix which represents all the received signals from all the different antennas of the SU. h SR is N × N matrix which denotes the channel gain between all the different antennas of the SU transmitter to all the trusted relay nodes. Z r is also N × N matrix that represents complex Gaussian noise at all relay nodes. During the second stage of transmission, the relay node, which maximizes the secrecy capacity of the message signal coming from the SU transmitter is allowed to transmit the amplified signal. The received message signals, from one of the best selected relay node for maximizing secrecy capacity of SU, at PU, SU receiver and eavesdropper are given as  y p = Ps h RP x + Z p , (11.5) yd = ye =

 

Ps h RD x + Z d ,

(11.6)

Ps h RE x + Z e .

(11.7)

Here, h RD , is the channel gain from one of the trusted relay nodes aimed at maximizing the secrecy capacity to the SU receiver, while h RE and h RP are channel gains from the same relay node to eavesdropper and PU, respectively.

11.2.3 Channel Representation In the above mentioned system model, the assumption made for the distribution of channel coefficients between two nodes is that they are zero mean, independent Gaussian random variables with variance σi,2 j , h i, j ∼ N (μ, σ 2 ). The signal, as it propagates in the channel, decreases with di,−αj factor, where α is the path loss exponent. The path loss exponent is the parameter which indicates the decrease in received signal strength with the increase in distance. The value of the path loss exponent depends on the specific surrounding or propagation environment.

11.3 Relay Selection Scheme and Problem Formulation With the appropriate selection of relay, the secrecy capacity can be maximized and, thereby, the physical layer security will be enhanced. The calculation of the instan-

120

V. Upadhyay and M. R. Bharti

taneous secrecy rate for a particular message at each of the trusted relay nodes in the network model shown in Fig. 11.1, can be done by 1 Cs (R) = { [log2 (1 + γSD + γRD ) − log2 (1 + γSE + γRE )]}. 2

(11.8)

Now the secrecy rate will be compared among different trusted relay nodes. The node having the maximum secrecy rate will be allowed to transmit an amplified message signal to the desired SU receiver. Here, γSD denotes the instantaneous signal-to-noise ratio (SNR) for the SU transmitter and the SU receiver link. γRD denotes the instantaneous SNR for relay to the SU receiver link. γSE denotes the instantaneous SNR for the SU transmitter to eavesdropper link. γRE denotes instantaneous SNR for relay to the eavesdropper link. The secrecy capacity can be expressed mathematically in a more simplified way as below. Cs (R) =

1 (1 + γSD + γRD ) log2 2 (1 + γSE + γRE )

(11.9)

The other parameter for detection of eavesdroppers in the network is based on the SOP. The SOP for traditional networks is usually defined as the probability that the secrecy rate is less than the target secrecy rate Rs , (Rs > 0). For CR networks, SOP is defined using its two components. The First component is the probability that secrecy rate is less than the target secrecy rate when the interference power at the PU is less than the interference temperature limit, the second component is the probability that the interference power at the PU is greater than the interference temperature limit. Thus, the secrecy outage probability for CR networks can be mathematically expressed as

PSOP =

 N



Pr {Csn

< (Rs )} ×

Pr (I PR

  R ≤ ) + [Pr (I P ≥ )]

(11.10)

n=1

I PR : Represents interference power at the PU.  : Represents the interference temperature limit. Csn : Represents secrecy capacity through the nth AF relay node. The interference power consists of the SU transmitter signal, noise, and relay signal at the PU. Therefore, while maximizing the secrecy rate, the interference power constraints at PU can also be taken into consideration. In this case, the problem becomes as Rmax = max(Cs (R)), (11.11) subject to I PR ≤ , which can be solved in a straightforward manner.

(11.12)

11 Physical Layer Security in MIMO Relay …

121

11.4 Simulation Results To obtain the average results, Monte Carlo simulations for 10000 independent trails have been performed. Apart from one SU transmitter with three antennas (MIMO) and one SU receiver with three antennas in the system, the other simulation parameters are given below. • • • • • • • • • • •

N = 3 relays has been considered for simulation. Each relay is having three receiving antennas and one transmitting antenna. The PU location is at (100, 0). The location of one eavesdropper is at (60, 0). The SU receiver is located at distance of (50, 0). The path loss exponent α = 3.5. Relay transmit power has been fixed to 15 W. The interference temperature limit has been set to  = −4 dB The target secrecy rate used to calculate the SOP is 0.35 bits/s/Hz. The transmission rate from SU transmitter is 2 bits/s/Hz. The three relays are located at distances (20, 0), (15 ,0), (25, 0), respectively, in the network.

The simulation result, analyzed in Fig. 11.2, is secrecy capacity versus transmit power of SU transmitter. In Fig. 11.3, SOP versus transmit power has been plotted, while in Fig. 11.4, the SU transmitter to eavesdropper distance has been varied and its impact on secrecy rate has been studied. All the simulation results are plotted for CS, OS, MIMO-CS, and MIMO-OS. The CS mainly comprises of transmitting of message signal from SU transmitter through any of the relay node irrespective of taking into consideration the presence of eavesdropper. The OS transmits the message through the particular relay node which is able to maximize the secrecy capacity by using the approach discussed already in this paper. In MIMO-CS, the antenna of the SU transmitter follows the same approach as in CS, just differing in the approach of sending messages through different transmitting antennas to any trusted relays in the network. The MIMO-OS follows the most enhanced approach of sending the messages. In this scheme, one of the antenna of MIMO sends the same message to every trusted relay node in a network, and then, by applying the algorithm (approach already discussed), it decides to send the message through one of the best selected relay node in the network in the second stage of transmission. In this work, MIMO-OS method has been analyzed for two cases, i.e., with two and three number of antennas. Figure 11.2 shows the secrecy capacity versus transmit power of the SU transmitter for different cases under consideration. It can be seen that the peak value of the secrecy capacity is attained at a significant power difference for the CS and the MIMO-OS. With the appropriate method of maximizing secrecy capacity, the improvement in secrecy capacity value for both MIMO-CS and MIMO-OS can also be seen in Fig. 11.2. In all cases, the secrecy capacity value first increases with an increase in transmit power, and then decreases due to an increase in interference power more than the

122

V. Upadhyay and M. R. Bharti

Fig. 11.2 Secrecy capacity versus transmit power of SU transmitter

Fig. 11.3 SOP versus transmit power of SU transmitter

interference temperature limit  at the PU. For MIMO system, less transmit power is used as compared to CS and secrecy capacity is also increased, which gives MIMO system a clear edge over CS for the purpose of enhancing the physical layer security of CR network. Figure 11.3 shows the SOP versus transmit power of the SU transmitter for different cases under consideration. It can be seen that for MIMO relay-based system, the

11 Physical Layer Security in MIMO Relay …

123

Fig. 11.4 Secrecy capacity versus SU transmitter to eavesdropper distance

SOP is lower as there are high chances of achieving the target secrecy rate, while in conventional system, the SOP is quite higher. The higher value of SOP signifies that either the system is unable to achieve the target secrecy rate or the interference power at the PU is greater than the interference temperature limit. With the increasing number of antennas in the MIMO relay-based system, the secrecy rate gets higher, which further increases the chance of achieving the target secrecy rate. In the considered system model, the target secrecy rate has been set to 0.35 bits/s/Hz for the calculation of SOP. The MIMO relay-based scheme shows the best while conventional system shows the worst SOP performance. Figure 11.4 shows the variation of secrecy capacity with the distance between the SU transmitter and the eavesdropper for different cases. The closer distance of SU transmitter from the eavesdropper leads to lower value of secrecy capacity. Hence, the closer distance increases the security vulnerabilities of the CR network even more. The relay nodes in the network are more vulnerable to the nearby eavesdroppers. Therefore the proper selection of relay node is one of the important aspects of the physical layer security. Due to this reason, the OS approach selects the relay node which can provide greater physical layer security. The MIMO-OS even gets a wider range of options for relay node selection for each of the SU transmitter antennas that are sending the message. In Fig. 11.4, the eavesdropper distance has been varied from (0, 5) to (0, 35) and it can be seen that as the distance increases, the value of secrecy capacity also increases up to a certain value before getting a flat curve since the eavesdropper position in the network after a certain distance becomes ineffective to the SU transmitter or relay.

124

V. Upadhyay and M. R. Bharti

11.5 Conclusion In this paper, the optimal relay selection for the MIMO relay-based system has been studied. The secrecy capacity and outage performance of CR networks has shown significant improvement. We also investigated our system performance against the varying positions of eavesdropper in the network with reference to the SU transmitter. In the proposed MIMO-OS, for each antenna of the SU transmitter, one relay is selected from the set of N relays to send a message to the SU receiver in the second stage. With the aim of improving the performance of the overall system, we have also taken care of the interference power constraint at the PU. For the above proposed approach, the simulation results show the improved performance for SU without effecting the performance of PU in the network. For future course of action, the work could be done in reducing the complexity of relay selection approach for the MIMO relay-based system. The work could also be done for improving the secrecy capacity when number of eavesdroppers in the network are increased.

References 1. Sakran, H., Shokair, M., Nasr, O., El-Rabaie, S., El-Azm, A.: Proposed relay selection scheme for physical layer security in cognitive radio networks. Iet Commun. 6(16), 2676–2687 (2012) 2. Haykin, S.: Cognitive radio: brain-empowered wireless communications. IEEE J. Sel. Areas Commun. 23(2), 201–220 (2005) 3. Salahdine, F., Kaabouch, N.: Security threats, detection, and countermeasures for physical layer in cognitive radio networks: a survey. Phys. Commun. 39, 101001 (2020) 4. Chorti, A., Perlaza, S.M., Han, Z., Poor, H.V.: Physical layer security in wireless networks with passive and active eavesdroppers. In: 2012 IEEE Global Communications Conference (GLOBECOM), pp. 4868–4873. IEEE (2012) 5. Amarasuriya, G., Li, Y.: Cognitive massive MIMO relay networks. In: 2017 IEEE Wireless Communications and Networking Conference (WCNC), pp. 1–6. IEEE (2017) 6. Sanguinetti, L., D’Amico, A.A., Rong, Y.: A tutorial on the optimization of amplify-andforward MIMO relay systems. IEEE J. Sel. Areas Commun. 30(8), 1331–1346 (2012) 7. Li, B., Zhang, M., Cao, H., Rong, Y., Han, Z.: Transceiver design for AF MIMO relay systems with a power splitting based energy harvesting relay node. IEEE Trans. Veh. Technol. 69(3), 2376–2388 (2020) 8. Kundu, C., Ghose, S., Bose, R.: Secrecy outage of dual-hop regenerative multi-relay system with relay selection. IEEE Trans. Wirel. Commun. 14(8), 4614–4625 (2015) 9. Khormuji, M.N., Larsson, E.G.: Cooperative transmission based on decode-and-forward relaying with partial repetition coding. IEEE Trans. Wirel. Commun. 8(4), 1716–1725 (2009) 10. Mo, Z., Su, W., Matyjas, J.D.: Amplify and forward relaying protocol design with optimum power and time allocation. In: MILCOM 2016—2016 IEEE Military Communications Conference, pp. 412–417. IEEE (2016)

Chapter 12

A Novel Approach for Bug Triaging Using TOPSIS Pavan Rathoriya, Rama Ranjan Panda, and Naresh Kumar Nagwani

Abstract In the field of software development life cycle, the maintenance phase is one of the focused steps. Normally, thousands of bugs are reported daily by testers. So, it is very important to fix that bug as soon as possible. In the current era, various project are there that work on the same project and if bug not fixed timely the other product can easily overtake the company business, so to fix the newly arrived bug the project manager finds the best developer to fix it and assign to the developer, and this process is called bug triaging. For bug triaging task automation, various methods had been carried out by various researchers machine learning, information retrieval, deep learning, etc., but the cons with that method were that they were not able to fix the problem simultaneously like bug tossing, load balancing, and developer availability. Hence to overcome that we have proposed a method called technique for order of preference by similarity to ideal solution (TOPSIS) which is based on multi-criteria decision-making (MCDM), which will consider the developer metadata with various criteria to automate the task of bug triaging. Based on the criteria, the parameter (closeness ratio) will be calculated, and based on the parameter value, the developer will be ranked for bug triaging.

12.1 Introduction Bug fixing is a vital step of a software project‘s maintenance and development phases. Certain large specific software projects should respond to consumer bugs as early as possible. The very first stage of the bug triaging process is to triage the bugs P. Rathoriya (B) Department of Information and Technology, National Institute of Technology, Raipur, India e-mail: [email protected] R. R. Panda · N. K. Nagwani Department of Computer Science and Engineering, National Institute of Technology, Raipur, India e-mail: [email protected] N. K. Nagwani e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_12

125

126

P. Rathoriya et al.

by allocating it to a developer who can truly fix the bug. Manual bug triaging is a time-consuming job as a large number of bugs reports received every day. When a bug is allocated to the incorrect developer, the expense and time required to fix the bug rise. Preferably, the bug triaging procedure should be time-effective and fast in taking action on newly reported bugs by allocating it to the most suitable and leastbusy developer [1]. Even so, with the emergence of issue tracking system systems, it is not uncommon to see thousands and thousands of bugs reported for fixing on a regular basis. With the rise of ubiquitous computing over the last few decades, large open-source software like OpenStack1 has become increasingly important for managing large pools of computing, storage, and networking resources. Similarly, Mozilla2, Eclipse3, and Jira4 are well-known bug reporting platforms. Many studies have been conducted in the past using various methods based on machine learning or deep learning, but the problem with that is that they were unable to solve load balancing and bug tossing problems at the same time, and they also do not consider the availability of the developers when assigning newly reported bugs. Real bug triaging is a complex multi-criteria decision-making procedure that needs comprehensive info of developer criteria like experience, the number of bugs fixed, their average bug fixing time, the number of bugs assigned, and so on. TOPSIS is a popular problemsolving technique in the multi-criteria decision-making paradigm [2, 3]. Hence, this method can be used to solve the bug triaging problem effectively using the features of developers. In this paper, Sect. 12.2 presents motivation, Sect. 12.3 presents some related work to bug triaging. In Sect. 12.4, the methodology is presented. Section 12.5 describes experimental and results. Section 12.6 covers some threats to validity, and Sect. 12.7 discusses the conclusion and future work.

12.2 Motivation The problems with previous research methods are that they mostly used historical data for training and did not take into account the availability of developers. Even in some studies, load balancing was missing, and the developer’s expertise was not considered in information retrieval methods, which resulted in the tossing of bugs and the time and expanse consuming.

12.3 Related Work Bug triaging is a problem of decision-making in which a suitable developer who can fix the bug is identified. Many researchers have carried out several studies. They have tried a variety of methods. As discussed in the following paragraphs, several authors used machine learning, information retrieval, topic modelling, and MCDM methodologies. In [4–10], various machine learning methodologies are used for bug

12 A Novel Approach for Bug Triaging Using TOPSIS

127

triaging. Sarkar et al. [9] suggested a method for bug triaging across five opensource projects: Bugzilla, Firefox, Eclipse, GCC, and Myly by utilizing five machine learning methods (SVM, expectation maximization (EM), NB, C4.5, and rules). Their method individual worked with developers who had previously fixed identical bugs for a specific project and did not take into account experienced developers who had freshly joined the projects. In Johnson and Zhang [11], they used the TF-IDF model by using the approximate stacked generalization method for classifier ensemble. Similarly, Li and Zhong [12] used the TF-IDF model with attribute description, title, and the appropriate feature selection with Naive Bayes. The existing machine learning mechanism faced challenges in that it is difficult to label bug reports with missing or insufficient label information, and most classification algorithms used in existing approaches are costly and inefficient with large datasets. Information retrieval techniques for automating the bug task are another set of approaches being explored by researchers. This bug dataset is considered as the document/textual form, and later this textual information is transformed as the word vector for bug triaging. Similarity, Zaidi and Lee [13] used the on-hot encoder method to convert the text data to vector matrix for bug triaging [1]. Lee and Seo [14] used the LDA method for text classification to identify the topics. Panda and Nagwani [15] used the intuitionistic fuzzy set-based bug triaging model. The issue with the information retrieval approach is that while a developer can be precisely identified based on the description, but the availability and expertise cannot be determined. Several studies have increased bug triaging accuracy by including additional information like components, products, severity, and priority. Nguyen et al. [16] used the topic mining-based method for bug triaging. Topic mining is generally based on the analysis of sematic characteristics and based on what makes the decision. Similarly, Jahanshahi and Basar [17] used the HLDA method by converting the correlated word into a chunk/phrase. Some bugs necessity be tossed due to the inexperience of the developer or sometimes the developer may be busy in fixing other. However, this may be reduced by selecting the most suitable developer based on the criteria like expertise on how much work has been allocated. Goyal and Sardana [18] used the MCDM method, namely the analytic hierarchy process (AHP) method, for bug triaging, in which the first newly reported bug term tokens are produced, and developers are ranked based on various criteria with different weightages. Gupta et al. [19] used a fuzzy method for order of preference by similarity to ideal solution (F-TOPSIS) with the bacterial foraging optimization algorithm (BFOA) and bar systems (BAR) to improve bug triaging results. It can be inferred from the above discussed method that present methods do not take into account the ranking of developers utilizing metadata and multi-criteria decision-making when selecting a developer. In fact, all of the parameters/features are not equal, hence the weighting and prioritizing of features should be done explicitly. Apart from that, developer availability is not taken into account in the current method. Hence, this paper identifies this gap and propose a bug triaging mechanism based on TOPSIS.

128

P. Rathoriya et al.

12.4 Methodology The detailed discussion of method given in this section presents a novel strategy for bug triaging. The method’s overall architecture is shown in Fig. 12.1. To begin, the bug data is gathered from open sources such as Kaggle and Bugzilla. The dataset contains 10,000 raw records and 29 columns of characteristics including information on bugs, such as the developer who fixed it, the date the bug was discovered, the date the bug was fixed, the bug ID, the bug summary, and so on. Kaggle provided the data for this project. Other open-source platforms, such as the Eclipse, Mozilla, Bugzilla, and others, are also accessible. Preprocessing tasks, such as text lowercasing, stop word removal, tokenization, stemming, and lemmatization, are applied to the bug summary in step two. The dataset’s developer metadata is extracted in the third phase. It includes information such as the developer’s name, the total count of bugs assigned to every developer, the number of bugs fixed by each developer, fresh bugs issued to each developer, the developer’s total experience, and the developer’s average fixing time to resolve all bugs. The developer names and bug summaries are also used to generate developer vocabulary. The newly reported preprocessed bug summary is matched with developer vocabulary in the fourth stage, which employs the cosine similarity [20] threshold filter. Developers are filtered from the developer vocabulary for next step five based on similarities. The developer metadata from step 3 is extracted in the fifth step, and only for the filtered developer from step 4. The analytic hierarchy process AHP [2] approach is used in step 6 to determine the criteria weight. The steps for measuring criteria weight for bug triaging are as follows:

Fig. 12.1 Proposed model architecture

12 A Novel Approach for Bug Triaging Using TOPSIS

129

Problem definition: The problem is described for which criteria are calculated, for example, the problem in this study is to select an expert developer to resolve newly reported bugs. Parameter selection: Appropriate parameter (criteria) are selected for finding their weight. The criteria are developer experience in year (E), total number of bugs assigned (A), newly assigned bugs (N), total bug fixed (R), and average fixing time (F). Create a judgement matrix (J): A squared matrix named A order of m × m is created for pairwise comparison of all the criteria, and the element of the matrix is the comparative importance of criteria to the other criteria. Jm∗m = ki j

(12.1)

where m is the count of criteria and a is the relative importance of criteria. Their entry of an element into the matrix follows the rules as Eqs. (12.2) and (12.3)  ki j = 1 k ji

(12.2)

kii = 1for all i

(12.3)

where i, j = 1, 2, 3,………, m. For relative importance of criteria [21], the data in Table 12.1 will be used. Normalized the matrix J. Then, find eigenvalue and eigenvector W t . Then, a consistency check of weight will be performed. It has the following steps: Calculate λmax by given Eq. (12.4). λmax =

1 n i th inKWt i=1 i th inW t n

The consistency index (CI) is calculate using the Eq. (12.5). Table 12.1 Relative importance of criteria

Importance value

Description

1

Identical importance

3

Reasonable importance

5

Strong importance

7

Very strong importance

9

Extremely strong importance

2, 4, 6, and 8

Middle importance values

(12.4)

130

P. Rathoriya et al.

CI =

(λmax − n) (n − 1)

(12.5)

where n is the count of criteria. The consistency ration (CR) is calculated by using Eq. (12.6). CI RI

CR =

(12.6)

where RI is random index [3]. If the consistency ratio is less than 0.10, the weight is consistent, and we can use the weight (W ) for further measurement in next step. If not, repeat from step 3 of AHP and use the same step. In step 7, the TOPSIS [22] model is applied for ranking the developer. The TOPSIS model has the following steps for developer ranking: Make a performance matrix (D) with the order m × n. Where n is the total count of criteria attributes and m is the alternatives (developers), the element of the matrix will be the respective value for the alternative according to the criteria. Normalize the matrix using the Eq. (12.7).  Ni j = ai j /

m k=1

ai2j

(12.7)

Multiply the normalized matrix (Ni j ) with the criteria weight shown in Eq. (12.8).     V = Vi j m×n = Wi Ni j m×n

(12.8)

Determine the positive idea solution (best alternative) (B* ) and the anti-idea solution (B− ) (worst alternative) using Eqs. (12.9) and (12.10). B∗ = B− =



  maxi Vi j | j ∈ J , mini Vi j | j ∈ J ‘ V1∗ , V2∗ . . . , Vn∗



  maxi Vi j | j ∈ J , mini Vi j | j ∈ J ‘ V1− , V2− . . . , Vn−

(12.9) (12.10)

Find the Euclidean distance between the available alternative and the best alternative called E * , and similarly, from the worst alternative called E − , using Eqs. (12.11) and (12.12).  E i∗

=

n j=1

 E i−

=

n j=1

 2 vi j − v ∗j 

vi j − v ij

(12.11)

2 (12.12)

Find the similarity to the worst condition (CC) or closeness ration using Eq. (12.13), the higher the closeness ration of an alternative, the higher the ranking

12 A Novel Approach for Bug Triaging Using TOPSIS

CCi =

E i∗

E i− + E i−

131

(12.13)

Generally, CC = 1 if the alternative has the best solution. Similarly, if CC = 0, then the alternative has the worst solution [23]. Hence for bug triaging, the developer who has the highest closeness will be ranked first, and the lowest closeness developer will have the last rank for bug triaging.

12.5 Experiment and Results For the experiment purpose, the dataset is taken form Kaggle it contains the 10,000 and 29 columns in which there is total 205 developer are there for fixing the bug, for experiment, we have taken top 20 developer and there bug assigned distribution is shown in Fig. 12.2. After that all the steps of proposed model is applied step by step and the weight of the criteria as calculated shown in Table 12.2 and its consistency is also checked it follow the rules as in [24] like less than 0.1 and the performance of model is evaluated using the matrices like, precision at k, recall at k, NDCG at k, and MAP at k shown in Tables 12.2, 12.3, 12.4 and 12.5. From Table 12.3, 12.4, and 12.5, it can observe that that highest f 1-score is 0.76, respectively, and the highest NDCG score is 0.847, and the highest mean average precision is 0.9. We have also applied the k-fold cross-validation (K = 10) on the model and get the mean accuracy is 92.196%. Figure 12.3 represents the mean

Fig. 12.2 Bug distribution to developer

132

P. Rathoriya et al.

Table 12.2 Weight of criteria Criteria

E

A

N

R

T

Weight

0.0497

0.102

0.069

0.482

0.482

Table 12.3 Bug triaging rank analysis @k k

1

2

3

4

5

6

7

8

9

10

P (%)

9

83

87

92

93

90

88

80

75

73

R (%)

34

39

46

52

58

61

67

71

67

63

F1 (%)

50

53

60

67

71

73

76

75

71

68

Table 12.4 NDCG score for top k developer rank K

1

2

3

4

5

6

7

8

9

10

NDCG

0.74

0.64

0.67

0.72

0.70

0.74

0.70

0.80

0.84

0.85

Table 12.5 Mean average precision (MAP) for top K developer rank K

1

2

3

4

5

6

7

8

9

10

MAP

0.9

0.76

0.66

0.66

0.69

0.73

0.76

0.78

0.80

0.81

average precision graph for top 10 developer rank, and Fig. 12.4 depicts the NDCG score for top 10 ranked developer. Fig. 12.3 MAP@K score line graph

12 A Novel Approach for Bug Triaging Using TOPSIS

133

Fig. 12.4 NDCG@K line graph

12.6 Threats to Validity The proposed method poses a threat due to the use of various criteria and there weight, if the criteria or its weights are changed, then the rank set of developer will be changed. Hence, it is required to weight should be assigned by the same person. Even if the criteria are chosen by different mankind the ideal solution will be different for different criteria, hence it will lead to change the overall rank of developer in bugs triaging.

12.7 Conclusion and Future Scope In this paper, an innovative approach for bug triaging based on MCDM method called TOPSIS is utilized for automation of bug triaging. In the approach, another MCDM method namely AHP is also used to check the consistency of criteria weight, because, the developer is selected based on criteria, hence it is required to weight should be defined rightly to check the performance of the model we used the k-fold crossvalidation (k = 10) and other method ranking evolutionary matrices like NDCG@k, F1-score at k, and MAP at k are also used and there values are, respectively, 84.7%, 76%, and 90%.

134

P. Rathoriya et al.

References 1. Chaitra, B.H., Swarnalatha K.S.: Bug triaging: right developer recommendation for bug resolution using data mining technique. In: Emerging Research in Computing, Information, Communication and Applications. Springer, pp. 609–618 (2022) 2. Issa, U., Saeed, F., Miky, Y., Alqurashi, M., Osman, E.: Hybrid AHP-Fuzzy TOPSIS approach for selecting deep excavation support system. Buildings 12(3), 295 (2022) 3. Singh, A., Biligiri, K.P., Sampath, P.V.: Development of framework for ranking pervious concrete pavement mixtures: application of multi-criteria decision-making methods. Int. J. Pavement Eng. 1–14 (2022) 4. Sun, X., Zhou, T., Wang, R., Duan, Y., Bo, L., Chang, J.: Experience report: investigating bug fixes in machine learning frameworks/libraries. Front. Comput. Sci. 15(6) (2021). https://doi. org/10.1007/s11704-020-9441-1 5. Malhotra, R., Dabas, A., Hariharasudhan, A.S., Pant, M.: A study on machine learning applied to software bug priority prediction. In: Proceedings of the Confluence 2021: 11th International Conference on Cloud Computing, Data Science and Engineering, pp. 965–970 (2021). https:// doi.org/10.1109/Confluence51648.2021.9377083 6. Agrawal, R., Goyal, R.: Developing bug severity prediction models using word2vec. Int. J. Cogn. Comput. Eng. 2, 104–115 (2021). https://doi.org/10.1016/j.ijcce.2021.08.001 7. Goyal, A., Sardana, N.: Feature ranking and aggregation for bug triaging in open-source issue tracking systems. In: Proceedings of the Confluence 2021: 11th International Conference on Cloud Computing, Data Science and Engineering, pp. 871–876 (2021). https://doi.org/10.1109/ Confluence51648.2021.9377053 8. Chowdhary, M.S., Aishwarya, R., Abinay, A., Harikrishna, P.: Comparing machine-learning algorithms for anticipating the severity and non-severity of a surveyed bug. In: Proceedings of the International Conference on Smart Technologies in Computing, Electrical and Electronics, ICSTCEE 2020, pp. 504–509 (2020). https://doi.org/10.1109/ICSTCEE49637.2020.9276756 9. Sarkar, A., Rigby, P.C., Bartalos, B.: Improving bug triaging with high confidence predictions at Ericsson. In: Proceedings—2019 IEEE International Conference on Software Maintenance and Evolution, ICSME 2019, pp. 81–91 (2019). https://doi.org/10.1109/ICSME.2019.00018 10. Jahanshahi, H., Cevik, M., Navas-Sú, J., Ba¸sar, A., González-Torres, A.: Wayback machine: a tool to capture the evolutionary behaviour of the bug reports and their triage process in open-source software systems. J. Syst. Softw. 111308 (2022) 11. Johnson, R., Zhang, T.: Supervised and semi-supervised text categorization using LSTM for region embeddings. In: International Conference on Machine Learning, pp. 526–534 (2016) 12. Li, Z., Zhong, H.: Revisiting textual feature of bug-triage approach. In: 2021 36th IEEE/ACM International Conference on Automated Software Engineering (ASE), pp. 1183–1185 (2021) 13. Zaidi, S.F.A., Lee, C.G.: One-class classification based bug triage system to assign a newly added developer. In: International Conference on Information Networking, pp. 738–741 (2021). https://doi.org/10.1109/ICOIN50884.2021.9334002 14. Lee, D.G., Seo, Y.S.: Improving bug report triage performance using artificial intelligence based document generation model. Human-centric Comput. Inf. Sci. 10(1) (2020). https://doi. org/10.1186/s13673-020-00229-7 15. Panda, R.R., Nagwani, N.K.: Classification and intuitionistic fuzzy set based software bug triaging techniques. J. King Saud Univ. Comput. Inf. Sci. (2022) 16. Nguyen, U., Cheng, K.S., Cho, S.S., Song, M.: Analyzing bug reports by topic mining in software evolution. In: Proceedings—2021 IEEE 45th Annual Computers, Software, and Applications Conference, COMPSAC 2021, pp. 1645–1652 (2021). https://doi.org/10.1109/COM PSAC51774.2021.00246 17. Jahanshahi, H., Ba¸sar, A.: DABT: a dependency-aware bug triaging method (2021). https:// doi.org/10.1145/3463274 18. Goyal, A., Sardana, N.: Optimizing bug report assignment using multi criteria decision making technique. Intell. Decis. Technol. 11(3), 307–320 (2017). https://doi.org/10.3233/IDT-170297

12 A Novel Approach for Bug Triaging Using TOPSIS

135

19. Gupta, C., Inácio, P.R.M., Freire, M.M.: Improving software maintenance with improved bug triaging. J. King Saud Univ. Comput. Inf. Sci. King Saud bin Abdulaziz University (2021). https://doi.org/10.1016/j.jksuci.2021.10.011 20. Guo, S., et al.: Developer activity motivated bug triaging: via convolutional neural network. Neural Process. Lett. 51(3), 2589–2606 (2020). https://doi.org/10.1007/s11063-020-10213-y 21. James, A.T., Vaidya, D., Sodawala, M., Verma, S.: Selection of bus chassis for large fleet operators in India: an AHP-TOPSIS approach. Expert Syst. Appl. 186 (2021). https://doi.org/ 10.1016/j.eswa.2021.115760 22. Rafi, S., Akbar, M.A., Yu, W., Alsanad, A., Gumaei, A., Sarwar, M.U.: Exploration of DevOps testing process capabilities: an ISM and fuzzy TOPSIS analysis. Appl. Soft Comput. 116, 108377 (2022) 23. Panigrahi, R., Borah, S.: Rank of normalizers through TOPSIS with the help of supervised classifiers. Retrieved from www.sciencepubco.com/index.php/IJET (2018) 24. Asemi, A. Ko, A., Asemi, A.: The AHP-TOPSIS based DSS for selecting suppliers of information resources. In: 2022 Second International Conference on Distributed Computing and High Performance Computing (DCHPC), pp. 104–116 (2022). https://doi.org/10.1109/DCH PC55044.2022.9732125

Chapter 13

Energy-Efficient Resource Allocation in Cognitive Radio Networks Kartik Verma and Manoranjan Rai Bharti

Abstract Through this paper, an attempt has been made to improve the energy efficiency of the resource allocation process for Orthogonal Frequency Division Multiplexing (OFDM)-based cognitive radio (CR) networks with imperfect spectrum sensing. Maximizing the energy efficiency of the multi-user CR system is the main objective, while simultaneously adhering to the total transmission power constraint along with each Primary User’s (PU) interference constraint, and also ensuring that the capacity of the system remains as high as possible. The errors in spectrum sensing have been taken into account, and then, the interference model is formulated. Windowing has been introduced to reduce the interference which is caused to PUs. The resource allocation process is a mixed-integer nonlinear programming problem. So, the resource allocation scheme is separated into two steps, i.e., sub-carrier assignment and power allocation, which means that the proposed algorithm provides suboptimal solution. Firstly, sub-carrier assignment is done, followed by power allocation. Fractal programming approach and sub-gradient method are used for the power allocation process. The simulation results show that there is an improvement in energy efficiency as compared to the previously used methods. Taking the sensing errors into account can also protect the normal communication of the PU. Windowing method used for reducing the interference, in turn, increases the capacity of the system.

13.1 Introduction We are living in the age of technological progress. The technology we are using today was not even thought of a few hundred years back. Similar is the case in the field of wireless communication. With the increase in users and the demand of better wireless communication technologies with each passing day, the number of devices K. Verma (B) · M. R. Bharti National Institute of Technology Hamirpur, Hamirpur, Himachal Pradesh 177005, India e-mail: [email protected] M. R. Bharti e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_13

137

138

K. Verma and M. R. Bharti

is increasing. Thus, the cost of operation is also increasing and the spectrum resources are becoming increasingly scarce. These are not the only costs that the world has paid while striving for technological advancements, as it has also led to much greater problems, like global warming and climate change. Wireless communication field is also contributing in climate change. With the increase in the number of users, the energy usage has increased to unsustainable limits leading to global warming and eventually climate change. According to the work done in [1], the usable spectrum utilization is very low. Spectrum utilization was prioritized in further studies for obvious reasons of cost and energy usage minimization; thus, the use of Cognitive Radio (CR) technology is proposed in [2] for better spectrum utilization by allowing the Secondary Users (SU) to access the spectrum hole vacated by the PU when it is not communicating. In [3–5], authors have investigated resource allocation scheme in Orthogonal Frequency Division Multiplexing (OFDM)-based CR networks for maximizing capacity and simultaneously adhering to the power and interference constraints while taking sensing errors into account. In recent years, emphasis has been given on green wireless communication. The governments around the world have come up with rigid environmental regulations in an attempt to minimize the energy usage. Thus, there is an emerging trend of addressing ‘energy efficiency’ as an important aspect of wireless communication technology. The authors in [6] have worked out an algorithm for maximizing energy efficiency in the CR networks, but the interference introduced into the PU is not taken into account. The authors in [7, 8] have proposed an algorithm to obtain the maximum energy efficiency while also considering the interference introduced into the PU. In [7], only one SU and PU are used and sensing errors are not considered. While in [8], multiple PUs and SUs are considered along with sensing errors due to imperfect spectrum sensing. In [9, 10], authors have introduced the concept of green wireless networks. In this paper, we strive to improve upon the work in [7, 8]. We have tried to maximize the energy efficiency while allocating the resources in a CR network and also taking the sensing errors into consideration. In the proposed work, the interference is also decreased with the application of windowing. Firstly, the interference model is formulated under imperfect spectrum sensing, and then, resource allocation is done with the objective function of maximizing the energy efficiency of the multi-user CR network. This primal problem is a mixed-integer nonlinear programming issue; therefore, a suboptimal solution is proposed to reduce the computational complexity by separating the power and sub-carrier allocation. Firstly, sub-carriers are distributed to SUs and initial power allocation is done. Then, the final power allocation is done. Still, the problem is quite complex as the interference constraints and total power constraint need to be considered while the final power allocation is being done to ensure reliable communication for each PU. In the proposed work, the final power allocation to sub-carriers is performed by using the sub-gradient method. Finally, the results of simulations show an improvement in energy efficiency. The rest of the paper is arranged as follows. In Sect. 13.2, system model of the CR network based on OFDM

13 Energy-Efficient Resource Allocation in Cognitive Radio Networks

139

with imperfect spectrum sensing is discussed. In Sect. 13.3, energy-efficient resource allocation process is described. In Sect. 13.4, simulation results are presented, and in Sect. 13.5, conclusions of this work are provided.

13.2 System Model We have considered a CR network comprising of M PUs and H SUs. A set of S sub-carriers is taken, which are orthogonal to each other and the total bandwidth is divided among S sub-carriers. Each sub-carrier has a bandwidth of B Hz. The PUs are licensed users and can access the spectrum at any time whereas the SUs can access it only when the PU vacates it. The CR network identifies the sub-carriers which are not occupied by the mth PU through spectrum sensing. The vacant sub-carriers’ set is denoted by Sum , and the occupied sub-carriers’ set is denoted by Som . The set of all vacant sub-carriers can be denoted by Su = Su1 ∩ Su2 · · · ∩SuM . The set of all occupied sub-carriers can be denoted by So = So1 ∪ So2 · · · ∪SoM . The probability that mth PU is using the sth sub-carrier is Psm . Typically, there are two sensing errors mis-detection and false alarm which are considered E smis and E sfal , respectively. The probability that the CR network detects that the sth sub-carrier is occupied by the mth PU and is truly occupied is denoted by ψsm . Also, the probability that the sth sub-carrier is indeed occupied by the mth PU and the CR deems it to be unoccupied is denoted as τsm . Thus, m

ψsm = Pr {Osm |O s }

 m    Pr O s |Osm Pr Osm  m     m  = Pr O s |Osm Pr Osm + Pr O s |Usm Pr {Usm }   1 − E smis Psm   , = 1 − E smis Psm + E sfal 1 − Psm

(13.1)

m

τsm = Pr {Osm |U s }

 m    Pr U s |Osm Pr Osm  m     m  = Pr U s |Osm Pr Osm + Pr O s |Usm Pr {Usm }

=

(13.2)

E mis P m ,  s s  E smis Psm + 1 − E sfal 1 − Psm

where Osm and Usm are considered to be the events when the mth PU has occupied m m and not occupied the sth sub-carrier, respectively. Also, O s and U s are the events that the sth sub-carrier is, respectively, occupied and not occupied by the mth PU as per information sensed by the CR networks. The transmitted signal is an OFDM

140

K. Verma and M. R. Bharti



sin (π f Ts ) 2 signal which has the power spectral density ϕ ( f ) = Ts for a default π f Ts rectangular window, where Ts denotes the duration of OFDMA symbol. Let g mj is the channel gain to the receiver of mth PU on jth sub-carrier from the CR system. s represents the interference power introduced into the mth PU on the Also, let I j,m jth sub-carrier when the CR network is transmitting on the sth sub-carrier having unity transmission power. We can write as j B−(s−1/2)B

s I j,m

=

g mj ϕ ( f ) d f.

(13.3)

( j−1)B−(s−1/2)B

Therefore, the interference power which is being introduced into the mth PU over the sth sub-carrier being used by a SU which has unit transmission power can be expressed by   s s Ism = ψ mj I j,m + τ jm I j,m . (13.4) j∈Som

j∈Sum

For a reliable communication between the SUs and to ensure that PUs communication does not get disturbed, the interference has to be kept at a minimum. In order to reduce the interference, the use of a windowing function has been proposed in this work. The windowing method is used to reduce the spectral side lobes of OFDM signal as shown in [11]. As a result, the PSD of the OFDM signal gets reduced to some extent. As can be seen from (13.3), the interference introduced into the PU is directly proportional to the PSD of the OFDM signal. Thus, windowing the OFDM signal reduces the interference. In this paper, raised cosine window has been used for the reduction of interference. The raised cosine window can be expressed as 

t i(t) = sinc Ts



π αt cos Ts ,  2αt 2 1− Ts

where α is the roll-off factor and 0 ≤ α ≤ 1 and the auto-correlation function of the raised cosine window is ⎡   ⎤ π αt πt  cos  cos ⎢ t α αt Ts Ts ⎥ ⎢ ⎥ R(t) = Ts ⎢sinc 2 − sinc   2 ⎥ . ⎣ Ts 4 T 2αt αt ⎦ s 1− 1− Ts Ts Let kh,s denote the channel gain between the hth SU which is on the sth sub-carrier and access point of the CR. The transmission power which has been allocated to the hth SU is denoted by ph,s . No is taken to be the additive Gaussian noise power. σh,s is

13 Energy-Efficient Resource Allocation in Cognitive Radio Networks

141

the signal-to-noise ratio (SNR) on sth sub-carrier by the hth SU, and σh,s = kh,s /No . Let C be the total throughput of the system which can be calculated as C=B

H  

  ηh,s log2 1 + ph,s σh,s ,

(13.5)

h=1 s∈Su

where ηh,s is a binary variable with values 1 or 0 depending on whether the sth sub-carrier is allocated to the hth SU or not. In CR networks, the energy consumed in the circuit components and during the transmission of signals comprises the total energy consumption. In accordance with [10], the circuit energy consumption, which is denoted as Pc , can be represented as Pc = Ps + βC, where Ps is static circuit power and β is a constant representing the consumption of power with unit transmission rate. After the formulation of the interference model, the main aim is to maximize the energy efficiency of the CR network while not violating the transmission power constraint and the interference constraints of the PUs. Thus, the objective problem can be formulated as OP1 :

max  H  p ,η h,s

s.t.

h,s

H  

h=1

C s∈Su

ηh,s ph,s + Pc

ηh,s ph,s ≤ Pth , ph,s ≥ 0, s ∈ Su

h=1 s∈Su H  

ηh,s ph,s Ism ≤ IthM , m = 1, 2, 3, . . . , M

h=1 s∈Su H 

ph,s ≤ 1, ηh,s ∈ {0, 1}, s ∈ Su

h=1

where Pth is total transmission power (threshold) available at secondary base station and Ithm is the interference threshold of the mth PU.

13.3 Resource Allocation In the following section, the method for resource allocation for maximization of energy efficiency in the Cognitive Radio networks is given. The maximization problem (OP1) mentioned above, being non-convex, belongs to a class of mixed-integer nonlinear programming problems, which is also termed as NP hard. Thus, a suboptimal solution is provided to this problem in which the two resource allocation steps, i.e., sub-carrier allocation and the allocation of power, are separately performed.

142

K. Verma and M. R. Bharti

13.3.1 Sub-carrier Allocation In the following subsection, we allocate sub-carriers to the SUs following a strategy which can roughly adhere to the interference constraints as well as power constraints and also remove the integer constraints simultaneously. In CR networks, which are OFDM based, if the sub-carrier is next to the sub-carrier used by the PU and greater the channel gain g mj , greater will be the interference introduced into the mth PU as per Eqs. (13.3) and (13.4). Hence, these factors should also be taken into account while allocating the sub-carrier. We have assumed that every sub-carrier which is vacant, i.e, s ∈ Su , shares equal responsibility to ensure normal communication to each PU. Firstly, for sub-carrier assignment, initial power has to be allocated to each sub-carrier and the sum of initial power allocation to all sub-carriers cannot be greater than Pth . Thus, the initial power allocation to each sub-carrier s ∈ Su is calculated as  Pth Ith IthM . ps = min , ,..., |Su | |Su | Is1 |Su | IsM 

(13.6)

Initial power allocation is followed by sub-carrier allocation, and it is done by solving the optimization problem OP1. It is seen that the interference constraints and the total power constraints have been satisfied. Thus, the allocation of the sub-carriers is done by maximization of the energy efficiency of the CR network. This means that the sth sub-carrier is assigned to that SU h for which it can provide maximum energy efficiency as per the initial allocation of power, i.e.,   log2 1 + ps σh,s h = argmax  , Psen h ps + + ξ log2 1 + ps σh,s |Su | 

(13.7)

where Psen is the sensing power. After the sub-carrier assignment, the SNR is updated over the sth sub-carrier σs = σh  ,s . This process continues until all the vacant subcarriers have been assigned to SUs. With the sub-carrier allocation, the constraints have been roughly satisfied and can be completely achieved after the final power allocation to sub-carriers is performed as given in the next subsection.

13.3.2 Power Allocation As stated earlier, the integer constraints have already been removed after sub-carrier allocation and the binary assignment variable ηk,n has already been fixed. It can be seen that the primal problem’s objective function is not convex, so the computational complexities make it difficult to solve the problem directly. Thus, to solve this problem, first of all, fractal programming as in [12] has been introduced and then the proposed optimal power allocation strategy can be realized.

13 Energy-Efficient Resource Allocation in Cognitive Radio Networks

143

The objective problem of maximization of energy efficiency can be converted into an equivalent minimization problem to facilitate the analysis and can be written as OP2 : s.t.

E n ( p) min p C ( p)  ps ≤ Pth , ps ≥ 0,

s ∈ Su

s∈Su



ps Ism ≤ Ithm ,

m = 1, 2, 3, . . . , M

s∈Su

where p is a set of power assigned to all sub-carriers, i.e., p = ( p1 , p2 , p3 , . . . ps ). E n ( p) is the total energy consumption, and C( p) is the total capacity of the CR network. E n ( p) and C( p) can be calculated by the following formulae: E n ( p) =



ps + Psen + β B

s∈Su



log2 (1 + ps σs ) ,

(13.8)

s∈Su

C( p) = B



log2 (1 + ps σs ) .

(13.9)

s∈Su

Both the energy consumed E n ( p) and the capacity C( p) functions are real-valued continuous functions and C( p) ≥ 0 in practical CR systems. Both E n and C are functions of p, and p ∈ R, where, R is the definitional domain of power p in optimization problems termed as OP1 and OP2. To obtain the solution of this non-convex problem, help of fractal programming [12] is taken and a new objective function derived from the minimization problem OP2 can be formulated as k( p, φ) = E n ( p) − φC( p),

(13.10)

where φ is a parameter with a positive value. Thus, OP2 can be converted into a new optimization problem and it can be written as OP3 : s.t.

mink( p, φ) p  ps ≤ Pth , ps ≥ 0,

s ∈ Su

s∈Su



ps Ism ≤ Ithm ,

m = 1, 2, 3, . . . , M

s∈Su

Let Q(φ) = min p {E n ( p) − φC( p)| p ∈ R} be the minimum value of k( p, φ) for each fixed φ. Thus, the relation between OP2 and OP3 is pointed out by the lemma given below

144

K. Verma and M. R. Bharti

Lemma 1 φ = E n ( p ∗ )/C( p ∗ ) = min p {E n ( p)/C( p)| p ∈ R} with one condition, i.e., Q(φ ∗ ) = Q(φ ∗ , p ∗ )

(13.11)

= min{E n ( p) − φ ∗ C( p)| p ∈ R} = 0 p

The detailed explanation of this process can be seen in [12]. From Lemma 1, it can be concluded that when φ = φ ∗ both OP2 and OP3 have the same optimal solution p, where the minimum value of the objective function in OP2 is φ ∗ . With the two given problems being equivalent, loop iteration algorithm is best method to solve the problems. This method involves the application of convex optimization theory for observing the scheme for optimal power allocation for a given φ, then, until (13.11) is satisfied φ is updated. The Lagrange function of OP3 whose objective function is a convex function of p when φ is fixed can be written as ⎛ L g = E n ( p) − φC( p) + γ0 ⎝

 s∈Su

⎞ ps − Ptot ⎠ +

M  m=1

⎛ γm ⎝



⎞ ps Ism − Ithm ⎠ ,

s∈Su

(13.12) where γ0 and γm are the dual variables, where γ0 is the Lagrangian variable for power constraint and γm is the Lagrangian variable for interference constraints. So, the derivative of L g with respect to ps is M  δL g σs = 1 + (β − φ) B + γ0 + γm Ism . δps (1 + ps σs ) ln 2 m=1

(13.13)

With the use KKT (Karush-Kuhn-Tucker) conditions, the optimal power allocation of problem OP3 as solved in [8] can be written as ⎡

⎤+ 1 − β) B (φ  ps = ⎣  − ⎦ . M m 1 + γ0 + m=1 γm Is ln 2 σn

(13.14)

There are M + 1 constraints on the CR system model, where M is the number of total interference constraints as there are M number of PUs, with each PU having different interference constraint and there is 1 transmission power constraint. Thus, in this work, M + 1 values of power are calculated for each sub-carrier satisfying one constraint at a time and choosing the power which is least among all. Choosing the least amount of transmitted power results in an improvement of energy efficiency.

13 Energy-Efficient Resource Allocation in Cognitive Radio Networks

145

 1 + (φ − β) B ps0 = − , (1 + γ0 ) ln 2 σs  + 1 (φ − β) B  psm =  − , 1 + γm Ism ln 2 σs 

 ps M =

.. . 1 (φ − β) B   − M 1 + γ M Is ln 2 σs

+ ,

where (.)+ = max(., 0)   ps = min ps0 , ps1 , ps2 , . . . , ps M .

(13.15)

It is found that with the help of the traditional water filling algorithm, we are unable to directly solve the primal problem, as the denominators in (13.14) and (13.15) contain linear combination of Lagrangian dual factors. Therefore, to realize the optimal solution for power allocation scheme a method called sub-gradient method [13] has been used. In this method, firstly a step sequence is designed, followed by the updation of γ toward the sub-gradient direction, where γ = (γ0 , γ1 , γ2 , γ3 , . . . , γ M ). In Eqs. (13.14) and (13.15), the corresponding update in the Lagrangian variables can be written as ⎡ ⎛ ⎞ ⎤+  γ0n+1 = ⎣γ0t − δ n ⎝ Pth − p s ⎠⎦ , (13.16) s∈Su

⎡ γmn+1

=

⎣γmn

⎛ −δ

n

⎝ Ithm

−



⎞⎤+ ps Ism ⎠⎦

, m = 1, 2, . . . , M.

(13.17)

s∈Su

where δ n is a step size sequence and is > 0. Generally, the step size δ n can be √ arbitrarily set to D/n or D/ n, where D is a constant and D > 0 and n is the number of iterations. δ n gets sufficiently small after an adequate number of iterations, and then, γ ∗ converges to the point that gives an optimal solution. After using the sub-gradient method, optimal power assignment for all sub-carriers for given φ is realized and then E n ( p) and C( p) can be calculated. Then until, Lemma 1 is satisfied, keep on updating φ = E n ( p)/C( p) and then the equivalent optimal allocation of power p ∗ of OP3 can be eventually achieved. The proposed resource allocation procedure has been described in Algorithm 1.

146

K. Verma and M. R. Bharti

Algorithm 1 Resource Allocation Process Input: 1. Enter the values of M, B, H, S, E sfal , E smis , β, α, Ithm , Psm , Psen , Pmax Resource Allocation: A. Sub-carrier Allocation: 2. Calculate the interference introduced into mth PU over the sth sub-carrier according to Eqs. (13.3) and (13.4) with and without the raised cosine window; 3. Calculate the initial power allocated to each sub-carrier according to Eq. (13.6) 4. According to Eq. (13.7), assign the sub-carrier s ∈ Su to the SU h  for which the energy efficiency is maximum; B. Optimal Power Allocation: Initialize φ = β+ ∈ where ∈> 0, Q(φ) = ∞ and also the tolerable error χ while |Q(φ)| > χ do Initialize γ and γ = [χ, χ, χ, . . . , χ]; while 2-norm ||γ /γ || > χ do Calculate ps for all s ∈ Su according to Eqs. (13.14) and (13.15); Update γ0 and γm where m = (1, 2, . . . , M) according to (13.16) and (13.17) and then compute γ = γn+1 − γn ; 11. end while 12. Compute Q(φ) and then update φ = E n ( p)/C( p) 13. end while 5. 6. 7. 8. 9. 10.

Output: 14. Return ηh,s , ps , φ ∗ , and then calculate the capacity, energy consumed and then energy efficiency of the CR network for the given system model.

13.4 Results and Discussions For the simulations, k and g are have been modeled as the independent channel gains and are considered to be Rayleigh random variables which are identically distributed having average of 0 dB. In the system model mentioned in Sect. 13.2, the value of M is taken as 2, i.e., the number of primary users is 2. The number of secondary users is taken to be 10, i.e., H = 10 SUs, and number of sub-carriers is taken S = 64. The sensing power Psen is the power consumed by network for spectrum sensing and is taken to be 0.02 W. The other parameters, i.e., the power of additive white Gaussian noise N0 , duration of OFDM signal Ts , and bandwidth of each sub-carrier B are 10−6 W, 5 µs, and 0.3125 MHz, respectively. The false alarm E sfal is uniformly is distributed over [0.01, 0.05] distributed over [0.05, 0.1], and mis-detection Q mis s uniformly. As in the primary network we have assumed two active PUs, the central location of 1st PU is c1 = 16, and central location of 2nd PU is c2 = 48. Thus, the 2 activity probability Psm is considered to be 0.8e−((s−cm ) )/50 . The simulation results have been averaged over 1000 channel realizations. Figure 13.1 shows the performance of the CR system in terms of energy efficiency versus Ith . It compares energy efficiency when power is calculated from Eq. (13.14) as done in [8] to when power is calculated from Eq. (13.15) with and without

13 Energy-Efficient Resource Allocation in Cognitive Radio Networks

147

Fig. 13.1 Energy efficiency versus interference threshold

the application of windowing. It is assumed that Pth = 0.1 W and the interference threshold is same for both the PUs, i.e., Ith1 = Ith2 = 0.1 mW. The unit transfer rate is considered to be β = 0.002 W/Mbps. From the figure, it can be concluded that the energy efficiency of the system improves with the proposed power allocation scheme given by Eq. (13.15) as compared to the power allocation strategy of [8] given in Eq. (13.14). However, with the application of windowing in the proposed approach given by Eq. (13.15) as well as the approach given by Eq. (13.14), the energy efficiency of the system is reduced. This is due to the fact that with the application of windowing technique, the SU sub-carriers can be allocated with higher powers, thus leading to a decrease in the energy efficiency of the system. By calculating allocated power by the proposed method, there is a decrease in allocated power, leading to an increase in energy efficiency. Figure 13.2 shows the performance of the CR system under consideration in terms of capacity versus Ith . It compares the capacity when power is calculated from Eq. (13.14) as done in [8] to when power is calculated from Eq. (13.15) with and without the application of windowing. It is obvious that with the increase in interference threshold the capacity increases up to a certain value then becomes constant. It can also be seen that with the application of windowing the capacity of the system can be increased but it comes at the cost of decreased energy efficiency (compare with Fig. 13.1). Thus, with the proposed method, energy efficiency increases but there is a decrease in capacity of the system. Therefore, by using the proposed method along with windowing to calculate power allocations, we can improve the energy efficiency of the system under consideration. By using the proposed method, even though capacity of the system decreases but it is quite comparable to the capacity obtained in [8].

148

K. Verma and M. R. Bharti

Fig. 13.2 Capacity versus interference threshold

Fig. 13.3 Interference introduced to the mth PU versus transmission power threshold

13 Energy-Efficient Resource Allocation in Cognitive Radio Networks

149

In Fig. 13.3, we plot the interference introduced to the mth PU versus Pth . It can be seen from the figure that the interference introduced into the mth PU is always less than Ith = 0.1 mW. It may also be observed that with the application of windowing technique, the interference introduced into the mth PU is decreased considerably, leading to higher capacity of CR network and also ensuring the reliable communication for the PUs.

13.5 Conclusions In this paper, energy-efficient resource allocation in cognitive radio networks with imperfect spectrum sensing is studied. As already illustrated, an energy-efficient approach is important due to the ongoing energy crisis in the world. As the problem of maximizing energy efficiency in a multi-user network is a mixed-integer nonlinear programming problem, a suboptimal approach is followed by first proceeding with sub-carrier allocation and then the power allocation. Fractal programming and subgradient method is used to realize the optimal power allocation process. Windowing technique has been applied to reduce the interference introduced into the PU system. Simulation results show that there is an improvement in energy efficiency by using the proposed method to calculate allocated powers, and capacity is also kept comparable by applying windowing.

References 1. Islam, M.H., Koh, C.L., Oh, S.W., Qing, X., Lai, Y.Y., Wang, C., Liang, Y.-C., Toh, B.E., Chin, F., Tan, G.L., et al.: Spectrum survey in Singapore: occupancy measurements and analyses. In: 2008 3rd International Conference on Cognitive Radio Oriented Wireless Networks and Communications (CrownCom 2008), pp. 1–7. IEEE (2008) 2. Mitola, J., Maguire, G.Q.: Cognitive radio: making software radios more personal. IEEE Pers. Commun. 6(4), 13–18 (1999) 3. Almalfouh, S.M., Stüber, G.L.: Interference-aware radio resource allocation in OFDMA-based cognitive radio networks. IEEE Trans. Veh. Technol. 60(4), 1699–1713 (2011) 4. Wang, S., Zhou, Z.-H., Ge, M., Wang, C.: Resource allocation for heterogeneous multiuser OFDM-based cognitive radio networks with imperfect spectrum sensing. In: 2012 Proceedings IEEE INFOCOM, pp. 2264–2272. IEEE (2012) 5. Kaligineedi, P., Bansal, G., Bhargava, V.K.: Power loading algorithms for OFDM-based cognitive radio systems with imperfect sensing. IEEE Trans. Wirel. Commun. 11(12), 4225–4230 (2012) 6. Pei, Y., Liang, Y.-C., Teh, K.C., Li, K.H.: Energy-efficient design of sequential channel sensing in cognitive radio networks: optimal sensing strategy, power allocation, and sensing order. IEEE J. Sel. Areas Commun. 29(8), 1648–1659 (2011) 7. Wang, Y., Xu, W., Yang, K., Lin, J.: Optimal energy-efficient power allocation for OFDM-based cognitive radio networks. IEEE Commun. Lett. 16(9), 1420–1423 (2012) 8. Gao, Y., Xu, W., Li, S., Niu, K., Lin, J.: Energy efficient resource allocation for cognitive radio networks with imperfect spectrum sensing. In: 2013 IEEE 24th Annual International

150

9. 10. 11. 12. 13.

K. Verma and M. R. Bharti Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC), pp. 3318– 3322. IEEE (2013) Chen, Y., Zhang, S., Xu, S., Li, G.Y.: Fundamental trade-offs on green wireless networks. IEEE Commun. Mag. 49(6), 30–37 (2011) Xiong, C., Li, G.Y., Zhang, S., Chen, Y., Xu, S.: Energy-and spectral-efficiency tradeoff in downlink OFDMA networks. IEEE Trans. Wirel. Commun. 10(11), 3874–3886 (2011) Abdulkafi, A.A., Sileh, I.K., Hardan, S.M.: Windowing techniques for reducing PAPR of OFDM in Li-Fi systems. J. Opt. Commun. (2019) Dinkelbach, W.: On nonlinear fractional programming. Manag. Sci. 13(7), 492–498 (1967) Ngo, D.T., Tellambura, C., Nguyen, H.H.: Resource allocation for OFDMA-based cognitive radio multicast networks with primary user activity consideration. IEEE Trans. Veh. Technol. 59(4), 1668–1679 (2010)

Chapter 14

Medical Internet of Things and Data Analytics for Post-COVID Care: An Analysis Salka Rahman, Shabir Ahmad Sofi, Suraiya Parveen, and Saniya Zahoor

Abstract The healthcare services across the world have been badly affected by the pandemic since December 2019. People have suffered in terms of medical supplies and treatments because existing medical infrastructure has failed to accommodate huge number of COVID infected patients. Further, patients with existing morbidities have been the worst hit so far and need attention. Therefore, there is a need of postCOVID care for such patients which can be achieved by using technologies such as Internet of Things (IoT) and data analytics. This paper presents medical IoT-based data analysis for post-COVID care. This paper, further, presents post-COVID data analysis to get an insight into the various symptoms across the different perspectives.

14.1 Introduction The coronavirus pandemic initiated in August 2019 from Wuhan city situated in China and within a few months the whole world got affected. According to W.H.O, coronavirus is defined as an infectious virus which is caused by SARC-CoV-2 virus. People who are infected will experience mild to moderate respiratory illness and will recover without and medical help but some, however, will need proper medical care and any age group can get affected by it and may get severely ill or even they can S. Rahman (B) Department of Computer Science and Engineering, Jamia Hamdard University New Delhi, Delhi 110062, India e-mail: [email protected] S. A. Sofi Department of Information Technology, National Institute of Technology Srinagar, Jammu and Kashmir 190006, India e-mail: [email protected] S. Parveen Department of Computer Science, Jamia Hamdard University New Delhi, Delhi 110062, India S. Zahoor Department of Computer Science, University of Kashmir Srinagar, Jammu and Kashmir 190006, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_14

151

152

S. Rahman et al.

die [1]. The pandemic came with many negative health impacts resulting in serious mental and physiological concerns such as anxiety, depression, loneliness, cardiac attacks, arrhythmia, and other health issues. The whole concept of normalcy that we used to live back in the days got totally changed and as a result the patient care and healthcare systems also got affected. The elderly people and others with existing morbidity are a major concern in situations like these. The general symptoms of COVID-19 are pneumonia, body ache, fever, rashes, diarrhoea, cough, fatigue, loss of taste or smell, breathlessness, and so on. Further, from the ongoing research, it has been found that 10–20% of COVID recovered patients experience post-COVID symptoms and these conditions are known as long COVID or post-COVID symptoms. The symptoms seen after recovery from COVID-19 includes fever, memory problem, difficulty in breathing, chest pain, anxiety, body ache, headache, tiredness, vomiting, nausea, organ damage, brain malfunctioning, brain fog, and so on. The duration of occurrence of the long COVID symptoms is still not clearly specified, however, after 3–4 months of recovery from COVID, people may experience post-COVID symptoms and the status of severity of the symptoms varies from person to person [2–5]. The availability for post-COVID data scenarios is quite limited, however, by analysing the symptoms in both the cases, it is evident that COVID and post-COVID symptoms possess similarity. Data collection is a method by which related data is gathered for a given problem. It is a major step involved in the research purposes or any other project in the company. Thus, a number of methods are available for different purposes such as surveys, polls, experimental observations, reports, etc. These methods provide a way by which data is gathered in an efficient way. Medical IoT/M-IoT is another approach for data collection used frequently for real-time analysis. It uses sensors to monitor the patient’s health status and stores the collected information in cloud infrastructure or any other servers or databases which ultimately help the doctors in keeping the track of the patient’s condition by simply using devices like smart phones, laptops, and tablets remotely. IoT devices and sensors are used in shaping the way that we live our lives today. IoT is a giant network of connected devices, these devices gather and share the data. In every physical device, sensors are embedded that provides the data which is then stored and pre-processed for valuable information extraction as per the requirement. IoT devices play a vital role in processing the patient’s condition, consider an example where a patient’s body temperature level needs to be monitored so, for that IoT-based device such as smart thermometer can be used for analysing the condition of the patient on a regular basis in real time. The collected data can be stored on the cloud and pre-processing of the same can be done over there, therefore, the doctors can then easily monitor and track the patient’s progress by simply using his smartphone. In critical situations like the pandemic, people avoid going for regular check-up’s as it is quite risky thus, IoT-based remote healthcare monitoring system can be used for keeping the health track of COVID and post-COVID patients. So, different IoT sensing devices like the fitness bands, oximeter, ECG/EEG sensors, qardiocore, and so on can be utilized for acquiring the data continuously so that the monitoring of patients can be done efficiently while patient’s stay at their respective homes and be safe.

14 Medical Internet of Things and Data Analytics for Post-COVID Care: … Fig. 14.1 Data analytics

153

Descripve analycs

Data analycs Predicve analyycs

Prescripve analycs

Medical data is expanding every day, to process and store this huge amount of data is a very complex task thus, healthcare data analytics is one of the approach that can be incorporated for the analysis of medical data especially for remote patient monitoring. Healthcare data analytics is the process of analysing data for decision making and improve performance of healthcare organizations. Thus, these approaches can be utilized for solving healthcare real-world applications. Further, the medical data collected on electronic health recorder needs proper exploratory data analysis (EDA) to analyse the data [6]. Data analytics are used to get insights out of healthcare data so that the hidden patterns can be extracted and utilized for treatments, to discover new drugs, diagnosis purposes, disease predictions research purposes, and so on in future. There are basically three major types of healthcare analytics, i.e. descriptive, predictive, and prescriptive. Descriptive analytics mainly focuses on existing data which has already happened and is available. Predictive analytics on the other hand focuses on what will happen next and finally prescriptive analytics tells us about the creation of future by providing suggestions that are best to avoid and also suggests what to incorporate. Refer Fig. 14.1. The paper is organized as: Sect. 14.2 gives the literature survey, Sect. 14.3 discusses post-COVID data analysis, Sect. 14.4 gives the discussion, and Sect. 14.5 gives the conclusions.

14.2 Literature Survey This section discusses the existing work done in the field of IoT, and data analytics for remote patient monitoring.

154

S. Rahman et al.

14.2.1 Internet of Things IoT-based systems are a connected network of devices and sensors such as smartphones, tablets, smart watches, EEC-ECG sensors, and so on, to exchange the data over the Internet. These devices are embedded with sensors that allows them to transmit the data. This method can be used for the patient monitoring scenarios for COVID and post-COVID conditions and some of the existing work done in this domain is mentioned here as: The authors in [7] propose an IoT-based system to monitor the quarantined patients. A predictive analysis is done in [8] on the downloaded dataset which are designed especially for COVID-19 quarantined patients. The authors in [9] propose a remote smart home healthcare support system (ShHeS) in which patients are remotely monitored by the healthcare professionals and even patient’s gets prescribed from the ease of their respective homes. In [10], a proposed model is designed for detecting hidden signatures of COVID19 by integrating biomedical sensors and AI. An IoT-based healthcare assistance system is proposed in [11] for COVID-19 patients. It is a survey-based research paper in which they have integrated multiple IoT associated technologies like remote monitoring, robot-based assistance, digital diagnostic, and so on which can be used for solving the problems faced during COVID-19 by the patients, health workers, and doctors. The authors in [12] propose an IoT-based conceptual architecture that focuses on scalability, interoperability, network dynamics, context discovery, reliability, and privacy issues faced in remote patient monitoring for the COVID-19 patients in hospitals and at home. An IoT-based system is proposed for safer movements in the pandemic using machine learning algorithms called SafeMobility that integrates IoT, fog, and cloud solutions to monitor the social distancing between people and control the capacity in them [13].

14.2.2 Data Analytics for Remote Patient Monitoring Medical data is growing exponentially and is very complex to store, process, and analyse insights from the big data by using traditional methods thus, advanced machine learning and deep learning techniques are used to analyse medical data efficiently. Healthcare data analytics is the process of analysing data for improved quality, decision making, and performance of healthcare organizations. In this way, the data is converted into precise, uniform, and timely information. Some of the existing work done in the field of ML and deep learning for the coronavirus is as: The authors in [14] conducted a detailed review focussing the utility of AI in the domain of imaging for COVID-19 patient care. A systematic meta-analysis is done for technical merit and clinical relevance of the collected data. The authors in [15] presented a classifier for COVID-19 detection which utilizes transfer learning for improving the performance and robustness of the DNN in vocal audio such as speed, cough, and breathe. The authors in [16] propose a ML-based model for the prediction

14 Medical Internet of Things and Data Analytics for Post-COVID Care: …

155

of anxiety during the COVID-19 pandemic in Saudi Arabia. The prediction models are developed by using support vector machine (SVM) classifier and J48 decision tree algorithm. The authors in [17] have proposed a method that classifies cough and is known as multi-criteria decision making (MCDM) method that uses ensemble learning techniques for COVID-19 cough classification. The authors in [18] propose an integrated method for the monitoring of COVID-19 by incorporating machine learning algorithms and IoT device. Data analysis for the post-COVID scenario can be seen helpful in extracting hidden patterns, insights, trends, and relationships. Researchers across the globe are working together from different domains in-order to have better understanding of the coronavirus. There has been work done in this domain as: The authors in [19] propose a meta-analysis on the hospitalized and non-hospitalized COVID recovered patients. The authors in [20] propose a review-based work that provides strategies to overcome long COVID. The authors in [21] propose a longitudinal analysis on 61 patients for eight months for the analysis of post-COVID symptoms, behaviour, and recovery on different patients.

14.3 Proposed Framework The proposed architectural design has five functional layers which are associated with each other, i.e. the output of one layer is taken as an input on the next layer and the process continues as shown in Fig. 14.2. The bottom is the data collection layer that provides the required data from the patients either through M-IoT devices, sensors, wearables, or through testing reports. The data gathered is sent to the next layer which is data aggregation layer at which the collected data gets combined/aggregated so that it can be pushed to the next layer for analysis. At the data analysis layer, the data collected is analysed by applying feature selection, feature engineering, and other data analytical techniques in-order to obtain the required relevant data. Once the pre-processing of the raw data is done, it is ready to go for the next layer which is modelling at which ML and deep learning algorithms are applied for prediction and the result obtained is sent to the last layer which is the patient monitoring layer at which the health conditions of the patient is monitored and examined remotely by the medical experts. The recommendations and suggestions provided by the experts are then sent to the patient as a feedback.

14.4 Post-COVID Data Analysis It is important to choose the correct data collection method as different tools are used for different research scenario. A dataset is a collection of related information that is gathered for a given problem. The dataset contains historical records or

156

S. Rahman et al.

Fig. 14.2 Architecture design

cases or observation and each case include numeric features that quantify a characteristic of an item. There are different methods available for collecting the data such as surveys, polls, experiments, reports, and so on. In the case of post-COVID patients, the availability of the data is very limited and the research is still ongoing. We have seen similarity of symptoms in COVID and post-COVID cases, however, exclusive datasets for post-COVID is a must, therefore, in our research, we have reached out directly to the PubMed and they have provided the required dataset. For the analysis purposes, we have utilized RapidMiner studio tool which is a data science platform developed by Ralf Klinkenberg, Ingo Mierswa, and Simon Fischer at Technical University of Dortmund in 2001. It is a fully automated data science tool that provides an integrated environment for machine learning, deep learning, data analytics, text mining, and predictive analytics [22]. The platform is used for commercial, business, educational, application development, and research purposes as it provides features such as data processing, optimisation, model validation, and results visualization. The preliminary analysis of the relevant features extracted from the dataset is discussed below. Figure 14.3 describes the age group of the patients in the dataset. The graph in the figure is plotted against different age groups vs. number of patients, i.e. the x-axis represents the three age groups and y-axis represents the total number of patients who are infected by the coronavirus. From the graph, it is clear that the age groups between 18 and 39 years have 164 infected patients, whereas the age group 40–64 has 205 infected patients and 65 and above possess 62 infected patients only. Thus,

14 Medical Internet of Things and Data Analytics for Post-COVID Care: …

157

we can conclude that 40–64 age group people are the most affected ones and they need to take more care. Figure 14.4 represents the status of recovery in patients. In the graph, the x-axis shows the recovery status as normal and not normal health status, whereas the y-axis represents the total number of patients. 320 patients recovered back to normal health, whereas 111 patients did not recovered. Significantly quite a large number of patients have recovered (Fig. 14.5).

Fig. 14.3 Age

Fig. 14.4 Recovery status

158

S. Rahman et al.

Fig. 14.5 Complications

The graph shows the status of complications in the patients as no or yes in the x-axis and the y-axis shows the total count of them. 354 patients have no complications, however, 77 patients possess complications and needs care. The graph is plotted against smoking status as non-smoker, ex-smoker, and smoker vs. total count. 245 people are non-smoker, 122 people are ex-smoker, and 61 people are smokers (Fig. 14.6).

Fig. 14.6 New symptoms

14 Medical Internet of Things and Data Analytics for Post-COVID Care: …

159

Figure 14.6 shows the presence of any new symptoms or ongoing symptoms in the patients at the acute phase of infection. The x-axis in the graph represents the status of any new symptoms present in a patient as yes or no, whereas the y-axis represents the total count in each of the categories. From the graph, 106 patients possess new symptoms/ongoing symptoms in the acute phase, however, 325 patients did not have any sign of new/ongoing symptoms at the acute phase of infection. Figure 14.7 demonstrates the recovery of symptoms in the infected patients. The Donut chart represents the symptom recovery status as recovered and not recovered. 38% of patients have not recovered, however, 61.5% of people have fully recovered. Thus, it is clear that majority of patients have recovered from their symptoms. Figure 14.8 demonstrates the status of fatigue in the patients. From the pie chart, it has been found that 45.3% of the people have no sign of fatigue in then, whereas 54.7% of people possess fatigue. Figure 14.9 demonstrates the presence of anxiety in patients. The graph represents the anxiety level in patients as no anxiety, mild to moderate anxiety, and severe to very severe anxiety. It has been found that 291 patients possess no anxiety, 103 possess mild to moderate, and 32 patients possess sever to very severe anxiety (Fig. 14.10). Figure 14.9 represents the status of existing ailments in the infected patients. The x-axis demonstrates the status of comorbidity as no or yes, whereas the y-axis represents the total count of the categories. 283 patients have been recorded for having no comorbidities in them, whereas 145 patients possess comorbidities (Fig. 14.11).

Fig. 14.7 Symptom recovery status

160

Fig. 14.8 Fatigue status

Fig. 14.9 Anxiety

S. Rahman et al.

14 Medical Internet of Things and Data Analytics for Post-COVID Care: …

161

Fig. 14.10 Comorbidity

Fig. 14.11 COVID complications

This figure represents the relation of medical complications experienced between the acute infection phase and completion of questionnaire to COVID-19 (medical evaluation = diagnosed by a physician; self-evaluated = attributed to COVID-19 by participant. The pie chart represents the status of COVID complications as unclear, medical evaluated, and self-evaluated. 14.3% of people have COVID-related complications which is self-evaluated, 35.1% have COVID-related complication who is medically evaluated, whereas 50.6% of people diagnosis are unclear (Fig. 14.12). The figure demonstrates the severity of symptoms in the patients during the acute phase. The x-axis shows the symptom severity as asymptomatic, mild to moderate,

162

S. Rahman et al.

Fig. 14.12 Symptom severity status

severe to very severe, and the y-axis shows the total number of patients. From the graph, 46 patients show asymptomatic condition, 221 patients possess mild to moderate severity of symptoms, and 164 patients possess severe to very severe symptoms and they need to take immense care (Fig. 14.13). This graph represents the presence of respiratory comorbidity in the patients. The x-axis shows the total number of patients, whereas the y-axis represents the status of respiratory comorbidity in the patients as yes or no. 388 patients have no respiratory comorbidity, whereas 33 patients possess respiratory comorbidity.

Fig. 14.13 Respiratory status

14 Medical Internet of Things and Data Analytics for Post-COVID Care: …

163

14.5 Discussion The post-COVID patient care has become more important because of the fact that many patients develop complications after their recovery from the COVID-19. There have been many deaths reported in the post-COVID era due to one or multiple organ failure which can be attributed to the unnoticed and unmonitored patients after COVID. Although, we have analysed only few parameters for the post-COVID scenario in our data analysis but the trend in that the patients develop complications in the post-COVID which is a matter of concern for all of us. It is, therefore, a need of the hour to identify the parameters which should be monitored continuously in the post-COVID era. The data has to be updated continuously and most importantly on regular basis. There should be data analysis to check the patients regularly and continuous if needed.

14.6 Conclusions Internet of Things provides a number of devices that sense the parameters necessary for the detection of symptoms in COVID and post-COVID scenarios. These include oximeter, fitness bands, ECG/EEG sensors, etc., that continuously monitor the patients. This leads to generation of enormous amount of medical data and processing of this data becomes a major challenge in such scenarios. Data analytics provides a viable solution to process the voluminous amount of monitored medical data. This paper discusses the role of IoT and data analytics in COVID/post-COVID scenarios. In this, the data analysis of post-COVID dataset is done using RapidMiner tool. The dataset is acquired directly from the PubMed which contains 42 attributes which are related to COVID and post-COVID scenarios such as fatigue, age, comorbidity, COVID complications, respiratory issues, and so on. The continuous IoT monitoring and data analysis provides an insight to understand the trends, patterns, and outliers in case of any abnormal symptom appears. This work can be intended to propose an IoT-based ML system for post-COVID care from family level to community level.

References 1. Coronavirus disease (covid-2019). https://www.who.int/health-topics/coronavirus#tab=tab_1 2. Post-Covid conditions (Long Covid). https://www.who.int/srilanka/news/detail/16-10-2021post-covid-19-condition 3. Episode 47, Post-Covid conditions. https://www.who.int/emergencies/diseases/novel-corona virus-2019/media-resources/science-in-5/episode-47---post-covid-19-condition 4. Post-Covid conditions. https://www.who.int/srilanka/news/detail/16-10-2021-post-covid-19condition

164

S. Rahman et al.

5. Episode 46, diabetes and covid-19. https://www.who.int/emergencies/diseases/novel-corona virus-2019/media-resources/science-in-5/episode-46---diabetes-covid-19 6. Raghupath, W., Raghupathi, V.: Big data analytics in healthcare: promise and potential. Health Inf. Sci. Syst. 2, 1–10 (2014) 7. Sharma, N.K., Gautam, D.K., Sahu, L.K., Khan, M.R.: First wave of covid-19 in India using IoT for identification of virus. Mater. Today Proc. (2021) 8. Hussain, S.A., Al Bassam, N., Zayegh, A., Al Ghawi, S.: Prediction and Evaluation of healthy and unhealthy status of COVID-19 patients using wearable device prototype data. MethodsX 101618 (2022) 9. Taiwo, O., Ezugwu, A.E.: Smart healthcare support for remote patient monitoring during covid19 quarantine. Inf. Med. Unlocked 20, 100428 (2020) 10. Hemamalini, V., Anand, L., Nachiyappan, S., Geeitha, S., Motupalli, V.R., Kumar, R., Ahilan, A., Rajesh, M.: Integrating bio medical sensors in detecting hidden signatures of COVID-19 with Artificial intelligence. Measurement 194, 111054 (2022) 11. Mukati, N., Namdev, N., Dilip, R., Hemalatha, N., Dhiman, V., Sahu, B.: Healthcare assistance to COVID-19 patient using internet of things (IoT) enabled technologies. Mater. Today Proc. (2021) 12. Paganelli, A.I., Velmovitsky, P.E., Miranda, P., Branco, A., Alencar, P., Cowan, D., Endler, M., Morita, P.P.: A conceptual IoT-based early-warning architecture for remote monitoring of COVID-19 patients in wards and at home. Internet of Things 18, 100399 (2022) 13. Yacchirema, D., Chura, A.: SafeMobility: An IoT-based System for safer mobility using machine learning in the age of COVID-19. Procedia Comput. Sci. 184, 524–531 (2021) 14. Born, J., Beymer, D., Rajan, D., Coy, A., Mukherjee, V.V., Manica, M., Prasanna, P., Ballah, D., Guindy, M., Shaham, D., Shah, P.L.: On the role of artificial intelligence in medical imaging of COVID-19. Patterns 2, 100269 (2021) 15. Pahar, M., Klopper, M., Warren, R., Niesler, T.: COVID-19 detection in cough, breath and speech using deep transfer learning and bottleneck features. Comput. Biol. Med. 141, 105153 (2022) 16. Albagmi, F.M., Alansari, A., Al Shawan, D.S., AlNujaidi, H.Y., Olatunji, S.O.: Prediction of generalized anxiety levels during the Covid-19 pandemic: a machine learning-based modeling approach. Inform. Med. Unlocked 28, 100854 (2022) 17. Chowdhury, N.K., Kabir, M.A., Rahman, M.M., Islam, S.M.S.: Machine learning for detecting COVID-19 from cough sounds: an ensemble-based MCDM method. Comput. Biol. Med. 145, 105405 (2022) 18. Deepa, N., Priya, J.S. and Devi, T.: Towards applying internet of things and machine learning for the risk prediction of COVID-19 in pandemic situation using Naive Bayes classifier for improving accuracy. Mater. Today Proc. (2022) 19. Fernández-de-Las-Peñas, C., Palacios-Ceña, D., Gómez-Mayordomo, V., Florencio, L.L., Cuadrado, M.L., Plaza-Manzano, G., Navarro-Santana, M.: Prevalence of post-COVID-19 symptoms in hospitalized and non-hospitalized COVID-19 survivors: a systematic review and meta-analysis. Eur. J. Int. Med. 92, 55–70 (2021) 20. Crispo, A., Bimonte, S., Porciello, G., Forte, C.A., Cuomo, G., Montagnese, C., Prete, M., Grimaldi, M., Celentano, E., Amore, A., de Blasio, E.: Strategies to evaluate outcomes in long-COVID-19 and post-COVID survivors. Infect. Agents Cancer 16, 1–20 (2021) 21. Aubry, A., Demey, B., François, C., Duverlie, G., Castelain, S., Helle, F., Brochot, E.: Longitudinal analysis and comparison of six serological assays up to eight months Post-COVID-19 diagnosis. J. Clin. Med. 10, 1815 (2021) 22. Intro to Rapidminer: a no-code development platform for data mining (with case study), https://www.analyticsvidhya.com/blog/2021/10/intro-to-rapidminer-a-no-code-develo pment-platform-for-data-mining-with-case-study/

Chapter 15

Mammography Image Classification and Detection by Bi-LSTM with Residual Network Using XG-Boost Approach Aman Chhabra and Manoranjan Rai Bharti

Abstract Early identification of breast cancer is crucial to increase the odds of a positive therapy, thus contributing decrease in mortality rate. Screen-based mammography and digitized mammography are two methods of mammography, out of which digitized mammography one is more secure. The proposed approach is based on ResNet-50 and Bi-LSTM network with extreme gradient boost (XG-boost), which resolves the problem of vanishing and exploding gradient, and contains the memory units in bi-direction making predictions easier than the existing approaches and provides an accuracy of 97.17% on the test set which is much higher than the accuracy of the existing approaches. It is vital to note that feature engineering is critical in this industry. Due to long short-term memory network (LSTM) economical memory, this research attempts to extract useful features using attention layers and to decrease error using the Bi-LSTM technique.

15.1 Introduction According to the latest statistics, breast cancer is found as the most oftenly diagnosed cancer among women worldwide [1–3]. Early identification is requisite to increase the chances of a positive therapy and, as a result, lowering down the fatality rates. Mammography, a screening procedure that uses cheap X-rays to reveal the diseased breast tissue, is by far the most widely used modality for the diagnosis of breast cancer [4–6]. A qualified medical specialist will normally analyze the resulting mammography. The gradual substitution of analog (film-based) mammography scanners with digitized mammogram scanners resulted in enhanced visible accessibility of a mammogram’s features. Even for skilled radiologists, obtaining a robust and reliable mammography categorization [7–10] remains a tough task. Complementary routines, such as double-reading (two separate experts marking the results) A. Chhabra (B) · M. R. Bharti National Institute of Technology Hamirpur, Hamirpur, Himachal Pradesh 177005, India e-mail: [email protected] M. R. Bharti e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_15

165

166

A. Chhabra and M. R. Bharti

Fig. 15.1 Procedure for obtaining a mammogram image [12]

or alternative modalities, such as biopsy, are a significant element of clinical practice for lowering the number of misclassifications. Mammography, oftenly known as a mammogram, is an important tool in the early identification and treatment of breast cancer. If the cancer is detected at early stage, the chances of a successful therapy increase. The mammography exam can be performed in two methods. The first is screen-based mammography, while the second is a digitized mammogram. Radiologists frequently employ both methods to identify breast cancer. However, the screen film-based mammography has several disadvantages, for example, the concealing effect of overlapping breast cells induced by projecting a 3D object over a 2D image. As a result, radiologists will not be able to provide a precise opinion on the presence of a malignant tumor [11, 12]. Asymptomatic women frequently receive false-positive results, while symptomatic women frequently receive false-negative results. As both false-negative and false-positive results can be hazardous, the radiologists must provide accurate results. As a result, in most hospitals, screen film-based mammography has been phased out in favor of digital mammography. The mammary image is acquired through a unique electronic X-ray indicator/detector in digital mammography, which turns the image into a computerized mammogram that can be viewed on a computer monitor and saved for further use. The procedure for obtaining a mammogram image is shown in Fig. 15.1.

15.1.1 Classification and Detection of Breast Masses The accompanying image processing methods are employed in digital mammograms for accurate and rapid mass identification, which enables radiologists to deliver correct diagnosis and treatment. The proposed technique was divided into different phases in this investigation in [13]. The very first stage involved detecting worrisome mass areas on mammograms, and the second phase involved classifying these areas as benign, probably benign, malignant, or probably malignant. Figure 15.2 illustrates the stages of a suggested technique.

15 Mammography Image Classification and Detection by Bi-LSTM . . .

167

Fig. 15.2 Schematic view for detecting and classifying breast masses using digital mammography as input

15.1.2 Pre-processing of Mammogram The procedure starts with the acquisition of digital mammography images at various sampling rates and quantization. To eliminate the noise from images, de-noising procedures are used, and afterward, contrast techniques are used to improve the mammography images. Numerous methodologies have been used at this level, including wavelet transforms, multi-resolution analysis, and region expanding techniques, among others.

15.1.3 Digital Mammogram Segmentation Segmentation is the practice of separating a picture into its component pieces or objects and then selecting the region of interests (RoIs) [14]. That is also the second key phase in mammography, where suspicious areas with masses are separated from ordinary parenchyma tissue. Separating the mammography picture in this manner breaks it into numerous non-overlapping sections, wherein the RoI is extracted. The suspected mass candidates are placed inside the designated RoI. Typically, suspicious regions are brighter than their surroundings and also exhibit a relatively uniform density, a uniform shape with variable sizes, and fuzzy edges.

168

A. Chhabra and M. R. Bharti

15.1.4 Mammogram-Based Feature Extraction It is a crucial stage inside the image processing approach that entails extracting picture attributes that statistically or subjectively reflect the objects in an image accurately. It is a high-level process for analyzing the shapes and designs of objects. Wavelet decomposition technique is used to extract features: The wavelet decomposition procedure is used to extract the mammography image features from selected RoI throughout this approach [15]. There seem to be five stages of processing: wavelet decomposition, co-efficient extraction, normalization, energy computation, feature reduction. Calculation of the roughness value for feature extraction: Breast cancer is detected in this work using a mammography scan. Each pixel’s roughness is determined [16]. By utilizing fractal analysis, the search zone is condensed. A region is considered to get a latent mass if indeed the computed roughness value drops between 2 and 3. Since they are not massless, all other sub-blocks with a roughness value of < 2 or > 3 must be eliminated.

15.1.5 Classification of Mammogram Detection of mass by applying acceptance/rejection criteria: Mammograms are classified according to whether the lump detected on the mammogram is malignant or benign. If indeed the mass is benign, the mammography is deemed normal. When mammography is abnormal due to a malignant tumor, it is termed to be abnormal. The following procedure aids in determining if the mass is benign or cancerous. Figure 15.3 shows the consequences of acceptance/rejection criteria application. Table 15.1 shows the different stages of criteria for acceptance and rejection of the mammogram.

15.2 Related Works Fam et al. [13] describe a unique method for automatically segmenting and classifying masses in mammograms to assist radiologists in making an accurate diagnosis. To begin, a mixture of many image enhancement methods was studied, namely guided imaging, median filtering, and contrast limited adaptive histogram equalization (CLAHE) to improve the visual aspects of the breast area and improve segmentation accuracy. Secondly, the densities of discrete wavelet coefficients density were developed as a method for detecting suspicious mass regions using the quincunx lifting methodology. Punithavathi and Devakumari [1] have discussed breast cancer

15 Mammography Image Classification and Detection by Bi-LSTM . . .

169

Fig. 15.3 Consequences of acceptance/rejection criteria application [12]

grading in digital mammography images. Digital image processing is extremely beneficial in the medical field, particularly when it comes to identifying and classifying mammography pictures. Breast cancer classification in mammography images is a critical stage in determining if a patient has malignant or non-cancerous masses. Kathale and Thorat [11] discussed many ways and procedures utilized to identify and segment the mammographic image’s RoI and anomalies. Dabass [14] provided a detailed review of these techniques, particularly for mammogram images. Vijayarajeswari et al. [7] described the categorization of mammograms using Hough transform-extracted features. The Hough transform is a two-dimensional transformation. This is used to segregate an image’s aspects of a specific shape. Miniaturized scale characterization and masses seem to be the two most critical threat indicators, and their automated detection is critical for early breast cancer detection. Jouirou et al. [4] gave a detailed overview of current MVIF breast cancer detection systems. It presented a standard hierarchical classification of mammography MVIF techniques. Ribli et al. [5] offered a CAD system that is based on the most effective object detection framework, faster recurrent-CNN (R-CNN), that automatically finds and classifies benign or malignant mammography abnormalities. The suggested technique achieves the best performance of the classifier on the publicly available INbreast database, with an AUC of 0.95. With an AUC of 0.85, the technique reported here placed second in the Digital Mammography DREAM Challenge. Sheba and Gladston [6] established an accurate classification technique for digital mammograms that distinguishes between healthy, benign, and cancerous breast parenchyma. The purpose of this article is to describe a computer-aided diagnosis system for the automatic identification and treatment of breast cancer using digital mammograms. Alkhaleefah and Wu [8] explained the idea of transfer learning, which entails utilizing the CNN’s power as a feature extractor to aid in identifying malignant from benign breast cancer images. Additionally, by changing the kernel width, the support vector machine (SVM) classifier relying on radial basis function (RBF) was tailored for its adaptability in fitting the data dimensional space suitably. The combination of CNN and RBF-based SVM generated strong results both for the dataset and application task

170

A. Chhabra and M. R. Bharti

Table 15.1 Acceptance and rejection criteria Criteria 1

Criteria 2

Criteria 3

The mass’s area is between 900 and 5000 pixels, which is worrisome. Regions that do not match this criterion are disqualified If the third-order moment (skewness) of any surviving region is negative, it is considered suspicious; otherwise, it is discarded If only the mean intensity surpasses a threshold level, then the remaining location stays suspicious. Regions that do not achieve this requirement will be rejected

used in this study. Mohamed and Salem [9] developed an automated approach for mammography mass categorization. The suggested method is comprised of three major steps. To begin, three distinct sorts of characteristics are distinguished from the mass. Then, using the T-test technique, the most significant features are considered. Finally, the classification phase is performed to differentiate benign or malignant masses using three classifiers: SVM, KNN, and ANN. Danala et al. [10] studied how and where to optimally build a computer-aided design (CAD) framework of contrast enhanced digital mammography (CEDM) images for breast mass classification. We constructed a CEDM database of 111 patients, 33 of whom were benign and 78 of whom were malignant. Each CEDM has two distinct sorts of images: dual-energy subtracted (DES) and low energy (LE). A CAD approach was used to segregate mass regions depicted independently on DES and LE images. Additionally, optimal segmentation findings obtained from DES images were projected to LE images or vice-versa.

15.3 The Proposed Method Bi-LSTM although employs a variety of algorithms to calculate the hidden information, and a long short-term memory (LSTM) is trustworthy when used with the standard recurrent neural network (RNN) structure. It overcomes the limitation of recurrent ANNs in managing long-distance dependency. An LSTM approach entails many memory units, each of which has three gates with a variety of functions. The next subsection provides settings for the criteria of the LSTM unit of sth patches using the feature vector S as input and the sth patches as an example. A specific calculation equation is used, which denotes the sigmoid function and denotes dot multiplication. The X t shows the forget gate: X t = α(Wx ws + Ux h s−1 + biasx ) The js is the input gate:

(15.1)

15 Mammography Image Classification and Detection by Bi-LSTM . . .

js = α(W j w j + U j h s−1 + bias j )

171

(15.2)

The u s variable represents the status of the candidate memory cell at the most recent time step, where tanh signifies the tangent hyperbolic function. u s = tanh(Wu ws + U j h s−1 + biasu )

(15.3)

The u s specifies the memory cell’s most recent state value. The value of X t was between 0 and 1. Dataset In this work, we have used the following datasets. Mammographic Image Analysis Society (MIAS) This dataset had 322 digital mammography. Digital Database for Screening Mammography (DDSM) This database, which was created in 1999, initially had 10,480 digital film mammography images.

15.3.1 Proposed Methodology: Flowchart The proposed methodology for mammography classification and detection can be described through the following steps: Step 1: In the first step, extract the dataset and preprocess it in pixel form. After preprocessing, apply segmentation using the k-means clustering approach. Step 2: After segmentation, extract features and apply sequence-based feature mapping Bi-LSTM. Step 3: Extract features input in Residual network 50, divided into five blocks and 10 CNN layers after non-linear mapping by the residual network. Step 4. Apply extreme gradient boost (XG-boost) base classifier and analyze the performance of the proposed model. The flowchart of the proposed methodology is given in Fig. 15.4.

15.4 Results and Discussion In this section, we present and discuss our results for the existing and proposed approaches. Figure 15.5 presents the confusion matrix of the existing approach representing the true (actual) label versus the predicted label. From the figure, we observe that: True-positive: 1482 are correctly predicted by the model.

172

A. Chhabra and M. R. Bharti

Fig. 15.4 Proposed methodology flowchart Fig. 15.5 Confusion matrix of existing approach

False-positive: 237 are wrongly predicted by the model. False-negative: 557 are wrongly predicted by the model. True-negative: 829 are correctly predicted by the model. Figure 15.6 presents the confusion matrix of the proposed approach representing the true (actual) label versus the predicted label. From this figure, we observe that: True-positive: 412 are correctly predicted by the model. False-positive: 6 are wrongly predicted by the model.

15 Mammography Image Classification and Detection by Bi-LSTM . . .

173

Fig. 15.6 Confusion matrix of proposed approach

Table 15.2 Comparison table for existing and proposed approach Approach Accuracy Precision Recall F-score Existing Proposed

74.43 97.17

74.95 97.19

74.43 97.17

73.23 97.17

ROC 83.42 99.89

False-negative: 16 are wrongly predicted by the model. True-negative: 343 are correctly predicted by the model. From Figs. 15.5 and 15.6, we can see that the proposed approach has less falsepositive and false-negative rates than the existing approach. Table 15.2 summarizes the accuracy, precision, recall, F-score, and ROC of the proposed and the existing approaches. The proposed approach obtains an accuracy of 97.17, which is significantly higher than the existing-VGG method’s accuracy of 74.43. Similarly, the precision, recall, F-score, and ROC each have a value of 97.19, 97.17, 97.17, and 99.89, respectively. According to the values given, the proposed strategy outperforms the existing approaches in all circumstances. Figure 15.7 presents accuracy and loss curves for the existing approach. The existing approach suffers from the problem of vanishing and exploding gradient. Hence, the accuracy and loss function get saturated after very few epochs. We calculate our results on 10 epochs because validation loss gets saturated after 10 epochs. The existing approach provides an accuracy of 74.43% on the test set. In Fig. 15.7, we can see that the model starts strongly overfitting after 6 epochs. Figure 15.8 presents accuracy and loss curves for the proposed approach. The proposed approach is based on ResNet-50 and Bi-LSTM network with XG-boost, which resolves the problem of vanishing and exploding gradient, and contains the memory units in bi-direction making predictions easier than the existing approach. We calculate our results on 30 epochs because validation loss gets saturated after

174

Fig. 15.7 Existing approach accuracy and loss

Fig. 15.8 Proposed approach accuracy and loss

A. Chhabra and M. R. Bharti

15 Mammography Image Classification and Detection by Bi-LSTM . . .

175

30 epochs. The proposed approach provides an accuracy of 97.17% on the test set which is much higher than the accuracy of the existing approach. The fluctuations around the curves are because of dropout layers and changes in learning rate during the training and validation process. In Fig. 15.8, we can see that the model starts strongly overfitting after 20 epochs.

15.5 Conclusion The use of biomedical imaging for the early identification and categorization of breast cancer is a critical part of medical diagnosis. Already existing methods of mammography have large false-positive and false-negative rates. The existing approaches suffer from the problem of vanishing and exploding gradient. The proposed approach is based on ResNet-50 and Bi-LSTM network with XG-boost, which resolves the problem of vanishing and exploding gradient, and contains the memory units in bidirection making predictions easier than the existing approaches and provides better results than the existing CNN models. Our suggested feature engineering methodology and machine learning-based ensemble classifier improved on DDSM and MIAS datasets by an average of 22.57% in precision, recall, and accuracy.

References 1. Punithavathi, V., Devakumari, D.: A framework on classification of mammogram images for breast cancer detection using image processing with data mining techniques. Int. J. Creative Res. Thoughts 8(2) (2020) 2. Al-Antari, M.A., Al-Masni, M.A., Choi, M.-T., Han, S.-M., Kim, T.-S.: A fully integrated computer-aided diagnosis system for digital X-ray mammograms via deep learning detection, segmentation, and classification. Int. J. Med. Inform. 117, 44–54 (2018) 3. Sinzinger, F.: Mammography Classification and Nodule Detection Using Deep Neural Networks (Dissertation) (2017) 4. Jouirou, A., Baâzaoui, A., Barhoumi, W.: Multi-view information fusion in mammograms: a comprehensive overview. Inf. Fusion 52, 308–321 (2019) 5. Ribli, D., Horváth, A., Unger, Z., Pollner, P., Csabai, I.: Detecting and classifying lesions in mammograms with deep learning. Sci. Rep. 8(1), 1–7 (2018) 6. Sheba, K.U., Gladston Raj, S.: An approach for automatic lesion detection in mammograms. Cogent Eng. 5(1), 1444320 (2018) 7. Vijayarajeswari, R., Parthasarathy, P., Vivekanandan, S., Basha, A.A.: Classification of mammogram for early detection of breast cancer using SVM classifier and Hough transform. Measurement 146, 800–805 (2019) 8. Alkhaleefah, M., Wu, C.-C.: A hybrid CNN and RBF-based SVM approach for breast cancer classification in mammograms. In: 2018 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pp. 894–899. IEEE (2018) 9. Mohamed, B.A., Salem, N.M.: Automatic classification of masses from digital mammograms. In: 2018 35th National Radio Science Conference (NRSC), pp. 495–502. IEEE (2018)

176

A. Chhabra and M. R. Bharti

10. Danala, G., Patel, B., Aghaei, F., Heidari, M., Li, J., Wu, T., Zheng, B.: Classification of breast masses using a computer-aided diagnosis scheme of contrast enhanced digital mammograms. Ann. Biomed. Eng. 46(9), 1419–1431 (2018) 11. Kathale, P., Thorat, S.: A review on methods utilized for classification of mammographic image. SAMRIDDHI J. Phys. Sci. Eng. Technol. 12(SUP 3), 99–102 (2020) 12. Priyanka, B.B., Kulkarni, D.A.: Digital mammography: a review on detection of breast cancer. Int. J. Adv. Res. Comput. Commun. Eng. 5(1), 386–390 (2016) 13. Fam, B.N., Nikravanshalmani, A., Khalilian, M.: An efficient method for automated breast mass segmentation and classification in digital mammograms. Iran. J. Radiol. 18(3) (2021) 14. Dabass, J., Arora, S., Vig, R., Hanmandlu, M.: Segmentation techniques for breast cancer imaging modalities—a review. In: 2019 9th International Conference on Cloud Computing, Data Science and Engineering (Confluence), pp. 658–663. IEEE (2019) 15. AbuBaker, A.A.: Mass lesion detection using wavelet decomposition transform and support vector machine. Environment. Int. J. Comput. Sci. Inf. Technol. (IJCSIT) 4(2), 33–46 (2012) 16. Rejani, Y., Selvi, S.T.: Breast cancer detection using multilevel thresholding (2009). arXiv preprint arXiv:0911.0490

Chapter 16

ECG Biometric Recognition by Convolutional Neural Networks with Transfer Learning Using Random Forest Approach Ankur and Manoranjan Rai Bharti Abstract Biometric authentication is amongst the most promising techniques for precise and automated recognition, which increases probability of identifying personnel with higher accuracy. Physiological and behavioural biometrics are the two primary categories of biometrics out of which physiological one is more secure. Electrocardiograms (ECG) biometric identification is still plagued by issues of generality and efficiency. In this paper, ECG of different individuals have been used to study biometric identification. A new convolutional neural network (CNN) technique has been proposed in this study to achieve human identification by ECG biometric identification with parameter tuning by transfer learning using an ensemble classifier, and thus, achieved an improvement in accuracy upto 10–15%.

16.1 Introduction Authentication or Identification of any person is a crucial part in high-security areas like labs, military areas, and many other places. Physiological and behavioural biometrics are the two primary categories of biometrics out of which physiological one is more secure. Behavioural biometrics include things like voice, gait, retina, fingerprint, face, and so on, whereas physiological biometrics comprise of things like electrocardiogram (ECG), electroencephalogram (EEG), and so on [1–3]. In human identification jobs, bio-potentials are becoming increasingly powerful and crucial. Already existing authentication systems like finger-print scanning, PIN code, iris scan, face scan can be falsified easily. Recent researches suggest that the ECG signal, which would be a critical component of clinical diagnostics, could be exploited as a novel biometric modality.

Ankur (B) · M. R. Bharti National Institute of Technology Hamirpur, Hamirpur, Himachal Pradesh 177005, India e-mail: [email protected] M. R. Bharti e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_16

177

178

Ankur and M. R. Bharti

16.1.1 ECG Based Biometric Systems The ECG biometrics signal of a typical heartbeat comprises distinct features such as T waves, P waves, and QRS-complexes. ECG biometrics is highly promising for human-based identification because of these appealing signal properties. The depolarization of the left and right atria is reflected in P-wave. It usually has a positive amplitude and appears to be smaller than 120 ms in duration. Depolarization of left and right ventricles is reflected by the QRS-complex, the QRS-complex remains for approximately 70–110 ms for a regular heartbeat. In the ECG waveform, it attains the largest amplitude value. It has a piece of greater frequency information than other ECG waves because of its steep slope, being riveted within the range between 10 and 40 Hz. The T -wave, which occurs at 300 ms after the QRS-complex, signifies the ventricular re-polarization. The T -wave’s position is completely reliant on pulse rate, this becomes thinner and nearer to the QRS-complex as the heart rate increases.

16.1.2 Biometric Identification System Based on Deep-ECG Six fiducial points are typically present in a normal ECG signal (Fig. 16.1). The QRScomplex is made of three points: Q, R, and S, each letter denotes a specific occurrence. Because the QRS-complex seems to be the region of electrocardiogram that would be less impacted by fluctuations owing to physical activity or emotional reactions, analysing it for biometric activities can be beneficial. As a result, Deep-ECG solely analyses the QRS-complex. Deep-ECG has three applications: (a) identity verification, (b) closed-set authentication, and (c) re-authentication on a regular basis. Deep-ECG Biometric Identification Procedures Signal Pre-processing: In this stage, [4] enhances the ECG signals and extracts m QRS-complexes that are most distinguishable. The noise can be eliminated by applying a Butterworth (high-pass) filter of third-order and a notch IIR filter with a 0.5 Hz cutoff frequency. An automatic labelling technique is used in the second stage to calculate the location of n R points (fiducial). Calculate the cross-correlation amongst QRS and each QRS-complex of H to produce the vector C. Vector V can be formed by combining the m QRS-complexes together with the highest possible data of vector-C. A feature-based vector V in operation is shown in Fig. 16.3 (Fig. 16.2). ECG CNN Feature-Extraction: The features calculated with the CNN are analysed by the Deep-ECG to use a Soft-max layer to return the client’s identity during training and the closed-set identification process. Architecture of CNN: The biometric verification process and structure of CNN utilized by DL-ECG is delineated in Figs. 16.2 and 16.4 respectively. It consists of 3 max-pooling layers, 6 convolution layers using rectified linear unit (ReLU) neurons, 3 Local Responses Normalization (LRN) levels, 1 dropout layer, 1 Soft-max layer

16 ECG Biometric Recognition by Convolutional Neural Networks . . .

Fig. 16.1 Heart’s ECG [3]

Fig. 16.2 Biometric verification process

179

180

Ankur and M. R. Bharti

Fig. 16.3 V represents a feature vector that contains m QRS-complexes. The vector V is consumed by CNN [4]

Fig. 16.4 CNN employed as a feature extractor to create biometric templates for re-authentication and identity verification periodically [4]

(for closed-set identification and training), and one fully-connected layer. ReLU neurons allow for the addition of nonlinearity for the network, hence favouring a more detailed depiction. LRN layers are beneficial when accompanied by ReLU neurons due to their boundless outputs. LRN enables the detection of high-frequency components with a high number of activations concerning their vicinity. Additionally, we employ dropout regularization to enhance generality and prevent co-evolution by arbitrarily resetting a part of activation functions to 0. The Soft-max layer appear as final layer of the network that would be utilized to train the CNN and carry out closed-set recognition. The same layer produces an integer tag according to identity of the user. Given a training set of samples from n u users, the CNN output is an integral value [1, 2, . . . n u ]. CNN layers use biases b to control the incoming signal x before convolutioning it using K filters stack f. Consequently, we get a signal symbolized mostly by symbol y as an output signal. x R H W D ,



f  RH W



D D 

, y R H



W  D 

(16.1)

Here D, W and H denote the dimensions of depth, breadth, and height. This is basic setup of convolutional layer, the output is calculated as follows for every coordinate (i, j, d):

16 ECG Biometric Recognition by Convolutional Neural Networks . . . 

yi  j  d  = bd  +

181



W  D H  

f i  j  d × xi  +i  −1, j  + j  −1,d,d 

(16.2)

i  =1 j  =1 d=1

In the majority of layers, it is required to pad the input signal x or to subsample the output signal. One can evaluate (Ph− , Ph+ , Pw− , Pw+ ) top-bottom-left-right paddings and strides in specifically (Sh , Sw ). In such circumstances, the output is usually calculated by using the following (with x implicitly extended with zeros as needed): 

yi  j  d  = bd  +



W  D H   i  =1 j  =1 d=1

f i  j  d × x Sh (i  −1)+i  −Ph− S|w( j  −1)+ j  −Pw−

(16.3)

The pooling layers minimize the feature space by employing the max-pooling operator. This operation determines the maximum response in an H  W  patch for every featured channel that uses the following procedure: yi  j  d  =

max

1≤i  ≤H  ,1≤ j  ≤W 

x|i  + i  − 1, j  − 1, d

(16.4)

The following operator is used by the LRN layers: ⎛ yi  j  d  = xi jk ⎝k + α



⎞−β xi2jt ⎠

(16.5)

tG(k)

where G(k) ⊂ {1, 2, 3, 4, . . . , D} is an equivalent input channels collection for every output channel k. In Deep-ECG, all LRN layers employ the values k = 1, D = 5, α = 2 × 10−4 and β = 0.75. An activation function is employed via ReLU layers that are not saturating, yi = max(0, xi ). The output of the Soft-max classifier is, ex j yi = n i=1

exi

(16.6)

Here, the total number of inputs to a neuron is represented by n. The length of the convolution kernels, number of layers, the stride, and the size of the max-pooling kernels were all empirically optimized. Training of CNN: CNN demands a huge quantity of samples in order to create a generalized model capable of performing a biometric identification. In order to increase the trained samples quantity, signals from distinct leads have been examined [4]. This is a process that is frequently used in conjunction with several other biometric modalities, like fingerprints and iris scans.

182

Ankur and M. R. Bharti

16.2 Related Works Alduwaile and Islam [2] have studied how a brief section of ECG signal can be efficiently utilized for biometric identification through the application of deep learning methods. A small CNN is used to raise the entropy of a brief section of a cardiac signal to improve generalization capability. Furthermore, it examines how distinct blind and feature-dependent parts of varying lengths affect the recognition system’s performance. Alduwaile and Islam [1] have presented a method for biometric identification of an individual using a solitary heartbeat ECG signal and a DL methodology. The authors examine the categorization of single heartbeat ECG signals separated depending upon R-peak and modified using wavelet compression (continuous) using a pretrained and light convolutional neural network. Bogdanov et al. [3] considered that biometric data is extremely difficult to forge and it seldom changes over time. However, widely used biometric identification systems like retina recognition, voice recognition, and fingerprinting, all have flaws. As a result, the authors have highlighted the most essential data in this regard. Li et al. [5] offered a different generic CNN technique, dubbed Cascaded CNN, for person recognition using electrocardiogram biometrics. Two CNNs are incrementally trained in our method. The very first CNN, fourier convolutional neural network, is used to draw out the features using ECG heartbeats, and the other, M-CNN, is utilized for biometrics classification. Cascading the M-CNN with F-CNN (trained) forms the Cascaded CNN, which serves as the ultimate identifying network. Kim et al. [6] presented an electrocardiogram signal’s two-dimensional coupling image-based personal identification system. The suggested technique feeds the network with the two-dimensional coupling image formed from three phases of ECG signal. The network is built on a CNN optimized for image processing. The waveforms of the two-dimensional coupling picture that serves as the network’s input data cannot be visually validated, but it has the benefit of augmenting the QRS-complex, which is a personally unique piece of information. Bajare and Ingale [7] suggested using one-dimensional convolutional neural networks to identify electrocardiogram signals for persons’ biometric recognition. The suggested system is designed to identify electrocardiogram signals using leads I and II separately, or a combination of leads (I and II). A 10-layer convolutional neural network model is created using an activation function (ReLU) to boost system’s execution. Wu et al. [4] offered a method for PIV, i.e. Personal Identity Verification based on two-dimensional convolutional neural networks using ECG signals. Presently, CNNs do exceptionally well in the area of image identification; to maximize this benefit, researchers transform electrocardiogram signals to 2D grayscale rather than the standard ECG. Guven et al. [8] described a novel fingertip ECG data gathering device that records the lead-I electrocardiogram signal via the thumb fingers (right and left). The suggested design is very sensitive, portable, dry contact, economical, and user friendly as it doesn’t need the use of bulky and unpleasant parts and components, like wet and sticky Ag/AgCl electrodes. Pinto et al. [9] examined unresolved difficulties and offer intense competition and applicable ECG biometric technology by capitalizing on present prospects. Hammad et al. [10] created two

16 ECG Biometric Recognition by Convolutional Neural Networks . . .

183

authentication mechanisms using two distinct level fusion algorithms: fusion algorithm based on feature and algorithm based on decision. CNN is used to extract features for individual modalities. From CNN, two layers are chosen by the authors which offered maximum accuracy on this step, with every layer acting like a distinct feature-descriptor. Then, they integrated them to create biometric templates using the suggested internal fusion technique. Abdeldayem and Bourlai [11] provided a novel strategy for establishing personal recognition systems that leverages the spectrotemporal dynamic properties of the electrocardiogram signal. They employ both generalized Morse Wavelets (CWT) and Short-time Fourier transform (STFT). This method adds to the information retrieved from the original electrocardiogram signal, which is then put into a 2D-convolutional neural networks that recovers subjectspecific and higher-level ECG-based features for every individual. da Silva Luz et al. [12] demonstrated how to derive meaningful representations for cardiac biometrics detection using deep learning approaches, notably convolutional networks. The authors specifically studied feature representation learning for the heart-biometrics using two main resources: heartbeat spectrogram and actual heartbeat signal.

16.3 The Proposed Method The proposed method is based on random forest which uses the advantage of the ‘wisdom of the crowd’. We firstly use CNN to take out the features using the dataset and then we pass it to Random Forest Algorithm. It uses the output from multiple different decision trees as a ‘vote’. Depending on the majority votes, the final class is decided as the output or predicted class.

16.3.1 Proposed Methodology: Flowchart The proposed methodology can be described by using the following steps and the flow chart given in Fig. 16.5. Step-1: Input the ECG dataset with a different type of signal and classes. The first step is pre-processing and generating features. Step-2: After combining all instances and making a matrix, input this matrix for CNN and nonlinear mapping. Step-3: Parameters of signal features tuned by transfer learning at nonlinear space. After feature mapping, apply random forest for training. Step-4: After training, apply testing and analyse the model performance. Dataset: In this work, Fantasia database is used which contains electrocardiogram (ECG) signals of young (21 years old to 34 years old ) and aged (68 years old to 85 years old) persons. Each subgroup of subjects has equal number of men and women.

184

Ankur and M. R. Bharti

Fig. 16.5 Proposed methodology flowchart

16.3.2 Transfer Learning Algorithm The features extracted from CNN are passed on to Random Forest algorithm for classification purpose using the following steps: Input: ECG dataset with labels Output: Classifier signals 1. 2. 3. 4.

Initialize Parameter W = W0 and θ = θ0 For Each Image, do forward pass of CNN on 3D-images. Define the layer-wise activation function Calculate the layer-wise patches, according to signal’s calculated cost-function for all signals φl[a,b,c] < −φ [a,b,c]l−1 + α[π [a,b,c]l−1 ] (16.7)

where, α < Learning value (0 to 1) 5. For all patches apply transfer learning field Y < −argminY E(Y, x n ; ω, θ ) − C(θ ; I i , ci )

(16.8)

16 ECG Biometric Recognition by Convolutional Neural Networks . . .

185

Y is a finite mapping of transfer learning. E(Y, x n ; ω, θ ) is pixel-wise optimization, C(θ ; I i , ci ) is neighbour signals. If Y = Y n , then, add. 6. Update CNN weights

θ < −W + δ (a,b,c)

(16.9)

7. Analysis by a trained classifier.

Table 16.1 Comparison table for proposed and existing approach’s performance Approach Accuracy Precision Recall F-score Proposed Existing

89.23 72.32

75.22 73.32

Fig. 16.6 Existing approach’s confusion matrix

89.12 74.43

75.35 69.23

186

Ankur and M. R. Bharti

16.4 Results and Discussion The different performance metrics that are used to check the efficiency of an algorithm are Accuracy, Precision, Recall, and F-measure. Table 16.1 summarizes the F-score, recall, precision, and accuracy of the proposed and the existing approaches. The proposed approach obtains an accuracy of 89.23%, which is significantly higher than the existing method’s accuracy of 72.32%. Similarly, the precision, recall, and F-score each have a value of 75.22, 89.12, and 75.35, respectively. According to the values given, the proposed strategy outperforms the existing approaches in all circumstances. Figures 16.7 and 16.6 illustrate the proposed approach’s confusion matrix and existing approach’s confusion matrix respectively. Figures 16.8 and 16.9 show existing approach’s (i.e. using CNN) accuracy and loss curves, respectively. In Fig. 16.9 we can observe that training loss is decreasing with epochs in the existing approach but if we observe the validation loss, it decreases in the beginning and after 3 epochs, it starts increasing. This is when the model starts to overfit. Overfitting can be handled by getting more training data, lowering the capacity of the model (by removing hidden layers) to memorize the training data, and using dropout layers, which will remove certain features randomly by setting them to zero. But in all the above-suggested

Fig. 16.7 Proposed approach’s confusion matrix

16 ECG Biometric Recognition by Convolutional Neural Networks . . .

187

Fig. 16.8 Accuracy versus epochs curve of existing approach

Fig. 16.9 Loss versus epochs curve of existing approach

solutions, the accuracy of the existing model reduces whilst working with large datasets. Figures 16.10 and 16.11 illustrate the proposed approach’s accuracy and loss curves, respectively. The proposed approach has a high accuracy of 89.23% whereas the existing approach produced an accuracy of 72.32% on the same large dataset.

188

Ankur and M. R. Bharti

Fig. 16.10 Proposed approach’s accuracy versus epochs curve

Fig. 16.11 Proposed approach’s loss versus epochs curve

16.5 Conclusion Recent researches have shown that ECG of a person is unique and can be used for authentication. CNNs have shown notable outcomes in the field of authentication using ECG of an individual. But, these do not work well when it comes to large dataset, new entries or features, etc. The proposed approach is based on Random Forest algorithm and produces better results than the existing CNN models with a large dataset. Extracted features from CNN are passed to Random Forest algorithm.

16 ECG Biometric Recognition by Convolutional Neural Networks . . .

189

Random forest uses the advantage of the ‘wisdom of the crowd’ and thus produces better results than the already existing method. Also, the proposed algorithm scales well when the new features and samples are added to the dataset.

References 1. Alduwaile, D., Islam, M.S.: Single heartbeat ECG biometric recognition using convolutional neural network. In: 2020 International Conference on Advanced Science and Engineering (ICOASE), pp. 145–150. IEEE (2020) 2. Alduwaile, D.A., Islam, M.S.: Using convolutional neural network and a single heartbeat for ECG biometric recognition. Entropy 23(6), 733 (2021) 3. Bogdanov, M., Filippova, A., Shakhmametova, G., Oskin, N.N.: Biometric authentication based on electrocardiogram. In: Biometric Systems. IntechOpen (2020) 4. Wu, J., Liu, C., Long, Q., Hou, W.: Research on personal identity verification based on convolutional neural network. In: 2019 IEEE 2nd International Conference on Information and Computer Technologies (ICICT), pp. 57–64. IEEE (2019). https://doi.org/10.1109/INFOCT. 2019.8711104 5. Li, Y., Pang, Y., Wang, K., Li, X.: Toward improving ECG biometric identification using cascaded convolutional neural networks. Neurocomputing 391, 83–95 (2020) 6. Kim, J.S., Kim, S.H., Pan, S.B.: Personal recognition using convolutional neural network with ECG coupling image. J. Ambient Intell. Humanized Comput. 11(5), 1923–1932 (2020) 7. Bajare, S.R., Ingale, V.V.: ECG based biometric for human identification using convolutional neural network. In: 2019 10th International Conference on Computing, Communication and Networking Technologies (ICCCNT), pp. 1–7. IEEE (2019) 8. Guven, G., Gürkan, H., Guz, U.: Biometric identification using fingertip electrocardiogram signals. Sig. Image Video Process. 12(5), 933–940 (2018) 9. Pinto, J.R., Cardoso, J.S., Lourenço, A.: Evolution, current challenges, and future possibilities in ECG biometrics. IEEE Access 6, 34746–34776 (2018) 10. Hammad, M., Liu, Y., Wang, K.: Multimodal biometric authentication systems using convolution neural network based on different level fusion of ECG and fingerprint. IEEE Access 7, 26527–26542 (2018) 11. Abdeldayem, S.S., Bourlai, T.: ECG-based human authentication using high-level spectrotemporal signal features. In: IEEE International Conference on Big Data (Big Data), pp. 4984– 4993. IEEE (2018) 12. da Silva Luz, E.J., Moreira, G.J., Oliveira, L.S., Schwartz, W.R., Menotti, D.: Learning deep off-the-person heart biometrics representations. IEEE Trans. Inf. Forensics Secur. 13(5), 1258– 1270 (2017)

Chapter 17

Design and Analysis of a Metamaterial-Based Butterworth Microstrip Filter Samiran Pramanik, Kamlendra Kumar Rathour, Edhot Chakma, and Chaitali Koley Abstract The creation of a Butterworth low pass filter (LPF) employing a metamaterial structure is presented in this study. The metamaterial structure improves the attenuation of a standard microstrip filter. In this study, a square uni-planer complementary split ring resonator is used as a form of unit cell in a metamaterial structure, and it is etched on the ground plane of a microstrip line. The fifth order LPF design is attained at the cut-off frequency of 3.5 GHz with attenuation of 20 dB. The proposed structure is fabricated and measured results are matched with simulation ones. The proposed design is useful in the microwave communication field.

17.1 Introduction The current era of wireless communication [1], with its growing demand, has increased the pressure on component quality in this frequency range. Due to the increased need for bandwidth, the majority of communication systems use microwave frequencies, and filters [2–4] are the most important and essential components for having high-quality frequency selective devices.

S. Pramanik (B) · K. K. Rathour · E. Chakma · C. Koley NIT Mizoram, Aizwal 796012, India e-mail: [email protected] C. Koley e-mail: [email protected] S. Pramanik College of Engineering and Management, Kolaghat, Purba Medinipur 721171, India © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_17

191

192

S. Pramanik et al.

Low pass filters (LPFs) were investigated in this study because they play an important role in microwave systems. Open stubs or stepped impedance microstrip lines [5–10] are commonly used in traditional LPF implementations. However, because these structures have a progressive rather than a sudden cut-off response, the rejection characteristic is limited in such LPFs. This limitation can be overcome by raising the passband insertion loss by adding new sections, however, this would increase the structure’s size. Metamaterial structures such as CSRR are increasingly being used to address these issues. Metamaterials (MTMs) [11–17] are defined as artificial, effectively homogeneous materials with cell sizes p and g that are significantly smaller than the guided wavelength λg and exhibit highly unusual properties such as negative effective permittivity r and permeability r that are not readily available in nature. Veselogo is the first to introduce left-handed materials (LHMs). In this paper, we have presented a fifth order Butterworth LPF with cut-off frequency of 3.5 GHz. Furthermore, the copper made ground plane is etched by using complementary double split ring resonator to enhance the attenuation from 20 to 60 dB. At the end, numerical and experimental results are compared to verify the effectiveness of the proposed structure.

17.2 Design and Analysis 17.2.1 Geometrical Pattern In Fig. 17.1, the developed microstrip low pass filter design is shown. The proposed structure consists of microstrip low pass filter with a ground plane of copper with conductivity of 5.8 × 107 S/m and thickness of 0.035 mm, printed on 0.8 mm thick FR4 substrate (εr = 4.4 and tan δ = 0.02). The modified ground plane is etched with the shape of complementary split ring resonator, enhance the performance of LPF using the concept of negative permittivity and permeability. The optimized dimensions are as follows: R0 = (8.6327 mm × 1.5295 mm), L 1 = (1.2894 mm × 0.2042 mm), C 2 = (3.0832 mm × 5.5528 mm), L 3 = (4.5222 mm × 0.2042 mm), C 4 = (4.9450 mm × 5.5528 mm), L 5 = L 3 , C 6 = C 2 , and L 7 = L 1 . Four CSSR cells are etched in the ground plane whose dimensions (as mentioned in Fig. 17.1c) are D0 = 6 mm, D1 = 3.6 mm, and s = 0.6 mm. The length, width, and height of substrate (as shown in Fig. 17.1b) are L = 40 mm, W = 20 mm, and H = 0.8 mm.

17 Design and Analysis of a Metamaterial-Based Butterworth Microstrip …

193

Fig. 17.1 Proposed Butterworth fifth order LPF a Top view of FR4 backed copper made modified ground plane, b Perspective view, and c Top view of metamaterial unit cell

194

S. Pramanik et al.

17.2.2 Design Procedure: Conventional Low Pass Filter The low pass filter can easily be designed by alternate sections of low and high impedance lines. These type of filters are known as stepped impedance filters. The general form of low pass filter contains high impedance lines which better approximates inductance and low impedance lines which better approximates capacitance. In this paper, these high and low impedance were consider as 20  and 120 , respectively. Firstly, a low pass lumped element design is selected, then values are normalized as c = 1 and g0 = 1, where c is cut-off frequency and g0 is input impedance. To get desired cut-off frequency and input impedance, these LC components are transformed as given by Eqs. 17.1 and 17.2.  0 gi 2π f c    0 g0 gi Ci = z0 2π f c 

Li =

z0 g0



(17.1) (17.2)

For stepped impedance, length for these components is calculated by Eqs. 17.3 and 17.4.   λgl (17.3) ll = 2π sin−1 (ωc L i /Z 0L )   λgc (17.4) lc = 2π sin−1 (ωc Ci /Z 0C ) The ωc L i (high impedance shunt susceptance) and ωc Ci in Eqs. 17.3 and 17.4 can be calculated with help of Eq. 17.5 and 17.6.    2πl L πlC + Z 0C tan ωc L = Z 0L sin λgc λgl     2 1 2πlC πl L + ωc C = sin tan Z 0C λgc Z 0L λgl 

(17.5) (17.6)

Parameters required for above design are given as follows: Parameter

Value

g1 = g7

0.4450

g2 = g6

1.2470

g3 = g5

1.8019

g4

2.0000

Cut-off frequency (fc )

3.5 GHz (continued)

17 Design and Analysis of a Metamaterial-Based Butterworth Microstrip …

195

Table 17.1 . Characteristic impedance

Z 0C = 20 

Z0 = 50 

Z 0L = 120 

Guided wavelength (mm)

λgc = 38.83

λg0 = 41.09

λgL = 43.70

Microstrip line width (mm)

W C = 5.55

W 0 = 1.52

W L = 0.2042

(continued) Parameter

Value

Relative dielectric constant (Er )

4.4

Height of substrate (h)

0.8 mm

Characteristic impedance (Z 0 )

50 

Corresponding microstrip width, (W0 )

1.5295 mm

Characteristic impedance of high impedance line, (Z 0L )

20 

Characteristic impedance of low impedance line, (Z 0C )

120 

Using above parameters, parameters listed in Table 17.1 and given equations the design for low impedance has been done.

17.2.3 Simulations and Analysis ANSYS HFSS along with Floquet port and master–slave conditions is utilized to simulate the designed structure. The structure was simulated for frequency sweep 1–8 GHz at solution frequency of 3.5 GHz and s-parameter response (in dB) has been obtained for the same. The response without of metamaterial is as displayed in represented in Fig. 17.2. The permittivity (εr ) and permeability (μr ) plots were obtained for unit metamaterial cell as shown in Fig. 17.3. Then, plot for design with metamaterial was obtained as shown in Fig. 17.4. The designed filters response shows a sharp attenuation. S 21 curve is close to 0 in passband and attains value of less than − 20 dB in stop-band which is a good response. Inferring to S11 , the normal return loss is − 10.47 dB and most extreme return loss is − 53.2 dB at 0.98 GHz. The cut-off occurs at 3.3 GHz (− 3 dB) which is close to intended cut-off frequency of 3.5 GHz. The filter lacks ripples and has sharper roll-off than normal Butterworth low pass filter. Metamaterial units etched in the ground plane have significantly improved the performance and selectivity of the filter.

196

S. Pramanik et al.

Fig. 17.2 Simulated S11 and S21 magnitudes in dB for low pass filter design (without metamaterial) in HFSS

Fig. 17.3 Simulated εr and μr magnitudes in dB for single metamaterial unit cell

17 Design and Analysis of a Metamaterial-Based Butterworth Microstrip …

197

Fig. 17.4 Simulated S 11 and S 21 magnitudes in dB for low pass filter design in HFSS

17.3 Fabrication and Measurement To experimentally affirm the Butterworth LPF, a model is constructed using the PCB technique, as shown in Fig. 17.5. The experiment was conducted using well-known free space techniques, as shown in Fig. 17.5c. Figure 17.6 depicts the measured Butterworth LPF reflection magnitudes (S11 and S21). The designed sample, as expected, exhibits LPF behaviour at 3.5 GHz. Except for minor deviations such as ripples and fabrication tolerance, the overall experimental results were in good agreement with the simulation results. Therefore, a Butterworth LPF having improved functionalities and considerably lower size were reported, as shown in the comparison of Table 17.2.

17.4 Conclusion In this paper, metamaterials are designed structures with exceptional electromagnetic characteristics. The isotropic metamaterial exhibits negative permittivity and penetrability, resulting in a negative index of refraction. In this case, CSRR units have been used which improved the selectivity of the filter. It gave sharper cut-off, stop-band less than − 20 dB. Furthermore, it lacks ripples. Metamaterial-based filters are stepping stone in designing subsystems for microwave frameworks.

198

S. Pramanik et al.

Fig. 17.5 Fabricated sample of Butterworth LPF a Top layer, b Bottom layer, and c Measurement setup of the fabricated sample

17 Design and Analysis of a Metamaterial-Based Butterworth Microstrip …

199

Fig. 17.6 Measured results of the fabricated sample

Table 17.2 Comparison table between this design and few other proposed low pass filters References

f c (GHz)

Max return loss in passband (dB)

Max insertion loss in stopband (dB)

Size (mm2 )

[14]

2.4

− 10

− 10

900

[15]

10.8

− 18

− 25

126

[16]

1.5

− 20

− 10

520

[17]

10.089

− 10

− 11.377

126

This paper

3.3

− 10

− 20

800

References 1. Pozar, D.M.: Microwave Engineering, 3rd ed., Wiley, Inc. (2005) 2. Fu, S.H., et al.: Compact miniaturized stepped impedance lowpass filter with sharp cutoff characteristic. Microw. Opt. Technol. Lett. 51(10), 2257–2258 (2009) 3. Hayati, M., et al.: Microstrip lowpass filter with high and wide rejection band. IEE Electron. Lett. 48(19), 1217–1219 (2012) 4. Rahman, A., et al.: Control of bandstop response of Hi-Lo microstrip low pass filter using slot in ground plane. IEEE Trans. Micro. Theory Tech. 52(3), 2539–2545 (2004) 5. Ahn, D., Kim, J.-S., Kim, C.-S., Qian, J., Itoh, T.: A design of the low-pass filter using the novel microstrip defected ground structure. IEEE Trans. Microw. Theory Tech. 49(1), 86–91 (2001) 6. Garcia-Garcia, J., Martin, F.F., Bonache, J., et al.: Stepped-Impedance low pass filter with spurious pass band suppression. Electron. Lett. 40(14), 881–883 (2004) 7. Sheen, J.W.: Compact semi-lumped low-pass filter for harmonics and spurious suppression’. IEEE Microw. Wirel. Compon. Lett. 10(3), 92–93 (2000)

200

S. Pramanik et al.

8. Tu, W.H., Chang, K.: Compact microstrip low-pass filter with sharp-rejection. IEEE Microw. Wirel. Compon. Lett. 15(6), 404–406 (2005) 9. Gupta, M., Upadhyay, D.K.: Design of a coupled-line microstrip butterworth low pass filter. In: 2019 6th International Conference on Signal Processing and Integrated Networks (SPIN), pp. 939–943 (2019) 10. Gupta, M., Kansal, M., Thyagarajan, S., Chauhan, P.S., Upadhyay, D.K.: Design of an optimal microstrip butterworth low-pass filter using colliding bodies optimization. In: Dutta, D., Kar, H., Kumar, C., Bhadauria, V. (eds) Advances in VLSI, Communication, and Signal Processing. Lecture Notes in Electrical Engineering, vol. 587. Springer, Singapore (2020) 11. Gangwar, K., Paras, Gangwar, R.P.S.: Metamaterials: characteristics, process and applications. Adv. Electron. Electr. Eng. 4(1), 97–106 (2014). ISSN 2231-1297 12. Marquez, R., Mesa, F., Martel, J., Medina, F.: Comparative analysis of edge- and broadsidecoupled split ring resonators for metamaterial design-theory and experiments. IEEE Trans. Antennas Propagat. 51, 2572–2581 (2003) 13. Ramakrishna, S.A.: Physics of negative refractive index materials. Rep. Progr. Phys. 68, 449– 521 (2005) 14. Ibrahim, A.A., Abdel-Rahman, A.B., Abdalla, M.A.: Design of third order band pass filter using coupled meta-material resonators. In: 2014 IEEE Antennas and Propagation Society International Symposium (APSURSI), pp. 1702–1703 (2014) 15. Abdalla, M.A., Arafa, G., Saad, M.: Compact UWB LPF based on uni-planar metamaterial complementary split ring resonator, In: 2016 10th International Congress on Advanced Electromagnetic Materials in Microwaves and Optics (METAMATERIALS) (2016) 16. B. Nasiri, Errkik, A., Zbitou, J., Tajmouati, A., Elabdellaoui, L., Latrach, M.: A novel design of a compact miniature microstrip low pass filter based on SRR. In: 2017 International Conference on Wireless Technologies, Embedded and Intelligent Systems (WITS) (2017) 17. Kurhe, N., Labade, R.P.: Metamaterial based ultra wide band low pass filter using new configuration of complementary split ring resonator cell. In: 2018 International Conference on Information , Communication, Engineering and Technology (ICICET) (2018)

Chapter 18

Internet of Things-Enabled Irrigation System in Precision Agriculture Shabir A. Sofi and Saniya Zahoor

Abstract Internet of Things (IoT) plays a very important role in increasing productivity and in achieving the success in the activity and practice of agriculture in many countries. The main purpose of using IoT in agriculture is that we can control water supplies in agriculture field without any wastage by using IoT sensors and can monitor the agriculture field continuously. IoT-enabled precision agriculture helps the farmers to take suitable precaution in case of any damage and diseases in the crops and plants. It monitors the growth of crop through IoT sensors that sense the various agricultural parameters such as pH of soil, its temperature, moisture content, etc., so that farmers can sow the seeds accordingly. This paper discusses the precision agriculture using Internet of Things. The paper also presents a case study of irrigation system that senses the agricultural parameters and waters the agricultural fields automatically. This paper further discusses the other applications of precision agriculture and possible challenges faced in precision agriculture.

18.1 Introduction IoT plays a very important role in increasing productivity and in achieving the success in the activity and practice of agriculture in many countries [1]. According to the research done by the UN Food and Agriculture Organization, the world is going to require to produce 70% more food in 2050 than it did in 2006 in order to suffice the need of food for rising population and to fulfil this demand farmers and agricultural fields are switching to IoT for enhancement of productivity, global market, less human intervention, minimum time and cost, etc. [2]. India’s financial resources are mainly S. A. Sofi (B) Department of Information Technology, National Institute of Technology Srinagar, Hazratbal, Kashmir, Jammu and Kashmir, India e-mail: [email protected] S. Zahoor Department of Computer Science, University of Kashmir, Hazratbal, Kashmir, Jammu and Kashmir, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_18

201

202

S. A. Sofi and S. Zahoor

dependent upon agriculture and the important hindrance that arises in traditional farming is shifting of climatic conditions from one to another which encompasses heavy or less rainfall, intensified storms, heat waves, etc., and these conditions affect the yield or productivity drastically. So in order to increase the productivity and to reduce the obstacles in agriculture, there is a need to use ingenious technology called IoT. IoT devices help in increasing the productivity and yield as they can not only monitor acid level present in the soil, temperature of soil but precision agriculture helps in monitoring of livestock productivity and health as well. Even IoT devices are able to provide information to the farmers regarding yield of crops, pest infection, yield of crops, nutrition of soil, etc. Hence, we can say that IoT devices are capable of providing information to farmers regarding all those variables which in one way or other way affect agriculture to a large extent. One of the purposes of using IoT in agriculture is to control water supply in agriculture field without any wastage by using sensors, monitor the agriculture field continuously and take suitable precaution in case it is prone to any damage. IoT sensors can sense the soil pH, temperature, moisture, etc., so that farmers can sow the seeds accordingly. Figure 18.1 shows the architectural diagram for controlling and monitoring of IoT-based agricultural field according to which in order to monitor the condition of a farm this kind of platform is designed so to obtain the real-time data frequently and automatically this kind of platform is provided so to get the real-time information frequently and in an automated way. The sensor senses data and data is forwarded to edge (gateway) and then cloud for further storage and analysis. The cloud platform comprises of database server and provides API in order to access the data, import the data from database so to visualize it and take appropriate action in order to prevent field from any kind of damage. Only the legitimate farmer can monitor real-time condition of agriculture field by accessing server database. The sensors that sense data are installed in an agricultural field and not only one or two sensors are installed but number of sensors is installed because data needs to be sensed from different locations of farm. The data that sensors sense can be data related to humidity, data associated with quantity of moisture contained in soil, data related to temperature of soil, data related to quantity of water present in soil, data related to insecticides or pesticides present in soil or needed in soil, data related to the crop monitoring. Currently, a number of traditional manual irrigation practices are still in use by farmers but this result in wastage of water. These practices irrigate land manually which results in the wastage of water. Mass irrigation is one of the methods that is used to irrigate the agricultural field. The disadvantage of the method is that the amount of irrigation exceeds the need of plant resulting in over usage of water. There has been research in automatic irrigation system to reduce water wastage and reduced labour [3]. Therefore, there is a need of IoT-enabled irrigation system for precision agriculture. The rest of the paper is organized as: Sect. 18.2 presents literature survey, Sect. 18.3 discusses a case study: smart irrigation in precision agriculture, Sect. 18.4 gives other applications of precision agriculture, Sect. 18.5 presents challenges and existing solutions, and Sect. 18.6 gives conclusions and future work.

18 Internet of Things-Enabled Irrigation System in Precision Agriculture

203

Fig. 18.1 IoT-based agriculture architecture

18.2 Literature Survey There has been recent work in IoT-based irrigation systems. Authors in [4] propose an IoT-based drip irrigation system in which irrigation is remotely controlled via a mobile application. Various sensors such as temperature, humidity, etc., are installed in the agricultural field to monitor environmental parameters. The sensed information is passed to the user on mobile application where data can be accessed and control over irrigation can be done remotely. The paper presents an automatic and smart watering mechanism for agricultural fields. Similarly, authors in [5] present an IoT-based irrigation system. It uses sensors embedded on microcontroller board to monitor water requirements and control irrigation in the agricultural fields. There has been work on optimizing the battery lifetime of devices used in IoTbased irrigation system [6]. Various agricultural parameters are monitored in this and these include soil moisture, soil humidity, soil pH, etc. Based on this data, the presented system takes decision to perform irrigation in agricultural field. Authors in [7] present a mobile application-based drip irrigation system. In their work, different parameters are monitored using sensors for humidity, light, temperature, etc. The sensed data is offloaded to the computer for further storage, analysis, and decision. The on and off drip is displayed using graphs on mobile application to the user. Similar work has been proposed in [8] as well. Taking into consideration, the fact of water crises around the world and at the same time its rising demand in the agricultural sector and in other fields compels us to provide the water in required quantities to these sectors. Water scarcity affects both the quality and the quantity of a crop and excessive use of water also because serious problems like deficiency of nutrients in the soil, hence giving rise to disease in the

204

S. A. Sofi and S. Zahoor

crops. It is not that easy to determine the amount of water required by the crop because various factors like type of the crop, type of the soil, moisture content present in the soil, and precipitation are responsible for determining that very particular thing. So taking into consideration, this thing IoT devices which make our agriculture smarter enough not only helps us to use the water in optimum way but also helps us to improve health of the crops as this technology makes use of the sensors that are being installed in an agricultural field which can sense the data regarding the water content present in the soil and this information when is visualized by the legitimate user drives him to take an appropriate action so to prevent critical loss in the yield of a crop that may occur due to scarcity of water and even prevents the deficiency of nutrients in soil which in turn means that plants are prevented from diseases as this may occur due to excess amount of water present in the soil. So the current problem of irrigation is expected to get changed by switching to IoT-enabled technology in agricultural field.

18.3 Case Study: IoT-Enabled Irrigation System in Precision Agriculture The presented system makes the efficient use of water. It will water plants on the basis of soil moisture level, thereby preventing over irrigation and under irrigation which are the reasons of water wastage and crop destruction. The proposed irrigation system will be very efficient in areas like house gardens, office premises, buildings, etc. This system uses devices like Arduino microcontrollers, sensors, Wi-Fi module is used to send the real-time data to cloud server for various resource managements. The objectives of the paper are to design a smart irrigation system to water plants with the use of devices like Arduino microcontrollers. Wi-Fi module is used to send the realtime data to the cloud server for various resource managements. Figure 18.2 shows the pin diagram for Arduino. Smart irrigation system using Arduino programmed to monitor agricultural parameters. Soil moisture sensor is connected to the A0 pin to the Arduino board as shown in Fig. 18.3. The system presents a smart irrigation mechanism. If the value of soil moisture is below a threshold, it sends the signal to the IoT devices to pump water to get the moisture back to normal level. The system consists of a motor pump; motor driver; and motor driver to water the fields (see Figs. 18.4 and 18.5).

18.4 Other Applications of Precision Agriculture Following discusses the other applications of precision agriculture [9].

18 Internet of Things-Enabled Irrigation System in Precision Agriculture

Fig. 18.2 Arduino pin diagram

Fig. 18.3 Soil sensor interface

Fig. 18.4 Block diagram for smart irrigation system

205

206

S. A. Sofi and S. Zahoor

Fig. 18.5 Schematic diagram for smart irrigation system

18.4.1 Soil Sampling and Mapping Soil is the stomach of a plant and obtaining information about soil of an agricultural field helps us to make important decisions, and our main focus is to obtain information regarding the nutrients present in the soil so that actions can be taken accordingly when soil is deficient in nutrients. The various factors that are necessarily responsible for analysing the amount of nutrients present in the soil are topography, type of soil, amount of fertilizer present or required, etc. These factors give clue about the physical, chemical, biological status of the soil and thereby helping to identify the factors that are limiting in nature so to deal with the crops accordingly. This helps us to sow different crop varieties in the soil according to the properties of soil that they match with. Today, we have different types of sensors available that sense the various properties of soil that include texture of the soil, water absorbing and retention capacity, absorption rate which minimizes erosion densification, salinization, acidification, and pollution. Drought is one of the major factor that reduces the yield of crop and many regions around the world are facing this issue at different levels so using remote sensing approach we can monitor the data regarding moisture content present in the soil frequently and hence helping us to analyse drought condition in the farm and accordingly the steps can be taken to control the damage.

18.4.2 Fertilizer A fertilizer is a chemical or natural substance that provides essential nutrients to the soil required for the growth of plant. A plant requires three macronutrients more importantly and those are nitrogen, potassium, phosphorus for growth of the leaves,

18 Internet of Things-Enabled Irrigation System in Precision Agriculture

207

for the growth of stem, and movement of water across the body of plant, for development of roots, fruits, and flowers, respectively, so any kind of deficiency or even excessive use of these nutrients leads to the adverse effects on the quality of soil, can make ground water poisonous as well. Also crops only absorb less than half of the nitrogen applied through fertilizer and rest is released out in the air which depletes the soil quality as well as cause unbalancing the climate. Therefore, the precision agriculture helps us to properly estimate the amount of nutrients required by the soil, thus preserving the ill effects on the soil and air. The measurement of nutrients required by the soil is not only expensive and time consuming but also requires more labour. Agriculture field that makes use of IoT overcomes these things by decreasing labour, reducing time, and increasing accuracy. Not only fertilization but fertigation and chemigation are other benefits of IoT.

18.4.3 Crop Disease and Pest Management Due to the failure of crops and yield reduction, many countries have faced loss of life due to non-availability of food as well as loss in economy. In order to control this loss, many pesticides and other agrochemicals became necessary part of agriculture sector. IoT-based intelligent devices allow farmers to use pesticides efficiently at the place that affected by pests. The innovative approaches in monitoring the disease in the crop and recognizing the area affected by the pests are based on processing the image of field where crops have been sown along with analysing the information regarding plant health, pest situation in the field which is obviously sensed by IoT sensors and thus allowing the farmers to take respective counter measure to prevent crop being getting attacked by the pests and diseases in the crop. Moreover, the advancement in robotic technology allows us to use the robots in agriculture with sensing devices and spraying nozzles with which it can locate the pest and deal with them accordingly by spraying the pesticide at that place and that action is triggered when legitimate user send command to do that specific action via application that user might be using.

18.4.4 Yield Monitoring, Forecasting, and Harvesting Yield monitoring is the process of analysing various factors related to the yield of farm for example mass of grain flow, content of moisture harvested, grain quantity of harvested crop, and quality of crop. On the other hand, crop forecasting is the mechanism of predicting the amount of production or yield beforehand. This prediction allows the farmer to take proper decisions and make correct planning regarding future. The right time for harvesting the crop which in turn helps us to surplus the production and quality of a crop can be specified properly by analysing the quality and maturity of the yield. A yield monitor can be installed on the harvester which in

208

S. A. Sofi and S. Zahoor

turn being linked with the application that particular legitimate user/farmer is using and via which harvest data and high quality yield maps are displayed to the farmer.

18.5 Challenges and Existing Solutions Following discusses challenges and existing solutions for precision agriculture [10].

18.5.1 Weak Internet Connectivity in Agriculture Fields Usually, most of the agricultural fields are situated in remote areas or we can say in rural areas where in the availability of Internet is not that much so the data transmission speed is obviously not strong. Also, links through which communication is possible may get affected due to crops or any other barrier. So such kind of thing facilitates the cost of forwarding the data or we can simply say it causes communication cost to increase and add in creating hindrance for switching to IoT-enabled technology in agriculture making it smarter enough and with the introduction of concept of big data such cost are increase tremendously. Existing Solution: In order to transmit data or information regarding agriculture field, we can make use of vacant TV frequencies as found by some of the prolific researchers. This kind of approach is mainly useful in the areas that are located remotely as in such areas due to poor TV reception which causes creation of white spaces in TV frequencies used for broadcasting such kind of white spaces are then available to us and we can make use of them. Also there exist some broadcast bands such as ultra-high frequency and very high frequency bands which possess capability to enhance the strength of Wi-Fi signals and hence making them more powerful. So this approach helps us in decreasing the cost, while connectivity gets increased ultimately helps us in improving farming technology.

18.5.2 High Hardware Costs Farmers primarily depend on the network of sensors which is distributed all over in the their fields as data regarding different variables such as temperature of the soil, moisture content present in the soil, pH level of the soil, water content present in the soil, insecticides or pesticides present in the soil, etc., needs to be sensed so take the correct step at right point of time. But there are only limited numbers of sensors that are cheaper in cost and rest are expensive. So those sensors that are cheaper in cost will obviously not have that much of functionality may be in terms of battery life, storage, processing power, etc., it means that such sensors will put limit on analysis of

18 Internet of Things-Enabled Irrigation System in Precision Agriculture

209

data, hence ultimately limits the actions that are required to be taken at the time when its required so to prevent any kind of damage, therefore, the productivity will also get affected that will in turn affect the profit. Talking about the tools that are required for adopting IoT-enabled technology in farming or agriculture they are costly and adopting automated machines in agriculture field would be very expensive than those machine that operate manually. But in order to get more productivity, more yield in short to get good output from field so to increase profit its necessary for the farmer to invest over such technology but at initial stage investment of such large amount will be more difficult for farmers. Existing Solution: Due to the reason that technologies which are required for precision agriculture be it the high end sensor or automatic machines, they are very expensive and a farmers face difficulty in affording such technologies so that they can use them in their farms. But there exist some solution to it like aerial vehicles possessing sensors can be used to improve geographical coverage and create accurate maps. We know that some countries or area do exist in the world that put restriction on using drone, it may be due to government protocols that are not allowing them to use it, or due to issues in battery life, due to its expensive cost, etc., so in such cases tethered eye helium balloons can be used. Basically, the data that is collected by sensors which are installed at various places in an agriculture field can be made more clear by using these aerial sensors that do generate images of these farms frequently. So this can be one of the approaches that enable us to deal with high hardware cost and thus allowing making more accurate data collection.

18.5.3 Disrupted Connectivity to Cloud Adopting IoT technology in agriculture so to make it smarter enough, obviously makes us to use something that will first sense the data and that’s nothing but a sensor, this data must be stored somewhere and that’s obviously a cloud where further analysis of data is done using big data analytics (because data that sensed from farm is in huge quantity), to get useful information from the data and accordingly correct actions can be taken in the field when required so to prevent the farm from any kind of harm that will ultimately affect yield and hence profit. But the problem lies at connectivity to the cloud via Internet as we know that farms or agriculture field are located in remote areas where Internet connection is not that strong and reliable which will support the streaming of such huge amount of data to the cloud plus their might be some source of barriers which will affect connectivity to cloud. Existing Solution: Some researchers have put forth the concept of farm beats that has been developed to deal with the non-reliable Internet connectivity between the agriculture field and the cloud. They include some capability that is available offline, the availability of specific IoT gateway, etc. These provide IoT connectivity that is end to end in nature and therefore allows this farm beats to provide different services in precision agriculture like providing accurate irrigation supply, accuracy

210

S. A. Sofi and S. Zahoor

in determining the pH, depicting the yield, prediction of climatic conditions, etc., even when its communication with the cloud is not reliable. So farmers need to switch to such technologies that ease the operation related to data, thereby enhancing the yield. This way farmer is able to discover all advantages of precision agriculture.

18.5.4 Lack of Infrastructure Switching to precision agriculture drives the farmer to make use of IoT-enabled technology in agricultural fields but due to the presence of communication infrastructure which is poor due to which farmers will not be able to take full support of this IoT technology. As agriculture fields are located usually at remote places where Internet connection is not reliable so the need to access data remotely regarding different variables of field such as temperature, moisture content present in the soil, and water content present in the soil reliably at any moment of time will face problem and in turn affect the monitoring of field and data related to it. So the overall affect will be on performing action in a farm that prevents any kind of damage, hence will ultimately affect yield thus profit. Existing Solution: Need and organization of mobile IoT nodes with space–time grouping and infrastructure that cloud provides where services are present which can perform processing of data and local or remote exchange of message and decision making. The services also necessitate the mobile network to fulfil the communication requirements of mobile IoT nodes.

18.5.5 Lack of Security The data that is collected from agriculture field is in large quantity and to secure such data is difficult because the security providing algorithm that are used in today’s world provide good amount of security but are not able to run on IoT devices with old equipment and another reason can be due to small size of an IoT device such strong security providing algorithm will not be able to run on these small devices, hence any user that in not legitimate can come and make fabrications or modification in this data which will affect the decisions that are required to be taken at right moment. Existing Solution: We can protect our IoT-based system from illegitimate access through access control techniques. This is very useful in applications that are critical in nature. We can also use security algorithms that are strong enough in nature on the devices that end user is using so as to prevent the production and expansion of forged messages.

18 Internet of Things-Enabled Irrigation System in Precision Agriculture

211

18.6 Conclusions and Future Work Internet of Things plays an important role in monitoring the agricultural parameters and getting information to farmers. Adoption of IoT in agriculture provides optimized solutions for many applications of agriculture. This paper presents an automated irrigation system which is effective enough to optimize utilization of water, human, and other resources. This system helps in irrigation in areas with low water level and leads to sustainability. This system is very volatile and low maintenance and could be adjusted according to various types of crops without much human efforts. The proposed system consists of a soil moisture sensor which senses the soil moisture and automatically irrigates the field. Automatic irrigation scheduling consistently has reduced the water wastage. Also the real-time data visualization helps in various resource management purposes and also keeps farmers aware about the current situation of the fields. The proposed system can be extend to include more agricultural parameters such as pH, humidity, and temperature sensors for better knowledge of the field. Also actuators such as automatic fertilization can be included in the system. Also, PIR sensor can also be used for intrusion detection. Further, using IoT-based precision agriculture system can be useful for farmers in the village where this facility can be implemented at the local fog level without many investments.

References 1. Sinha, B.B., Dhanalakshmi, R.: Recent advancements and challenges of Internet of Things in smart agriculture: a survey. Future Gener. Comput. Syst. 126, 169–184 (2021) 2. Pierce, F.J., Nowak, P.: Aspects of precision agriculture. Adv. Agron. 67, 1–85 (1999) 3. García, L., et al.: Deployment strategies of soil monitoring WSN for precision agriculture irrigation scheduling in rural areas. Sensors 21(5), 1693 (2021) 4. Khan, K.A., et al.: A study on development of PKL power. In: Computational Intelligence and Machine Learning, pp 151–171. Springer, Singapore (2021) 5. Haji, S.H., Sallow, A.B.: IoT for smart environment monitoring based on python: a review. Asian J. Res. Comput. Sci. 57–70 (2021) 6. Rastogi, R., et al.: Sensor-Based irrigation system: introducing technology in agriculture. In: The Smart Cyber Ecosystem for Sustainable Development, p. 153 (2021) 7. Numajiri, Y., et al.: iPOTs: Internet of Things-based pot system controlling optional treatment of soil water condition for plant phenotyping under drought stress. Plant J. (2021) 8. Hassan, A., et al.: A wirelessly controlled robot-based smart irrigation system by exploiting Arduino. J. Robot. Control (JRC) 2(1), 29–34 (2021) 9. Srilakshmi, A., et al.: A comparative study on Internet of Things (IoT) and its applications in smart agriculture. Pharmacognosy J. 10(2) (2018) 10. Srinivasan, C.S.K.: Smart farming-challenges and their solution on agriculture using IoT. Ann. Rom. Soc. Cell Biol. 3983–3996 (2021)

Chapter 19

A Hybrid Feature Selection Framework for Breast Cancer Prediction Using Mutual Information and AdaBoost-RFE Himanshu Dhoke and Aakanksha Sharaff

Abstract Every year, the number of deaths from breast cancer rises drastically. It is the most common kind of cancer and the biggest reason for mortality in women throughout the world. In real-world applications like breast cancer classification, it is difficult to cope with high-dimensional and unbalanced datasets. To counter this problem, feature selection is used as a strategy for preprocessing, with the classification performance and computing efficiency taken into account. This research introduces a hybrid feature selection framework that incorporates filter and wrapper techniques to obtain optimal feature subset. In the first stage, the filter technique mutual information (MI) was used to arrange the features according to their relevance due to its computational efficiency. It cannot, however, discard less important features. Therefore, a wrapper method recursive feature elimination (RFE) employed to eliminate redundant features in the second phase. For RFE, Adaptive Boosting (AdaBoost) is used as the estimator. By using a random forest classifier, the suggested feature selection approach is compared to existing feature selection methods. With a 79.41% feature reduction ratio, the suggested MI-AdaBoost-RFE technique selects seven features out of 32. The suggested method has a 97.36% accuracy and an AUROC of 0.974, which is higher than other recent studies.

H. Dhoke (B) · A. Sharaff National Institute of Technology, Raipur, Chhattisgarh, India e-mail: [email protected] A. Sharaff e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_19

213

214

H. Dhoke and A. Sharaff

19.1 Introduction Among the most crucial tasks in machine learning and data mining is classification, which involves identifying class labels based on information about various attributes. Datasets with a huge number of attributes are common in classification. Not all of features, however, are essential to classification [1]. Irrelevant and duplicated features might affect classification performance. Furthermore, a huge amount of features causes the curse of dimensionality, which is a key hurdle in classification tasks. A common strategy to solve these issues is to use dimensionality reduction techniques. There are two most common feature reduction strategies, each with its own set of benefits and drawbacks [2]. The first technique generates a new low-dimensional feature set from the original feature space, preferably consisting of uncorrelated features. Furthermore, the majority of the data contained in the original feature space is likely to be preserved in the new feature subset. Feature synthesis, also known as feature extraction, is a concept that describes how each new feature is a function of all previous ones [3]. Feature selection or feature subset selection is another reduction strategy. The process of choosing a subset of significant features from a huge number of original features is known as feature selection [4]. The subset of important features chosen should be suitable for describing the complete information, which is described by original features. By eliminating unrelated and redundant features, feature selection technique may increase the performance of classification and reduce the computation overhead. When developing a feature selection algorithm, two major difficulties must be addressed [5], the first is the most efficient way to find the best feature subset, and the second is how to identify a subset’s goodness. The remainder of this paper is organized into the following sections listed below. The prior research on the feature selection problem is examined in Sect. 19.2. The methodology we used for our research is defined in the Sect. 19.3. The findings of the research are presented in Sect. 19.4, while Sect. 19.5 summarizes the findings and conclusions with future work.

19.2 Literature Survey This section summarizes previous research on filter and wrapper-based hybrid feature selection algorithms. To turn a huge quantity of data into a useful data distribution, feature selection is crucial. Rouhi et al. [6] proposed the R-m-GA method to decrease the size of microarray data by integrating Relief, minimum redundancy, and maximum relevance (mRMR) along with the genetic algorithm. Sharaff et al. [7] evaluated the effectiveness of two feature selection strategies, info-gain and chi-square, on several classification algorithms to build a classifier for spam email filtering. In comparison with the chi-square feature selection approach, info-gain outperforms it.

19 A Hybrid Feature Selection Framework for Breast Cancer Prediction …

215

Many researchers use mutual information as one of the most prevalent feature selection approaches. It is a statistical model that uses information from one feature to determine the dependence between the two features. To overcome the problems of selection of features, Hancer et al. [8] suggested a technique for selecting features on basis of two variants of differential evolution, i.e., single and multi-objective and mutual information that could be applied to continuous datasets. Lu et al. [9] combined the mutual information maximization a filter approach with the adaptive genetic algorithm a wrapper approach to create a hybrid feature selection methodology in order to minimize the size and number of duplicated samples in gene expression data. Wrapper techniques are more efficient in terms of performance, but they are more computationally costly than filter approaches; researchers used the RFE wrapper approach extensively. Rtayli et al. [10] introduced the new feature selection method in which the random forest classifier was upgraded to identify frauds after RFE embedded with SVM was applied for predicting features after normalizing the dataset using synthetic minority oversampling and hyper-parameter tuning utilizing gridsearchCV approach. Chiew et al. [11] suggested a feature selection approach; in that approach, the features are automatically identified by utilizing an unique technique cumulative distribution function gradient, after that selected features are forwarded to the second stage, which uses a function perturbation ensemble to generate the best features. Boosting algorithms help poor learners increase their accuracy. Even though the training data is the same, various learning algorithms perform better with different predictors. Priscilla et al. [12] presented a two-phase feature selection approach that utilized mutual information to rank the features according to their relevance in the first phase. Then, to pick the best feature space, XGB-RFE, a wrapper approach based on fivefold cross-validation, is used. Although there has been other research on integrating filter methods with wrapper approaches in the literature, the focus of this work was on determining whether cancer is benign or malignant from a dataset with huge dimensions. The suggested technique combines mutual information a filter technique along with a wrapper approach recursive feature elimination to eliminate the unnecessary attributes, reducing computational time and increasing performance of proposed model. In this study, AdaBoost, an effective learner, is implemented in the recursive feature elimination.

19.3 Methods The proposed hybrid feature selection model is described in this section. To determine the optimal feature subset, a filter technique mutual information approach is used in the first step. The features acquired in the first step are sent to the RFE in the second phase, which is then cross-validated to remove any remaining irrelevant features and improve classification accuracy [10].

216

H. Dhoke and A. Sharaff

19.3.1 Mutual Information Mutual information is a powerful filter technique for determining how variables relate. When the MI value between two different random variables is more than zero, the variables are substantially dependent to each other [13]. When the mutual information value is zero, the variables are considerably independent to each other. As illustrated in Fig. 19.1, Eq. 19.1 represents theoretical approach to mutual information between two different random variables U and V for n observations. I(U ; V ) = H (V ) − H (V |U )  p(u, v) . = p(u, v) log ueU,veV p(u) p(v)

(19.1)

where H(V ) shows the level of ambiguity in a random variable V, and H(V|U) denotes the conditional entropy of U given V, and p(.) represents the probability mass function. Equation 19.2 shows how MI is proportional to the entropies of the given random variables U and V. If the MI value between two discrete random variables is larger than zero, the two variables are strongly dependent; otherwise, the variables are statistically independent. The shortcomings of this empirical approach include evaluating conditional independence for all attributes and calculating probability density functions. ⎧ ⎨ H (U ) − H (U |V ) I (U ; V ) = H (V ) − H (V |U ) (19.2) ⎩ H (U ) + H (V ) − H (U, V ). Fig. 19.1 Relation among mutual information and entropy

19 A Hybrid Feature Selection Framework for Breast Cancer Prediction …

217

19.3.2 Recursive Feature Elimination Recursive feature elimination (RFE) is a well-known wrapper kind feature selection technique, which means it employs a unique machine learning model to assist in feature selection. RFE identifies a set of attributes by having started with all of the features in the training set and subsequently removing the ones that aren’t needed until only the appropriate number of features remains. RFE eliminates collinearity and feature dependency. RFE employed a cross-validation strategy to choose the most appropriate scored attributes by using the estimator numerous times for every iteration to exclude the less essential attributes. In the RFE technique, cross-validation is utilized to help the estimator learn and create the best feature score weights. It aids in the training of the model in order to prevent data from being over fit [14]. The decision coefficients are obtained by extracting the appropriate attributes, which are subsequently preserved for the next training iteration. With each iteration of the RFE method, the size of the feature decreases recursively. When it comes to selecting the best features from the first n features on the basis of feature ranking, RFE with cross-validation outperforms traditional RFE.

19.3.3 Proposed Hybrid MI-AdaBoost(w) -RFE Feature Selection Method A framework of the presented hybrid feature selection approach is illustrated by Fig. 19.2. Firstly, the breast cancer dataset is collected from UCI machine learning repository. Then, it is sent to preprocessing stage. The data for training is delivered to the proposed model’s first stage following the data preprocessing stage. Even though mutual information is a popular statistical approach in numerous studies, this filter technique may rank features by identifying relationships between them. As a result, MI has chosen the most important features. The best predictors are chosen for the next step of processing via RFE along with cross-validation based on feature ranking. The best subset of features obtained from the first stage is used to filter the features with the highest correlation and relevance in the second phase. On the basis of the decision scores produced by the classifier, the RFE approach recursively removes the less significant features at each iteration [1]. The major objective of the research is to keep the model from being overfit. Hence, AdaBoost was chosen as a strong model. The AdaBoost method employs weak learners in the form of extremely little decision trees, which are added sequentially to the ensemble. After then, each model in the series seeks to improve on the previous model’s predictions. This is accomplished by focusing the training dataset on training cases where previous models failed to predict correctly [15]. The hyper-parameters, on the other hand, have a significant influence on the model’s performance; thus, they are fine-tuned to identify the model’s optimum parameters.

218

H. Dhoke and A. Sharaff

Fig. 19.2 Flowchart for the suggested MI-AdaBoost(w)- RFE method

There are two phases involved in the proposed hybrid feature selection model. The MI filter approach is being used in the first stage to choose the subset of best features. To enhance computational efficiency and classification accuracy, less important features have been discarded. In the second stage, the remaining unnecessary and correlated attributes from the previous stage are removed using recursive feature elimination method. The dataset is split into k folds, with k-1 folds being the training set and remaining onefold being the testing set, and the scored values for n observations are determined using a cross-validation procedure. This procedure is carried out k times, with every fold being validated only once. As a result, it exceeds existing strategies for reducing high-dimensional features and avoiding overfitting in models. The random forest classifier is being adopted to evaluate the optimum feature set’s effectiveness. Figure 19.3 depicts feature ranking achieved by the suggested model.

19.3.4 Data Description and Preprocessing In this experiment, the Wisconsin breast cancer (diagnostic) dataset from the UCI machine learning repository was utilized. The dataset’s features are generated by using a digitized picture of a fine-needle aspirate sample of breast cancer. The dataset

19 A Hybrid Feature Selection Framework for Breast Cancer Prediction …

219

Fig. 19.3 Feature importance graph for suggested MI-AdaBoost(w) -RFE method

includes total 569 instances (357 benign and 212 malignant), 2 classes (62.74% benign and 37.26% malignant), and 32 integer-valued attributes. After eliminating the dataset’s constant, missing, and duplicate values [5]. The label encoder technique is utilized in order to transform the categorical values of the target variable into numerical form.

19.4 Experimental Results and Discussion The suggested hybrid feature selection technique MI-AdaBoost(w) -RFE is compared with the existing feature selection methods in dealing with an unbalanced and huge dimension dataset in this section. Hyper-parameter tuning was performed on the

220

H. Dhoke and A. Sharaff

Table 19.1 Confusion matrix

Predicted values Actual values

Positive

Negative

Positive

TP

FN

Negative

FP

TN

machine learning model to reduce overfitting and increase model performance. The proposed model’s efficiency is compared to that of other feature selection strategies as well as the original features. This experiment was carried out on a Windows 10 platform. Python programming language was used to implement the algorithm.

19.4.1 Evaluation Matrix In the binary classification problem with an unbalanced dataset, determining appropriate evaluation measures is a major challenge. For this study, the following assessment measures were used to assess the classifier’s performance. Although accuracy is a widely used statistical measure for calculation, it cannot be utilized as a key metric to assess performance when the minority class is small. As a result, various parameters like as F1-score, recall, precision, and AUROC are taken into account [4]. The classification metrics are first obtained using Table 19.1 confusion matrix. True Positives (TP) are instances that have been correctly identified as false. True Negatives (TN) are instances that have been properly predicted as true. False Positives (FP) are instances that were predicted to be false but turned out to be true. False Negatives (FN) are instances that predicted to be true but are actually false.

19.4.2 Model Verification To ensure that the suggested feature selection approach is legitimate, the classification performance of all models using entire features and derived features from alternative feature selection approaches is evaluated. The suggested feature selection method outperforms the outcomes of the other feature selection techniques employed in the experiment. Table 19.2 compares the feasibility of the proposed MI-AdaBoost(w) RFE approach based on metrics such as accuracy, recall, precision, F1-score, and AUC-ROC score. Figure 19.4 depicts the classification performance of methods using original features as well as features created by MI and the wrapper method AdaBoost-RFE, after that the suggested MI-AdaBoost(w) -RFE. Each feature selection approach selects a distinct number of features nevertheless, when contrasted to other methods, the suggested technique selects the fewest features.

19 A Hybrid Feature Selection Framework for Breast Cancer Prediction …

221

Table 19.2 Comparison of the suggested MI-AdaBoost(w) -RFE method to benchmark feature selection approaches Methods

Features Selected

F1 Score

Recall

Precision

Accuracy

Original

32

93.18

97.61

89.13

94.73

MI

15

94.25

97.61

91.11

95.61

RFE

10

95.23

95.23

95.23

96.49

MI-AdaBoost-RFE

7

96.47

97.61

95.34

97.36

Fig. 19.4 Classification performance of feature selection methods

We can observe that the proposed strategy MI-AdaBoost(w) -RFE achieved an F1score of above 96% with the optimal amount of features and outperformed other approaches by attaining a classification accuracy of over 97%. Table 19.2 shows a comparison of performance indicators for the feature selection methods. The ROC curve for the breast cancer dataset is depicted in Fig. 19.5 in relation to the various feature selection strategies utilized. The suggested MIAdaBoost(w) -RFE takes much less time to execute when compared to certain existing feature selection methods and with the entire features. These exceptional findings suggest that the proposed MI-AdaBoost(w) -RFE approach can assist to improve the classifier’s efficiency by decreasing the complexity caused because of high dimension of data. The study’s effectiveness is determined by the outcome of the results. Because mutual information does the ranking in the first phase, RFE computation time is lowered by reducing the less significant feature subset. The suggested approach reduces the dimensionality, which reduces the execution time for all classifiers.

222

H. Dhoke and A. Sharaff

Fig. 19.5 ROC curve for the breast cancer dataset

19.4.3 Shortcomings Although the suggested technique has certain advantages, it also has some drawbacks. The current MI filter approach can only rank the feature importance because of its constraints. As a result, MI filter approach works well for datasets with low dimensions, but it is unfeasible for datasets with high dimensions. When selecting the amount of ranked features from the mutual information in the first stage, it might be difficult to determine the limit. For datasets with high dimensions, the existing wrapper technique recursive feature elimination is more time consuming to compute. Proposed method used a combination of MI and RFE to overcome these constraints. Because in the second stage, the highly rated features from the previous stage are retrieved and fed to RFE. Hence, calculation time is reduced for RFE. Because of the combined feature selection technique’s feature elimination, the findings are promising.

19.5 Conclusion and Future Work This research presents a unique hybrid feature selection strategy called MIAdaBoost(w) -RFE to improve the classification rate of breast cancer disease by identifying the most useful subset of features. A comparison with various feature selection methods is performed to authenticate the effectiveness of the suggested feature selection approach. According to the findings, the proposed technique decreased the dataset’s dimensionality by focusing on the most significant attributes, hence improving the performance of classifier by reducing computational time. The

19 A Hybrid Feature Selection Framework for Breast Cancer Prediction …

223

outcomes of the experiment are impressive in terms of F1-score, with AdaBoost outperforming other techniques with a classification accuracy of over 97%. When comparing the execution times of the entire features and the suggested approach for the model, it was observed that the strategy we presented reduces the execution time of AdaBoost while attaining a 0.974 AUROC score. Hence, this study proves the effectiveness of the proposed approach in classification of breast cancer disease. Finally, the enhancements gained from applying the suggested framework can be extended to other disease datasets in the future, such as brain tumor and stroke. To reduce the RFE’s temporal complexity, further sophisticated optimization techniques can be applied.

References 1. Lamba, R., Gulati, T., Jain, A.: A hybrid feature selection approach for Parkinson’s detection based on mutual information gain and recursive feature elimination. Arab. J. Sci. Eng. (2022). https://doi.org/10.1007/s13369-021-06544-0 2. Mehmood, M., Alshammari, N., Alanazi, S.A., Ahmad, F.: Systematic framework to predict early-stage liver carcinoma using hybrid of feature selection techniques and regression techniques. Complexity. 2022, (2022). https://doi.org/10.1155/2022/7816200 3. Tiwari, A., Chaturvedi, A.: A hybrid feature selection approach based on information theory and dynamic butterfly optimization algorithm for data classification. Expert Syst. Appl. 196, 116621 (2022). https://doi.org/10.1016/j.eswa.2022.116621 4. Amini, F., Hu, G.: A two-layer feature selection method using genetic algorithm and elastic net. Expert Syst. Appl. 166, 114072 (2021). https://doi.org/10.1016/j.eswa.2020.114072 5. Kalita, D.J., Singh, V.P., Kumar, V.: Two-way threshold-based intelligent water drops feature selection algorithm for accurate detection of breast cancer. Soft Comput. 26, 2277–2305 (2022). https://doi.org/10.1007/s00500-021-06498-3 6. Rouhi, A., Nezamabadi-Pour, H.: A hybrid feature selection approach based on ensemble method for high-dimensional data. In: 2nd Conference on Swarm Intelligence and Evolutionary Computation CSIEC 2017—Proceeding, pp. 16–20 (2017). https://doi.org/10.1109/CSIEC. 2017.7940163. 7. Sharaff, A., Nagwani, N.K., Swami, K.: Impact of feature selection technique on email classification. Int. J. Knowl. Eng. 1, 59–63 (2015). https://doi.org/10.7763/ijke.2015.v1.10 8. Hancer, E., Xue, B., Zhang, M.: Differential evolution for filter feature selection based on information theory and feature ranking. Knowl.-Based Syst. 140, 103–119 (2018). https://doi. org/10.1016/j.knosys.2017.10.028 9. Lu, H., Chen, J., Yan, K., Jin, Q., Xue, Y., Gao, Z.: A hybrid feature selection algorithm for gene expression data classification. Neurocomputing 256, 56–62 (2017). https://doi.org/10.1016/j. neucom.2016.07.080 10. Rtayli, N., Enneya, N.: Enhanced credit card fraud detection based on SVM-recursive feature elimination and hyper-parameters optimization. J. Inf. Secur. Appl. 55, 102596 (2020). https:// doi.org/10.1016/j.jisa.2020.102596 11. Chiew, K.L., Tan, C.L., Wong, K.S., Yong, K.S.C., Tiong, W.K.: A new hybrid ensemble feature selection framework for machine learning-based phishing detection system. Inf. Sci. (Ny) 484, 153–166 (2019). https://doi.org/10.1016/j.ins.2019.01.064 12. Priscilla, C.V., Prabha, D.P.: A two-phase feature selection technique using mutual information and XGB-RFE for credit card fraud detection. Int. J. Adv. Technol. Eng. Explor. 8, 1656–1668 (2021). https://doi.org/10.19101/IJATEE.2021.874615

224

H. Dhoke and A. Sharaff

13. Shao, Z., Yang, S.L., Gao, F., Zhou, K.L., Lin, P.: A new electricity price prediction strategy using mutual information-based SVM-RFE classification. Renew. Sustain. Energy Rev. 70, 330–341 (2017). https://doi.org/10.1016/j.rser.2016.11.155 14. Lin, X., Yang, F., Zhou, L., Yin, P., Kong, H., Xing, W., Lu, X., Jia, L., Wang, Q., Xu, G.: A support vector machine-recursive feature elimination feature selection method based on artificial contrast variables and mutual information. J. Chromatogr. B Anal. Technol. Biomed. Life Sci. 910, 149–155 (2012). https://doi.org/10.1016/j.jchromb.2012.05.020 15. Wang, H., Liu, S.: An effective feature selection approach using the hybrid filter wrapper. Int. J. Hybrid Inf. Technol. 9, 119–128 (2016). https://doi.org/10.14257/ijhit.2016.9.1.11

Chapter 20

2D-CTM and DNA-Based Computing for Medical Image Encryption Mobashshirur Rahman and Piyush Kumar

Abstract Medical imaging plays an important role in the proper treatment of a patient’s disease. Nowadays, the virtual diagnosis of patients is becoming popular, so these images are sent from one place to another via the network for diagnosis purposes. So, the medical images need to be secured from any illegal access and modification. To handle this situation, a novel key generation with medical image encryption and decryption algorithms is proposed based on chaos, DNA computing, and Mersenne Twister (MT). The key is generated using a proposed new 2D-Chaotic Tan Map (2D-TCM). Then, the image is passed through two levels of confusion and diffusion based on DNA computing, chaos, and MT to get a final, highly secured encrypted image. Extensive comparative evaluation of the results is performed on Xray images and MRI images using different security metrics, such as key space, key sensitivity, histogram analysis, entropy, and correlation analysis. The results show that the proposed model achieved a better result in comparison with other related techniques.

20.1 Introduction Nowadays, after the breakthrough of the novel COVID-19, the treatment of patients virtually is of utmost popularity. When examining the patient virtually, the medical image plays an important role in the proper diagnosis of the patient’s disease. Medical images are also important for analysis and research purposes. There are many modalities of medical images which are currently popular, such as Magnetic Resonance Imaging (MRI), X-ray, Computed Tomography (CT), and ultrasound. However, it is also important to protect these medical images from cyber attackers. The attackers may use medical images to leak the privacy of patients or also alter the data, which may lead to false diagnoses. Therefore, there are many schemes related to M. Rahman (B) · P. Kumar National Institute of Technology Patna, Patna, India e-mail: [email protected] P. Kumar e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_20

225

226

M. Rahman and P. Kumar

securing the medical images that are proposed, such as image steganography, image watermarking [1], and image encryption [2]. In all of these schemes, currently, the encryption method is the most popular for securing the data. DNA computing [3], Chaos, and cryptography are frequently used together to provide security. A number of technique is proposed for securing the medical image using encryption process. Ravichandran et al. [2] proposed DNA and chaos-based confusion and diffusion process to get encrypted image. Belazi et al. [4] have proposed DNA-based computing, SHA-256, and chaos-based encryption and decryption scheme for color as well as gray-level medical image. Zhang [5] proposed piecewise linear chaotic map (PWLCM) DNA-based confusion and diffusion operation for key generation and encryption. Traditional approaches have several disadvantages: high algorithm complexity, lack of proper evaluation, and encryption time. A novel model based on Deoxyribonucleic Acid (DNA), Mersenne Twister (MT), and 2D-Chaotic Tan map is proposed to efficiently secure the medical image from cyber-attacks. In this approach, two levels of confusion and diffusion are performed based on the proposed 2D-Chaotic Tan Map (2D-CTM) equation, and DNA computing is performed to get the final cipher image. The major contribution of this proposed approach can be summarized as: • A new 2D-chaotic map is proposed for generating private keys for confusion and diffusion processes. • In the proposed method, a private key is employed for confusion, followed by DNAbased diffusion with the private key. Finally, MT-based confusion is performed to get the final encrypted image. • Comparative result analysis is carried out using different security measures, such as entropy, key space, differential attack, correlation analysis, and robustness analysis. The result showed that the proposed model outperforms the other existing related models. The paper has 5 sections. The next section is about recent related works. In Sect. 20.3 , the proposed methodology is discussed. Section 20.4 has the results and discussion part. And finally, the paper is concluded in Sect. 20.5.

20.2 Literature Survey Belazi et al. [4] have proposed a novel medical image security algorithm based on DNA encoding, SHA-256, and bit-level pixel confusion and diffusion. The logisticChebyshev chaotic map and sine-Chebyshev chaotic map are used for generating key streams. Extensive result analysis is carried out using different security metrics. The algorithm has achieved better time complexity as well in comparison with other existing algorithms. However, the author has not considered the issue related to the storage cost of medical images. Zhang [5] proposed a new DNA and chaos-based image encryption algorithm. In this paper, the author has used piecewise linear chaotic

20 2D-CTM and DNA-Based Computing for Medical Image Encryption

227

map (PWLCM) to generate key sequences. Two levels of DNA-based confusion and diffusion are performed to get the final encrypted image. Abdelfatah [6] has proposed a color image encryption with authentication technique based on MT, hashing, digital signature, and conic curve. The image is divided into red, green, and blue then diffusion and hashing along with digital signature to get final encrypted image. Hosny et al. [7] have proposed a chaos-based image encryption scheme. In this proposed work, the authors have performed a block-based followed by a pixel-level confusion process. The zig-zag-based scan is used for pixel scrambling. Then, the diffusion process is carried out by a 1D-logistic map to get the final encrypted image. However, the 2D map can be used for diffusion and confusion processes to get better security. Guan et al. [8] have proposed hyperchaos and DNA-based pixel scrambling and confusion process in frequency domain to get secured encrypted image. Priyanka and Kumar Singh [9] have proposed a lightweight medical image encryption scheme. In this scheme, singular value decomposition is used along with a pixelbased confusion process to get the final encrypted image. The scheme achieves a good result in less time. Yousif et al. [10] have proposed a new image encryption algorithm based on bit-level replacement, chaos, and DNA-based encoding techniques. First, bit replacement is performed to generate two images, and then, the diffusion process is performed with the key. Finally, DNA-based encoding and decoding are performed to get the encrypted image.

20.3 Methodology Medical images are sent over the Internet and kept on servers for diagnosis and analysis purposes. However, there is a need to secure medical images from unauthorized access or modification. Therefore, a novel algorithm is proposed for efficiently securing digital images on the Internet from cyber-attacks. In this approach, two levels of confusion and diffusion of the original image are performed using a new 2D-chaotic map, DNA computing, and MT, to get a highly secure encrypted image.

20.3.1 Key Generation Key generation is the most important step for getting a secure encrypted form of the original image. In this approach, a new 2D-chaotic map is designed to generate the key stream. The 2D-Chaotic Tan Map (2D-CTM) is given by the following equation. 

pn+1 = k + pn ∗ tan(m) mod 1, qn = k + m ∗ (1 − pn+1 ∗ tan(m)) mod 1,

where pn , k , and m is the initial control parameter value.

(20.1)

228

M. Rahman and P. Kumar

In the above equation, the first initial value for pn , k, and m is chosen to obtain the initial sequence. Using this map, two sequences are generated, named as X1 and Y 1. The X1 is used for the diffusion process, and the Y 1 is used for the confusion process in this algorithm. The algorithmic steps for the key generation are shown in Algorithm 1. Algorithm 1: Proposed Key Generation Steps

1 2 3 4 5 6 7 8 9 10

Input : s, k, and m are initial control parameter. Also, give desired size of the key as parameter Output: Secret keys sequences, X1 and Y1 key of given size Function GenerateKey(Mi ): X1 ← initialize empty array for key sequence 1 Y1 ← initialize empty array for key sequence 2 for i in range(size) do s = k + ((s) ∗ tan(m)) % 1 X 1.append(int ((s ∗ pow(10, 10)) % (256))) s = k + (m ∗ (1 − s ∗ sin(m))) % 1 Y 1.append(int ((m ∗ pow(10, 10)) % (256))) end return X1, Y1

20.3.2 Encryption The encryption is the final step for generating the cipher image from its original form. So, after the successful generation of the private key, the encryption is performed. First, a pixel-level shuffling of the original image is performed using the Y 1 sequence generated using 2D-CTM. S_image[Y 1[i]] = Original_image[i]

(20.2)

where S_image denote modified image, Original_image is the original image, and i denotes the index value. After getting the shuffled image, encode the pixels of the image into DNA sequences using rule Rule 1. Also, encode the value of key X1 into the DNA sequence. Rule 1 ← (Key[i]%8) + 1

(20.3)

Rule 2 ← (i%8) + 1

(20.4)

where i ∈ integer and key is private encryption key. After converting the values of both images and keys to DNA sequences, perform a DNA-based XOR operation to obtain a diffused image (Tables 20.1 and 20.2).

20 2D-CTM and DNA-Based Computing for Medical Image Encryption

229

Algorithm 2: Image Encryption

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29

Input : Medical images Mi , where i = 1, 2, 3...n, and Key X1,Y1 along with their DNA encoded values Output: Encrypted images Ci , where i = 1, 2, 3...n. Function Encrypt_Image(Mi ): Y 1, X 1 = 2D − C T M() sequence2 = Mer senne_T wister () sum ← sum of pixels of original image R1 ← (K ey[i]%8) + 1 R2 ← (i%8) + 1 D N A Ri ← Encodepi xel with r ule Ri Decode Ri ← DecodeD N A with r ule Ri size = M*N X1 D N A ← DNA based encode form of key X1 for i from 0 to size) do S_image [Y 1[i] − 1] = Mi [i] end for i from 0 to size do Encode_dna[i] = D N A R1 end for i from 0 to size do Encode_dna[i] = Encode_dna[i] XOR X1 D N A [i] end for p in Encoded_dna do Enc_dna= Decode R2 end for j from 0 to size do Encr ypted_image [sequence2 [ j] − 1] = Enc_dna[ j] end for j in range(M*N) do Ci pher _image [ j] = Ci pher _image [ j] ⊕ X 1 ⊕ (Sum % 256) end return Ci

Now, after the diffusion process, decode the DNA sequence back to the integer values. Now, perform MT-based confusion of the cipher image [6]. Finally, the XOR operation is performed among the sum of pixels of the original image, X1 and the cipher image to get a highly secured final encrypted image. The complete block diagram of encryption process is shown in Fig. 20.1.

20.3.3 Decryption The reverse step of encryption needs to be followed to recover the original image from its encrypted form. First, reshuffling of pixels is performed based on the Y 1 sequence. Then, perform DNA-based encoding of pixels. The DNA-based XOR operation is carried out with the X1 key. After that, decode the pixels of the image

230

M. Rahman and P. Kumar

Table 20.1 DNA-based rule for encoding and decoding Rules A T 1 2 3 4 5 6 7 8

00 00 11 01 10 01 10 11

11 11 00 10 01 10 1 00

Table 20.2 Rule for XOR operation of DNA sequences XOR A T T C A G

T C A G

A G T C

C

G

10 01 10 11 00 00 11 01

01 10 01 00 11 11 00 10

C

G

G A C T

C T G A

Key Generator ORIGINAL IMAGE

2D-TCM Y1

X1

SHUFFLED IMAGE

DNA BASED XOR RULE 1

XORED DNA SEQUENCE

ENCODED DNA SEQUNECE

RULE 2

X1 MT BASED SHUFFLING

SUM OF PIXELS

ENCRYPTED IMAGE

Fig. 20.1 Proposed encryption flow

DECODING DNA

DECODED SEQUENCE

20 2D-CTM and DNA-Based Computing for Medical Image Encryption

231

into integers. Now, reshuffle using the MT-based sequence used during encryption. Finally, perform an XOR-based operation between the image, X1, and the sum of pixels of the original image to get back the original image.

20.4 Performance Analysis A novel medical image encryption technique is proposed based on chaos, MT, and DNA computing. The 2D-CTM is used for key generation, and a DNA-based operation is performed to get the final encrypted image. The performance of the proposed model is evaluated thoroughly using different security evaluation metrics.

20.4.1 Experimental Setup and Dataset Details The proposed scheme is tested on an Asus laptop with an Intel Core i5 processor, 8 GB of RAM, a 512 GB SSD hard disk, and a 4 GB Nvidia graphics card. The result evaluation is performed using different modalities of medical images, namely X-ray and MRI images. All these images are taken from the Medpix website [11].

20.4.2 Results and Discussion 2D-Chaotic Tan Map (2D-TCM) In this subsection, the discussion of 2D-TCM and related chaotic map is performed. The Sine map [12] is chaotic 1D map which can be described by the following equation. Sn+1 = μ sin(π Sn ) (20.5) where μ ∈ [0.87, 1] for chaotic behavior. A new 2D map is proposed which is also based on trigonometry tan(x), which is given by the following equation. 

pn+1 = k + pn ∗ tan(m) mod 1, qn = k + m ∗ (1 − pn+1 ∗ tan(m)) mod 1,

(20.6)

where pn , k, and m is the initial control parameter value. Here, each time initial value pn get updated and added to the key sequence. Now, the bifurcation diagram is drawn for the Sine map and the 2D-CTM in Fig. 20.2a, b, respectively. It can be seen that the proposed 2D-CTM has no pattern as compared to the Sine map. Thus, the 2D-CTM is very sensitive to initial conditions as well.

232

M. Rahman and P. Kumar

Fig. 20.2 Bifurcation diagram of a sine map and b 2D-CTM, c Lyapunov exponent diagram of 2D-CTM in range [0, 4]

Fig. 20.3 a Original image, b encrypted image with p = 0.93, c decrypted image with slight modification in p value as 0.930000000000001, and d decrypted image with correct p value

Now, the Lyapunov Exponent (LE) is calculated for 2D-CTM, and it is shown in Fig. 20.2c. The LE of 2D-CTM has a positive value just above 0 and is positive. And a positive LE means the map is chaotic at that point. So, the proposed map has a large range of chaotic behavior from (0−∞). Thus, the sequence generated using 2D-CTM has more chaotic behavior and an unpredictable trajectory. So, it can be perfectly used for the key generation process. Key Space Key space is defined as the total size of the key used for the encryption process. It determines the restiveness of the model against exhaustive search attacks. In the 2D-CTM-based key generation process, the initial parameters pn , k, and m have a precision of at least 214 . So, the resulting key space becomes (1014 )3 = 1042 , which is large enough to prevent exhaustive search attacks [12]. Thus, the key generated by the proposed model has a restiveness toward exhaustive search attacks. Key Sensitivity The key sensitivity shows the susceptible algorithm with a slight modification in key value. The 2D-CTM is highly sensitive to a minute change in the initial parameters of the key. For evaluation of 2D-CTM toward key sensitivity, a secret key is generated with an initial parameter of pn = 0.93, k = 90, and u = 4. And the given image is encrypted. Now, again the decryption key is generated, keeping all the parameters the same except slightly modifying the value of pn from 0.93 to 0.930000000000001. Now, with this key, the decryption is carried out. The result of decryption is shown in Fig. 20.3. And it is clearly visible that with a slight modification to the initial parameters, the decryption process completely failed.

20 2D-CTM and DNA-Based Computing for Medical Image Encryption

233

Fig. 20.4 Histogram of original image and its corresponding cipher image for a x-ray image and b MRI image Table 20.3 Information entropy comparative analysis for cipher images Method Guan et al. [8] Akkasaligar and 2D-CTM Biradar [13] Entropy

7.992

7.890

7.999

Histogram Analysis The histogram analysis shows the frequency of different pixel values. If the pixel distribution of the cipher image is uniform, this means that the cipher image can prevent statistical attacks as the attackers are not able to find any pattern in the image. Figure 20.4 shows the histograms of original images and their corresponding cipher images. Entropy Information entropy measures the level of randomness present in the cipher image. It is calculated using Eq. 20.7. The ideal value of entropy is 8. The entropy value for the cipher images of X-ray and MRI is calculated as 7.9992 and 7.9993, which is very close to the ideal value. Table 20.3 shows the comparative analysis of the information entropy of the proposed scheme with other related schemes. IE(x) = −

n 

p(xi ) log2 p(xi ).

(20.7)

i=1

where IE stands for Information Entropy, and p(xi) is probability of pixel xi in the image. Correlation Analysis Correlation analysis determines the way in which adjacent pixels are related to each other. If there is a pattern among pixels, then the attackers may launch an attack using this pattern. So, ideally the correlation coefficient should be 0, i.e., it means there is no correlation between adjacent pixels. The correlation of the X-ray and MRI images is calculated, and the comparative result is shown in Table 20.4.

234

M. Rahman and P. Kumar

Table 20.4 Performance comparison for correlation coefficients Methods Correlation Guan et al. [8] Akkasaligar and Biradar [13] 2D-CTM

Horizontal 0.0011 0.0194

Diagonal 0.0001 0.0195

Vertical −0.0002 0.0195

0.0009

−0.0002

−0.0021

20.5 Conclusion A novel algorithm based on DNA computing, chaotic map, and Mersenne twister is proposed for securing medical images from cyber-attacks. In the proposed algorithm, a new chaotic map, 2D-CTM, is proposed to generate a private key for the confusion and diffusion process. After key generation, the encryption process is carried out by performing confusion based on sequences generated by 2D-CTM followed by DNA encoding. The DNA-based XOR operation is carried out between the key and the DNA-encoded image. Then, the image is decoded and MT-based shuffling is performed. Finally, the XOR-based encryption is performed among the image, key, and sum of pixels of the original image, to get a highly secured encrypted image. The comparative result evaluation is performed on X-ray images and MRI images considering different security metrics, such as key space, histogram analysis, key sensitivity, entropy, and correlation analysis. The result of the proposed model achieves a better result than other existing models.

References 1. Evsutin, O., Melman, A., Meshcheryakov, R.: Digital steganography and watermarking for digital images: a review of current research directions. IEEE Access 8, 166589–166611 (2020) 2. Ravichandran, D., Banu S.A., Murthy, B.K., Balasubramanian, V., Fathima, S., Amirtharajan, R.: An efficient medical image encryption using hybrid DNA computing and chaos in transform domain. Med. Biol. Eng. Comput. 59(3), 589–605 (2021) 3. Namasudra, S., Chakraborty, R., Majumder, A., Moparthi, N.R.: Securing multimedia by using DNA-based encryption in the cloud computing environment. ACM Trans. Multimedia Comput. Commun. Appl. 16(3s), 1–19 (2021) 4. Belazi, A., Talha, M., Kharbech, S., Xiang, W.: Novel medical image encryption scheme based on chaos and DNA encoding. IEEE Access 7, 36667–36681 (2019) 5. Zhang, Y.: The image encryption algorithm based on chaos and DNA computing. Multimedia Tools Appl. 77(16), 21589–21615 (2018) 6. Abdelfatah, R.I.: A color image authenticated encryption using conic curve and Mersenne twister. Multimedia Tools Appl. 79(33–34), 24731–24756 (2020) 7. Hosny, K.M., Kamal, S.T., Darwish, M.M.: A color image encryption technique using block scrambling and chaos. Multimedia Tools Appl. 81(1), 505–525 (2021) 8. Guan, M., Yang, X., Hu, W.: Chaotic image encryption algorithm using frequency-domain DNA encoding. IET Image Process. 13(9), 1535–1539 (2019)

20 2D-CTM and DNA-Based Computing for Medical Image Encryption

235

9. Priyanka, Kumar Singh, A.: FastMIE: faster medical image encryption without compromising security. Measurement 196, 111175 (2022) 10. Yousif, S.F., Abboud, A.J., Alhumaima, R.S.: A new image encryption based on bit replacing, chaos and DNA coding techniques. Multimedia Tools Appl. 81, 27453–27493 (2022). https:// doi.org/10.1007/s11042-022-12762-x 11. MedPix: U.S. National Library of Medicine [Online]. Available: https://medpix.nlm.nih.gov/ home. Accessed: 04-May-2022 12. Mansouri, A., Wang, X.: A novel one-dimensional sine powered chaotic map and its application in a new image encryption scheme. Inf. Sci. 520, 46–62 (2020) 13. Akkasaligar, P.T., Biradar, S.: Selective medical image encryption using DNA cryptography. Inf. Secur. J. Glob. Perspect. 29(2), 91–101 (2020)

Chapter 21

A Review of Financial Fraud Detection in E-Commerce Using Machine Learning Abhay Narayan, S. D. Madhu Kumar, and Anu Mary Chacko

Abstract Online shopping, banking, financial institutions, and government all use e-commerce. Fraudulent conduct is seen in a variety of industries and in our daily lives. The key to lowering losses is to create powerful and efficient fraud detection system. For the identification of fraud, a wide range of techniques and methods are used today. The machine learning methodologies utilized in e-commerce fraud detection, the weaknesses of such methods, and avenues for new research in this area are reported systematically in this review paper.

21.1 Introduction E-commerce is the process of purchasing or selling goods or services through the Internet. Because of the convenience and quickness with which people can purchase online, it is becoming increasingly popular. Meanwhile, with the rise of e-commerce, there has been a significant increase in fraud activities leading to cheating of the users who use e-commerce domain for buying, selling, shopping, etc. The total volume of activities, such as clicks, purchases, and reviews, are used by most e-commerce platforms to generate a ranking index and a reputation factor for merchants. Products with a higher ranking index will often appear at the top of the search results, and consumers choose items with a good reputation. Regular strategies to improve these metrics include producing high-quality goods, giving outstanding services, and promoting. All of these require a significant amount of effort. It encourages some unscrupulous merchants to advertise their products through spam operations. There are three types of frauds found in general in e-commerce: Transactions Fraud: Fraud transactions are acts in which merchants illegally gain the rights and interests of e-commerce platforms. A. Narayan · S. D. Madhu Kumar (B) · A. M. Chacko National Institute of Technology Calicut, Calicut, India e-mail: [email protected] A. M. Chacko e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_21

237

238

A. Narayan et al.

Opinion fraud: Spammers assess items with a significant prejudice. Typically, they post fake reviews in order to mislead consumers’ purchasing decisions. Ad-Click Fraud: Click fraud is the illegal act of clicking on pay-per-click (PPC) adverts in order to boost site income or deplete a company’s advertising budget. The detection of fraudulent activity is a great challenge. Fraudsters attack in a variety of ways, and there is no one-size-fits-all methodology or approach to prevent fraud. In this review paper, we present a comprehensive survey of different fraud detection methods, various evaluation measures used for checking these methods, major challenges, and research directions in this filed. The remainder of this paper is organized as follows: Sect. 21.2 discusses the different fraud detection techniques. Section 21.3 reviews the different machine learning-based state-of-the-art methods in financial fraud identification and current research works in e-commerce fraud detection. In Sect. 21.4, review of community-based fraud detection approaches is outlined. Section 21.5 gives an overview of different evaluation measures used to evaluate a machine learning technique. In Sect. 21.6, challenges in fraud detection are laid down. Section 21.7 concludes the paper.

21.2 Fraud Detection Techniques In this section, various fraud detection techniques which are existing currently are discussed.

21.2.1 Rule-Based Fraud Detection A rules-based model is a set of rules that may be used to spot fraudulent activity. A single rule is made up of a set of circumstances that, if met, classify an action as fraudulent or potentially fraudulent.

21.2.1.1

Strengths and Vulnerabilities

The advantages of rule-based systems are as follows: I. If a certain rule generated an alert for a specific action, the reason for the alert is quite clear. II. There is no need to collect training datasets for machine learning algorithms because the rules are operational from the start. III. The People working in the e-commerce firm can easily develop rules because they are already familiar with the fraud activities and the preventive or related activities to be taken after the fraud detection.

21 A Review of Financial Fraud Detection in E-Commerce …

239

The disadvantages of rule-based systems are as follows: I. As new fraud patterns arise, new rules must be devised to reverse engineer the assaults of fraudsters. II. The cost of maintenance increases as the number of rules grows. III. The human understanding of rule-based systems is limited (due to manual development of rules and necessary maintenance).

21.2.2 Machine Learning for Fraud Detection The drawbacks of rule-based systems are addressed by machine learning models. They flourish in contexts with high data volume and dimensionality. Decision trees, random forests, gradient boosting, and neural networks are examples of algorithms that are meant to uncover complicated, nonlinear patterns using hundreds (if available) of data. Such an approach necessitates a shift in perspective. For starters, deploying machine learning models necessitates the usage of high-quality, labeled historical data as a training dataset. The more data the e-commerce Firms have (in terms of the number of transactions and the number of features capturing transaction attributes), the better will be the model’s performance.

21.2.2.1

Strengths and Vulnerabilities

The advantages of machine learning-based systems are as follows: I. Automatic fraud pattern recognition—The algorithm is responsible for determining what constitutes a fraud. It is our job to provide it as much information as we can (in form of a feature vector). II. Minimum Manual Intervention—Many of the procedures are programmable. Companies with well-developed machine learning pipelines spend the majority of their time exploring new features and algorithms while monitoring the performance metrics of existing models using monitoring applications. The disadvantage of machine learning-based systems is that a substantial quantity of historical data is required to run ML models. In the next section, we present a summary of the existing work on the use of machine learning in economic fraud detection.

21.3 Literature Survey Recent strategies have begun to concentrate on developing machine learning models to automatically discover inherent patterns from earlier fraud data, which can be

240

A. Narayan et al.

classified into three: supervised, unsupervised, and semi-supervised fraud detection systems.

21.3.1 Supervised Fraud Detection Previously, supervised algorithms for detecting online fraud were mostly used, based on the premise that fraud activities follow specific patterns. In general, these approaches gather statistical characteristics from many sources, such as previous behavior and user profiles, and then employ conventional classifiers such as neural networks [16] and SVM [28] for fraud detection. To detect fraudulent transactions, these methods extracted the most important features from online payment data and employed numerous supervised machine learning algorithms. Ensemble learning methods were used to create a classifier. The detailed review is given in Table 21.1.

21.3.2 Unsupervised Fraud Detection The unreasonable assumption that all fraud data label information is available is made by supervised fraud detection algorithms. As a result, unsupervised approaches are being studied in order to automatically identify unrevealed patterns in fraud data. Shchur et al. [40] addressed the fraud detection issue in temporal graphs and detected suspicious behaviors using unsupervised approaches such as dense sub-block finding. Sun et al. [42] introduced FraudVis, a system for visualizing unsupervised fraud detection methods from several viewpoints, including temporal, intra-group correlation, and inter-group correlation. The detailed review is given in Table 21.2.

21.3.3 Semi-supervised Fraud Detection Labeled fraud data is frequently scant in practical e-commerce platforms, necessitating the use of effective semi-supervised fraud detection methods. Several researchers have recently started to use graph-based approaches for semi-supervised fraud detection [32, 48, 50]. The paper [48] presented a collaboration-based multi-label propagation technique for exploiting label correlations and accelerating it by directly recognizing communities on the user-item interaction network. Li et al. [27], to identify spam reviews on Xianyu App, use users and goods as nodes in a bipartite network and associate reviews as edge characteristics. Based on its local heterogeneous information and global context, a heterogeneous GNN is suggested to classify whether a review is spam or not. Binbin et al. [17] consider credit payment service customers, merchants, and devices

21 A Review of Financial Fraud Detection in E-Commerce …

241

Table 21.1 Supervised learning-based fraud detection techniques Machine learning

Method adopted

Problem considered

References

Advantage

Disadvantage

Supervised learning

Support vector machine

E-commerce, credit card fraud

[28]

Accurate and trustworthy

It is necessary to comprehend and label the input. For the training phase, more computation time is required

Neural network Fraud transactions

[16]

Neural network Spam reviews

[31]

Ensemble learning

[10]

Fraud transactions

Support vector Click fraud machine, Naive Bayes

[6]

KNN, random forest

Click fraud

[18]

Bayesian network classifiers

Fraud transaction

[19]

Logistic regression

Fraud transaction

[47]

Support vector machine

Spam reviews

[24]

Support vector machine

Spam reviews

[20]

Decision tree

Transaction fraud

[25]

Naive Bayes

Spam reviews

[35]

Random forest

Click fraud

[7]

Logistic regression Naive Bayes

Dark market analysis transaction fraud

[2]

Ensemble learning

Click fraud transaction fraud

[36]

Supervised learning-torank unit

Rank influential products spam review

[1]

242

A. Narayan et al.

Table 21.2 Unsupervised learning-based fraud detection techniques Machine Method used Problem References Advantage learning considered Unsupervised learning

K-means

Fraud transactions

HMM

Fraud [33] transactions Spam reviews [52]

Unsupervised matrix iteration algorithm Ensemble learning Fuzzy clustering HMM

Click fraud

[38]

Unknown patterns and characteristics of data are simple to find

Disadvantage Complex in terms of computation. Because of the unlabeled input, it is less accurate

[44]

Fraud [5] transactions Fraud [4] transactions Lexicon-based Spam reviews [23] method StatisticsSpam reviews [9] based unsupervised clustering Autoencoder Click fraud [43] GAN Fraud [12] transaction Mixture Spam review [26] models LSTM Transaction [22] fraud

as various types of nodes and their interactions as edges in heterogeneous graph, and propose a metapath-based heterogeneous graph embedding approach dubbed HACUD to identify cash-out users. Details are summarized in Table 21.3.

21.4 Community-Based Fraud Detection on E-Commerce The majority of existing efforts concentrate on detecting individual fraud users or transactions. However, in large-scale e-commerce networks, fraudulent activity is

21 A Review of Financial Fraud Detection in E-Commerce …

243

Table 21.3 Semi-supervised learning-based fraud detection techniques Machine Method used Problem References Advantage learning considered Semisupervised learning

Graph-based

Malicious promotion

[50]

Neural network Convolutional Neural network Deep Feed forward NN Thresholdbased detection Co-training by features approach NN

Click fraud

[43]

Transaction fraud

[13]

Transaction fraud Spam review

[25]

Spam review

[53]

Malicious account Convolutional Spam review neural networks RNN Transaction fraud

[32]

It reduces the amount of annotated data used

Disadvantage Iteration results are not stable. It has low accuracy

[29]

[27]

[37]

becoming more complex, and organized groups have arisen. Wang and Paschalidis [49] looked into the detection of anomalous bots in cyberspace. They suggested a twostage detection technique, with a flow-based anomaly detector, detecting individual aberrant nodes and a graph-based detector determining a fraud community based on their temporal connections. Sarma et al. [39] developed a bank anti-fraud guard system using on community-based detection algorithms to detect inherent trends that lead to fraud occurrences. Gangopadhyay and Chen [14] created two different kinds of methods for detecting exclusive tiny fraud communities for physicians-patients graphs in a healthcare context. The research [45] formalized the problem of identifying crowd fraud in online advertising. Investigators studied crowd fraud behaviors and discovered three crucial characteristics: synchronization, moderateness, and dispersivity. To detect the fraud community, the researchers also presented data construction, coalition filtering, and DP-means clustering. In [30], the authors presented a novel fraud community detection architecture that leverages the strengths of both offline and online data perspectives, mostly spatial-temporal data that is maintained offline. The detailed review of community-based fraud detection is given in Table 21.4.

244

A. Narayan et al.

Table 21.4 Community-based fraud detection approach Machine learning Method used Problem considered

References

Unsupervised learning Semi-supervised learning Semi-supervised learning Unsupervised learning

DP-means clustering Multi-view graph neural network Graph analytics, SVM

Crowd fraud detection [45] Group of malicious [30] sellers on e-commerce Transactional fraud [34]

Network analysis, k-modes clustering

Semi-supervised multi-label learning

Label propagation algorithm

Unsupervised multi-label learning Semi-supervised

Reviewer2Vec algorithm Graph based

Unsupervised supervised

Expectation maximization algorithm

Identify candidate groups for transactional fraud Detects communities on the user-item graph for spam behaviors Detect online fraud reviewer groups Online users to detect auction fraud Detect online fraud reviewer groups

[46]

[48]

[11] [3] [51]

21.5 Evaluation Measures Different mathematical measurements have been employed in research investigations to evaluate the outcome of suggested algorithms. The performance of the suggested methods has been evaluated using conventional criteria on the basis of receiver operating characteristic (ROC) or precision-recall (PR) curves for those with adequate labeled data. In machine learning, ROC curves are widely used to show the findings of binary situation issues [8, 21]. However, ROC does not give much insight into severely skewed datasets, whereas PR curves provide a more meaningful view of an algorithm’s performance. When it comes to fraud detection, the number of negative samples far outnumbers the number of positive cases. As a result, a significant change in the number of false positives might result in just a minor change in the false-positive rate utilized in the ROC analysis [8, 21]. Furthermore, by comparing false positives to true positives rather than true negatives, precision reflects the influence of many negative cases on the algorithm’s performance. The F-measure is the second most used performance metric. Given the nature of fraud concerns, this preference for accuracy over the next metric is expected. Accuracy emphasizes true negative, which is irrelevant in fraud detection. In an unequal class distribution, a measure that places a stronger emphasis on false negative (FN) and false positive (FP) is more useful. The F-measure is recommended in these situations because it balances accuracy and recall, resulting in a more accurate estimation of a fraud detection model.

21 A Review of Financial Fraud Detection in E-Commerce …

245

Goix [15] proposed two new ways for evaluating the efficacy of anomaly detection approaches using dimensional data without data labels, namely excess-mass (EM) and mass-volume (MV) curves. Fraud community structure identification is a challenging problem. In community detection, the availability of appropriate criteria for evaluating whether a technique identifies substantial groupings of nodes that best suit the underlying community organization is a critical issue. To assess the accuracy of community discovery algorithms, the Normalized Mutual Information (NMI) method has been frequently employed [41].

21.6 Challenges in Fraud Detection We have highlighted in this paper many important problems that are to be addressed by the research community while focusing on effective solutions to detect the fraud activities in the e-commerce sector. We are summarizing the major challenges here. • Existing e-commerce fraud detection systems confront a problem: They have a hard time adapting to the ongoing mutation of fraud patterns since they rely heavily on “fraudster seeds” established by domain specialists. This is the most difficult to combat since fraudsters are continually looking for new and inventive methods to get past the systems and perform the crime. As a result, updating fraud detection models with evolving patterns to detect becomes critical. The model’s performance and efficiency suffer as a result of this. As a result, machine learning models must be updated on a regular basis or else they would fail to achieve their goals. • Class Imbalance: Only a small number of clients are attempting to defraud eCommerce systems. As a result, there is an imbalance in the categorization of fraud detection models, making it more difficult to develop them. Because detecting the fraudsters frequently entails refusing some valid actions, the result of this difficulty is a bad user experience for genuine clients. • Model Interpretations: This restriction is related to the idea of explainability, because there has been little study on the impact of model explanations on the quality and efficiency of fraud investigator choices. • E-commerce businesses are now growing into other sectors, such as new nations or platforms. Correlation of user behaviors across multi-platforms is neglected.

21.7 Conclusion In this paper, we gave a broad review of financial fraud detection procedures based on three types of spam activity: Transactions Fraud, Opinion Fraud, and Click Fraud. From the perspective of the research community working on financial fraud detec-

246

A. Narayan et al.

tion, we have outlined a review of financial fraud detection procedures. The unprecedented epidemic that rocked the global financial system has intensified digital change, which has resulted in more subtle forms and more clever financial fraud schemes. In terms of data, the current tendency is to use more panoramic data for identifying fraudulent activities. From basic quantitative data to the contemporary multi-source unstructured data, the data employed in fraud detection procedures has evolved. Deep learning approaches have recently acquired popularity due to their versatility and breakthrough efficacy in identifying financial fraud in terms of the model. The graph-based detection technique is a new way to look at multi-source data on fraud activity. Digital money transactions and behaviors of the stakeholders are growing more intelligent and complex as technology advances. The present condition of financial fraud and the research attempts to detect them, particularly in e-commerce, is examined in this paper. We find that there is an urgent need for the research community to focus on this important area of fraud detection and prevention in the e-commerce area.

References 1. Al-Nabki, M.W., Fidalgo, E., Alegre, E., Chaves, D.: Content-based features to rank influential hidden services of the tor darknet. arXiv preprint arXiv:1910.02332 (2019) 2. Al Nabki, M.W., Fidalgo, E., Alegre, E., De Paz, I.: Classifying illegal activities on tor network based on web textual contents. In: Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, pp. 35–43 (2017) 3. Bangcharoensap, P., Kobayashi, H., Shimizu, N., Yamauchi, S., Murata, T.: Two step graphbased semi-supervised learning for online auction fraud detection. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pp. 165–179. Springer (2015) 4. Batani, J.: An adaptive and real-time fraud detection algorithm in online transactions. Int. J. Comput. Sci. Bus. Inform. 17(2), 1–12 (2017) 5. Behera, T.K., Panigrahi, S.: Credit card fraud detection: a hybrid approach using fuzzy clustering & neural network. In: 2015 Second International Conference on Advances in Computing and Communication Engineering, pp. 494–499. IEEE (2015) 6. Beránek, L., N`ydl, V., Remeš, R.: Click stream data analysis for online fraud detection in e-commerce. In: INPROFORUM 2016 (2017) 7. Berrar, D.: Learning from automatically labeled data: case study on click fraud prediction. Knowl. Inf. Syst. 46(2), 477–490 (2016) 8. Davis, J., Goadrich, M.: The relationship between precision-recall and roc curves. In: Proceedings of the 23rd International Conference on Machine Learning, pp. 233–240 (2006) 9. Deng, S.: Deceptive reviews detection of industrial product. Int. J. Serv. Oper. Inform. 8(2), 122–135 (2016) 10. Dhankhad, S., Mohammed, E., Far, B.: Supervised machine learning algorithms for credit card fraudulent transaction detection: a comparative study. In: 2018 IEEE International Conference on Information Reuse and Integration (IRI), pp. 122–125. IEEE (2018) 11. Dhawan, S., Gangireddy, S.C.R., Kumar, S., Chakraborty, T.: Spotting collective behaviour of online frauds in customer reviews. arXiv preprint arXiv:1905.13649 (2019) 12. Fiore, U., De Santis, A., Perla, F., Zanetti, P., Palmieri, F.: Using generative adversarial networks for improving classification effectiveness in credit card fraud detection. Inf. Sci. 479, 448–455 (2019)

21 A Review of Financial Fraud Detection in E-Commerce …

247

13. Fu, K., Cheng, D., Tu, Y., Zhang, L.: Credit card fraud detection using convolutional neural networks. In: International Conference on Neural Information Processing, pp. 483–490. Springer (2016) 14. Gangopadhyay, A., Chen, S.: Health care fraud detection with community detection algorithms. In: 2016 IEEE International Conference on Smart Computing (SMARTCOMP), pp. 1–5. IEEE (2016) 15. Goix, N.: How to evaluate the quality of unsupervised anomaly detection algorithms? arXiv preprint arXiv:1607.01152 (2016) 16. Guo, Q., Li, Z., An, B., Hui, P., Huang, J., Zhang, L., Zhao, M.: Securing the deep fraud detector in large-scale e-commerce platform via adversarial machine learning approach. In: The World Wide Web Conference, pp. 616–626 (2019) 17. Hu, B., Zhang, Z., Shi, C., Zhou, J., Li, X., Qi, Y.: Cash-out user detection based on attributed heterogeneous information network with a hierarchical attention mechanism. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 33, pp. 946–953 (2019) 18. Iqbal, M.S., Zulkernine, M., Jaafar, F., Gu, Y.: Fcfraud: fighting click-fraud from the user side. In: 2016 IEEE 17th International Symposium on High Assurance Systems Engineering (HASE), pp. 157–164. IEEE (2016) 19. Jagdish, S., Singh, M., Yadav, V.: Credit card fraud detection system: a survey. J. Xidian Univ. 14(5) (2020) 20. Jalther, D., Priya, G.: Reputation reporting system using text based classification 21. Jeni, L.A., Cohn, J.F., De La Torre, F.: Facing imbalanced data–recommendations for the use of performance metrics. In: 2013 Humaine Association Conference on Affective Computing and Intelligent Interaction, pp. 245–251. IEEE (2013) 22. Jurgovsky, J., Granitzer, M., Ziegler, K., Calabretto, S., Portier, P.-E., He-Guelton, L., Caelen, O.: Sequence classification for credit-card fraud detection. Expert Syst. Appl. 100, 234–245 (2018) 23. Kamalesh, M.D., Diwedi, H.K.: Extracting product features from consumer reviews and its applications. Int. J. Appl. Eng. Res. 10(2), 2345–2351 (2015) 24. Khurshid, F., Yan Zhu, X., Zhuang, M.A., Ahmad, M.: Enactment of ensemble learning for review spam detection on selected features. Int. J. Comput. Intell. Syst. 12(1), 387 (2018) 25. Kim, E., Lee, J., Shin, H., Yang, H., Cho, S., Nam, S.-k., Song, Y., Yoon, J.-a., Kim, J.i.: Champion-challenger analysis for credit card fraud detection: hybrid ensemble and deep learning. Expert Syst. Appl. 128, 214–224 (2019) 26. Kumar, N., Venugopal, D., Qiu, L., Kumar, S.: Detecting anomalous online reviewers: an unsupervised approach using mixture models. J. Manag. Inf. Syst. 36(4), 1313–1346 (2019) 27. Li, A., Qin, Z., Liu, R., Yang, Y., Li, D.: Spam review detection with graph convolutional networks. In: Proceedings of the 28th ACM International Conference on Information and Knowledge Management, pp. 2703–2711 (2019) 28. Li, C., Ding, N., Dong, H., Zhai, Y.: Cs-svm. Int. J. Mach. Learn. Comput. 11(1) (2021) 29. Li, Y., Lin, Y., Zhang, J., Li, J., Zhao, L.: Highlighting the fake reviews in review sequence with the suspicious contents and behaviours. J. Inf. Comput. Sci. 12(4), 1615–1627 (2015) 30. Li, Z., Hui, P., Zhang, P., Huang, J., Wang, B., Tian, L., Zhang, J., Gao, J., Tang, X.: What happens behind the scene? Towards fraud community detection in e-commerce from online to offline. In: Companion Proceedings of the Web Conference 2021, pp. 105–113 (2021) 31. Liu, Y., Pang, B., Wang, X.: Opinion spam detection by incorporating multimodal embedded representation into a probabilistic review graph. Neurocomputing 366, 276–283 (2019) 32. Liu, Z., Chen, C., Yang, X., Zhou, J., Li, X., Song, L.: Heterogeneous graph neural networks for malicious account detection. In: Proceedings of the 27th ACM International Conference on Information and Knowledge Management, pp. 2077–2085 (2018) 33. Lucas, Y., Portier, P.-E., Laporte, L., He-Guelton, L., Caelen, O., Granitzer, M., Calabretto, S.: Towards automated feature engineering for credit card fraud detection using multi-perspective HMMs. Future Gener. Comput. Syst. 102, 393–402 (2020) 34. Molloy, I., Chari, S., Finkler, U., Wiggerman, M., Jonker, C., Habeck, T., Park, Y., Jordens, F., Schaik, R.V.: Graph analytics for real-time scoring of cross-channel transactional fraud. In:

248

35. 36.

37.

38. 39.

40.

41. 42.

43.

44. 45.

46.

47.

48. 49. 50.

51. 52.

53.

A. Narayan et al. International Conference on Financial Cryptography and Data Security, pp. 22–40. Springer (2016) Nandhini, G., Muthamilselvan, T., Alagiri, I., Thanapal, P.: Spam review detection on ecommerce sites. Int. J. Civ. Eng. Technol. (IJCIET) 9(7), 1167–1174 (2018) Portnoff, R.S., Huang, D.Y., Doerfler, P., Afroz, S., McCoy, D.: Backpage and bitcoin: Uncovering human traffickers. In: Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 1595–1604 (2017) Roy, A., Sun, J., Mahoney, R., Alonzi, L., Adams, S., Beling, P.: Deep learning detecting fraud in credit card transactions. In: 2018 Systems and Information Engineering Design Symposium (SIEDS), pp. 129–134. IEEE (2018) Santos, L.J.S., Ocampo, S.R.: Bayesian method with clustering algorithm for credit card transaction fraud detection. Roman. Stat. Rev. 1(1) (2018) Sarma, D., Alam, W., Saha, I., Alam, M.N., Alam, M.J., Hossain, S.: Bank fraud detection using community detection algorithm. In: 2020 Second International Conference on Inventive Research in Computing Applications (ICIRCA), pp. 642–646. IEEE (2020) Shchur, O., Bojchevski, A., Farghal, M., Günnemann, S., Saber, Y.: Anomaly detection in car-booking graphs. In: 2018 IEEE International Conference on Data Mining Workshops (ICDMW), pp. 604–607. IEEE (2018) Strehl, A., Ghosh, J.: Cluster ensembles—a knowledge reuse framework for combining multiple partitions. J. Mach. Learn. Res. 3, 583–617 (2002) Sun, J., Zhu, Q., Liu, Z., Liu, X., Lee, J., Su, Z., Shi, L., Huang, L., Xu, W.: Fraudvis: understanding unsupervised fraud detection algorithms. In: 2018 IEEE Pacific Visualization Symposium (PacificVis), pp. 170–174. IEEE (2018) Thejas, G.S., Boroojeni, K.G., Chandna, K., Bhatia, I., Iyengar, S.S., Sunitha, N.R.: Deep learning-based model to fight against ad click fraud. In: Proceedings of the 2019 ACM Southeast Conference, pp. 176–181 (2019) Thejas, G.S., Dheeshjith, S., Iyengar, S.S., Sunitha, N.R., Badrinath, P.: A hybrid and effective learning approach for click fraud detection. Mach. Learn. Appl. 3, 100016 (2021) Tian, T., Zhu, J., Xia, F., Zhuang, X., Zhang, T.: Crowd fraud detection in internet advertising. In: Proceedings of the 24th International Conference on World Wide Web, pp. 1100–1110 (2015) Tselykh, A., Knyazeva, M., Popkova, E., Durfee, A., Tselykh, A.: An attributed graph mining approach to detect transfer pricing fraud. In: Proceedings of the 9th International Conference on Security of Information and Networks, pp. 72–75 (2016) Varmedja, D., Karanovic, M., Sladojevic, S., Arsenovic, M., Anderla, A.: Credit card fraud detection-machine learning methods. In: 2019 18th International Symposium INFOTEHJAHORINA (INFOTEH), pp. 1–5. IEEE (2019) Wang, H., Li, Z., Huang, J., Hui, P., Liu, W., Hu, T., Chen, G.: Collaboration based multi-label propagation for fraud detection. In: IJCAI, pp. 2477–2483 (2020) Wang, J., Paschalidis, ICh.: Botnet detection based on anomaly and community detection. IEEE Trans. Control Netw. Syst. 4(2), 392–404 (2016) Weng, H., Li, Z., Ji, S., Chu, C., Lu, H., Du, T., He, Q.: Online e-commerce fraud: a largescale detection and analysis. In: 2018 IEEE 34th International Conference on Data Engineering (ICDE), pp. 1435–1440. IEEE (2018) Xu, C., Zhang, J.: Towards collusive fraud detection in online reviews. In: 2015 IEEE International Conference on Data Mining, pp. 1051–1056. IEEE (2015) Yu, C., Zuo, Y., Feng, B., An, L., Chen, B.: An individual-group-merchant relation model for identifying fake online reviews: an empirical study on a Chinese e-commerce platform. Inf. Technol. Manag. 20(3), 123–138 (2019) Zhang, W., Chaoqi, B., Yoshida, T., Zhang, S.: Cofea: a novel approach to spam review identification based on entropy and co-training. Entropy 18(12), 429 (2016)

Chapter 22

ReEDNet-An Encoder–Decoder Framework for Single Image Dehazing Konark Keshaw, Abhishek Pandey, Gopa Bhaumik, and M C Govil

Abstract Single image dehazing based on deep learning techniques has made significant progress. Most existing approaches retrieve haze-free images by estimating the transmission map and global atmospheric light, which are difficult to compute. In this paper, an encoder–decoder-based deep network ReEDNet is designed for single image dehazing. Further, a Cross Feature Aggregation (CrFeAt) block is introduced that utilizes skip connections to preserve the spatial feature to the last layer. ReEDNet is a light network that can be deployed in a resource-constrained environment. A quantitative analysis of the proposed network is performed on the RESIDE dataset. Extensive experiments show that the proposed ReEDNet achieves better performance compared to the state-of-the-art approaches.

22.1 Introduction Image dehazing focuses on retrieving clear images from hazy images. Haze is a phenomenon that occurs in images due to particles, fog, mist or even low lighting conditions. It can happen due to various weather conditions, including low light and rain. Both artificial and natural causes produce haze under photographic conditions. As a result, image dehazing becomes an ill-posed task. Image dehazing is a crucial challenge in the field of computer vision. It is essential to preserve the image’s visual aesthetics. Hazy images often affect the performance of the computer vision algorithms developed for object detection and tracking. It is inconvenient to train models with hazy images because it is difficult to capture images in various hazy weather conditions, particularly for relatively large datasets. Thus, image dehazing is applied in computer vision domains such as aerial imagery, classification, image/video retrieval, remote sensing and video analysis and recognition. Dehazing through a single image is a hard task. Various efforts have been made in the past in this field. The dehazing methods are divided into two categories: traditional and learning-based. Traditional cutting-edge methods such as Dark Channel Prior K. Keshaw · A. Pandey · G. Bhaumik (B) · M. C. Govil National Institute of Technology Sikkim, Sikkim, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_22

249

250

K. Keshaw et al.

(DCP) [1] and Color Attenuation Prior (CAP) [2] have yielded impressive outcomes. DCP suggests that haze-free images have a value near zero in at least one of the color channels over a small path, which increases the depth of the haze. This concept is used to compute scene transmission. Traditional methods are a good place to start, but most are limited to specific atmospheric conditions, such as atmospheric light and are ineffective in low light conditions. Moreover, traditional methods are also computationally complex. Recently, deep learning has achieved significant attention in various domain [3–5]. Learning methods based on deep neural networks, such as DeHazeNet [6], MsCNN [7] and AOD-Net [8], have recently made significant progress in the dehazing process. Most DNN-based methods have highly complicated network architecture, which necessitates high-end processing hardware, limiting their use in real-world applications. All methods have been limited to dehaze images captured under specific conditions. However, image degradation caused by haze is more general than daytime or night time conditions. The atmospheric scattering model, proposed by McCartney [9] in 1976, is commonly used in computer vision to describe the formation of a hazy image and can mathematically represented as: I (x) = R(x)τ (x) + i(x)(1 − τ (x))

(22.1)

where I is the observed haze image, R(x) is the haze-free image to be retrieved, while i(x) and τ (x) are two critical parameters. The first, i(x) denotes a global environmental illumination. The term τ (x) characterizes the scene transmittance. It denotes the proportion of light emitted by the object under observation that reaches the camera. The first part of Eq. (22.1) is known as the direct transmission and the second part is called the airlight. A global atmospheric light and medium transmission map are calculated based on the above model, and R can be retrieved from I . Calculating these parameters is a significant part of the algorithms. Recent deep neural network methods based on the atmospheric model have attempted to compute τ (x) and i(x). However, because calculating τ (x) and R(x) independently is difficult, it has resulted in an expensive task. The environmental illumination in these models can differ from pixel to pixel. For example, when τ (x) is close to 1, the effect of R(x) in the hazy image becomes negligible. In this paper, we propose a learning-based network, namely ReEDNet, an convolutional encoder–decoder dehazer network with a different approach that is solely focused on producing a haze-free image. The proposed network produces a clean image rather than calculating any other parameters and then using those parameters to produce the dehazed image. Nevertheless, it has been shown in literature that CNNbased models can produce effective results without the use of an atmospheric model. The proposed network reduces computational complexity by reducing the number of layers in the architecture. ReEDNet localizes the hazy region in the image by varying the receptive fields using the encoder–decoder scheme. Further, an intermediate block is introduced, which allows the shallow features from the downsampling part to flow to a higher level in the upsampling part, which is essential for feature preservation.

22 ReEDNet-An Encoder–Decoder Framework for Single Image Dehazing

251

The main contributions of this paper are summarized as follows: • We develop an end-to-end dehazing network ReEDNet that uses the encoder– decoder framework to yield the best trade-off between performance and parameters for image dehazing. ReEDNet outperforms the existing approaches of dehazing in terms of PSNR and SSIM, thereby reducing the computation cost. • We introduced an intermediate block, namely Cross Feature Aggregation (CrFeAt) block, consisting of residual connections for feature preservation. The CrFeAt block retains the spatial information lost due to repetitive downsampling in the encoder–decoder up to the top layer. • The performance of the proposed ReEDNet is evaluated on benchmark RESIDE datasets containing indoor, outdoor and I/O hazy images, and the experimental results show the effectiveness of the proposed network. The rest of the paper is organized as follows: In Sect. 22.2, an overview of the proposed ReEDNet is presented. Section 22.3 discusses the comparative analysis of network. Section 22.4 gives a detailed analysis of the experimental results, the dataset used, Quantitative and Qualitative analysis. Finally, Sect. 22.5 presents the conclusion of the paper.

22.2 The Proposed Network This section describes the architecture of the proposed haze relevant feature attention network for image dehazing in detail. First, an overview of the architecture is provided. The encoder block, Cross Feature Aggregation(CrFeAt) block, and decoder block are then discussed in detail.

Fig. 22.1 Detailed workflow of the proposed network architecture for single image dehazing

252

K. Keshaw et al.

22.2.1 The Proposed ReEDNet We proposed a deep end-to-end network ReEDNet based on the encoder–decoder framework trained to take on a hazy image and restore its haze-free version. ReEDNet mainly consists of three blocks: encoder block, Cross Feature Aggregation (CrFeAt) block and decoder block. The encoder block takes hazy images as input and encodes the features. The output from the encoder block is then fed to three consecutive Cross Feature Aggregation (CrFeAt) blocks. The output is then further fed to the decoder block to retrieve the haze-free image. The overview of the network architecture is shown in Fig. 22.1. The network architecture and its three building blocks are explored in more detail in the following subsections. The Encoder Block: ReEDNet takes a hazy image as input of shape (512 × 512 × 3) consisting of an RGB channel. The first and second layers consist of 64-channel convolutional blocks with a 3 × 3 filter size. A two-step downsampling layer follows, encoding the input image into a (256 × 256 × 64) volume. A residual is taken from this pooling layer, which is fed into the decoder block. This whole block of two convolution layers and a pooling layer is repeated. This results in an encoded volume of (128 × 128 × 64). The resulting (128 × 128 × 64) volume is then fed to the CrFeAt block. (3,3,64 (x))) (22.2) ψ(x) = λ(2,2) [(3,3,64 1 1 (3,3,64 (ψ(x)))) ϒe (x) = λ(2,2) [(3,3,64 1 1

(22.3)

Cross Feature Aggregation (CrFeAt) Block: In literature, various encoder–decoder methods are introduced in order to improve the performance of CNN-based models, but repeated downsampling in the network degrades spatial details. To preserve spatial details, skip connections are used between encoder and decoder blocks, and even in the CrFeAt block of the proposed ReEDNet to enhance performance. CrFeAt is comprised of five convolution layers, each with a depth channel of 64 and a kernel size of 3 × 3. The output of the sequential convolution layers is aggregated with the previous layer and activated with ReLU functions. Experimentation has revealed that three consecutive CrFeAt blocks improve the efficiency of the proposed ReEDNet. Figure 22.1 also depicts an overview of the CrFeAt block. (3,3,64 (3,3,64 (3,3,64 (3,3,64 (x)))))) ϒr (x) = η(x + 3,3,64 1 1 1 1 1 ϒres (x) = ϒr (ϒr (ϒr (x)))

(22.4)

(22.5)

The Decoder Block: The structure of the decoder block is similar to that of the encoder block. To begin, it is comprised of two (128 × 128 × 64) volumes of convolutional layers. The output is then upsampled by a factor of two, yielding (256

22 ReEDNet-An Encoder–Decoder Framework for Single Image Dehazing

253

× 256 × 64) volume. This block of two convolutional layers and one upsampling layer is repeated, yielding a volume of (512 × 512 × 64). The encoder block generates a residual connection with the output features. The final upsampled (512 × 512 × 3) volumes are combined with the inputs and fed into ReLU [10] to generate the final dehaze image. The residual from the input and encoder phase aids in the capture of overall boundary features and limits spatial degradation. Consider an input hazy RGB image Hx , the outcome Dout of ReEDNet can be represented mathematically f, f,d represents convolution operation with kernel using Eqs. (22.2–22.11) where s (k,k) denotes kernel size k × k, η denotes the ReLU size f × f and depth channel d, λ activation function. (22.6) β(y) = y + ϒres (y) (3,3,64 (β(y)))) θ(y) = λ(2,2) (3,3,64 1 1

(22.7)

γ(y) = θ(y) + ψ(y)

(22.8)

ϒd (y) = λ(2,2) (3,3,64 (3,3,64 (γ(y)))) 1 1

(22.9)

α(y) = 3,3,64 (ϒd (y)) 1

(22.10)

Dout = Hx + α(Hx )

(22.11)

22.3 Analysis of Network Existing deep networks, such as DeHazeNet [6] and MsCNN [7] produce effective results but are inefficient enough to be used in real-time dehazing due to their high computational complexity. Several existing methods use transmission maps and atmospheric light estimates to design a network for image dehazing. This approach adversely influences dehazing performance because calculating transmission maps and atmospheric light is computationally expensive and produces inaccurate estimates. Therefore, we present an end-to-end network based on encoder–decoders that does not require any additional parameter estimations. We used skip connections between the encoder and decoder and from the input to preserve the spatial details. ReEDNet resolves the shortcomings of the existing methods and reduces the complexity of dehazing process.

254

K. Keshaw et al.

22.4 Experiments and Results and Analysis The proposed ReEDNet is evaluated on RESIDE benchmark dataset. The following subsections discuss in detail the implementation details, dataset used, evaluation metrics, quantitative and qualitative analysis of the proposed ReEDNet.

22.4.1 Implementation Details ReEDNet is trained with Nvidia GeForce RTX 2080 GPU with Xeon processor, 16core CPU, and 11 GB RAM under Cuda 10.0. on Tensorflow-GPU 2.0. For calculating loss, we have used a combination of MSE loss and perceptual loss [11]. A lot of CNN methods have used only MSE loss but it has been seen that only MSE loss produces blurry outputs. Recent research has shown that using multiple loss functions in a DNN-based dehaze network will help improve the quality of the result. We chose a compound loss because of these considerations. For accelerated training, the Adam optimizer [12] is used with a standard learning rate of 1 ×10−4 . This optimizes the training of ReEDNet. We train ReEDNet for 20 epochs. The batch size is set to 16 to balance the training speed and the memory consumption on the GPU.

22.4.2 Dataset The proposed ReEDNet is evaluated on synthetic and real-world data. RESIDE [13, 14] is a popular synthetic dataset that is divided into five subsets: Indoor Training Set (ITS), Outdoor Training Set (OTS), Synthetic Objective Testing Set (SOTS), RealWorld Task Driven Testing Set (RTTS) and Hybrid Subjective Testing Set (HSTS). RESIDE-ITS contains 13,990 synthetic images created from 1399 clean images from the depth datasets NYU-Depth V2 and Middlebury stereo. In addition, the proposed network is tested on the NTIRE-2018 I-O hazy dataset. The NTIRE-2018 I–O hazy dataset is divided into two parts: I-Haze [15] and O–Haze [16]. The sample images from the dataset are shown in Fig. 22.2.

22.4.3 Evaluation Metric The Peak Signal-to-Noise Ratio (PSNR) and the Structural Similarity Index (SSIM) are used to validate the performance of the proposed ReEDNet as evaluation metrics, which are commonly used as criteria to evaluate image quality in the image dehazing task. PSNR is the ratio of a signal’s maximum possible power to the power of corrupting noise affecting the fidelity of its representation. A high PSNR value indicates low noise, and vice versa.

22 ReEDNet-An Encoder–Decoder Framework for Single Image Dehazing

255

Fig. 22.2 Sample images from the dataset Table 22.1 Comparative results of the proposed ReEDNet and existing dehazing methods over indoor hazy images from SOTS dataset DCP [1] CAP [2] DehazeNet [6] MsCNN [7] AOD-Net [8] ReEDNet PSNR SSIM

16.62 0.81

19.05 0.83

21.14 0.85

17.57 0.81

 PSNR = 20log10

MAX f √ (MSE)

19.06 0.85

20.67 0.79

 (22.12)

The Structural Similarity Index (SSIM) determines the similarity of two images. The premise of structural information is that pixels have a lot of interdependencies, especially when they are spatially close. The structural information about the objects in the visual scene is preserved in these dependencies.

256

K. Keshaw et al.

Fig. 22.3 Visual representation of the comparative analysis of proposed ReEDNet with existing state-of-the-art approaches Table 22.2 Comparative analysis of the proposed ReEDNet and existing dehazing methods over outdoor hazy images from SOTS dataset DCP [1] CAP [2] DehazeNet [6] MsCNN [7] AOD-Net [8] ReEDNet PSNR SSIM

18.54 0.71

23.04 0.81

26.84 0.82

21.73 0.83

24.08 0.87

19.77 0.79

Table 22.3 Comparative analysis of the proposed ReEDNet and existing dehazing methods over hazy images from I–O Hazy dataset DCP [1] CAP [2] DehazeNet [6] MsCNN [7] AOD-Net [8] ReEDNet PSNR SSIM

15.60 0.68

14.16 0.59

14.04 0.63

16.39 0.70

16.63 0.70

14.42 0.65

22.4.4 Quantitative Analysis Table 22.1 gives the SSIM and PSNR scores of proposed ReEDNet and other existing methods for the SOTS indoor test set, whereas Tables 22.2 and 22.3 present a comparative analysis for SOTS outdoor I/O hazy datasets. The comparative analysis shows that the proposed ReEDNet outperforms the existing approaches, namely DCP and MsCNN for SOTS indoor dataset. Similarly, for SOTS outdoor dataset, ReEDNet outperforms DCP and MsCNN. In comparison with the complexity and number of parameters of state-of-the-art methods, ReEDNet performs significantly better and is substantially lighter and faster.

22.4.5 Qualitative Analysis The visual representation of the efficacy of proposed ReEDNet is illustrated in Fig. 22.3, which depicts the qualitative results of three state-of-the-art deep learning

22 ReEDNet-An Encoder–Decoder Framework for Single Image Dehazing

(a) PSNR

257

(b) SSIM

Fig. 22.4 Graphical representation of the comparative analysis of proposed ReEDNet with existing approaches

algorithms and two prior-based algorithms on the RESIDE dataset. The comparative analysis of the existing networks and the proposed ReEDNet in terms of PSNR and SSIM is presented graphically in Fig. 22.4 .

22.5 Conclusion This work presents an end-to-end encoder–decoder-based deep network ReEDNet for single image dehazing. Further, a residual connection-based block (CrFeAt) is designed for effective spatial feature learning.The experimental results show that the CrFeAt block improves system performance when compared to other state-of-the-art methods. Furthermore, a combination of MSE and perceptual loss is used as the loss function to accelerate training process. Experimental analysis on standard RESIDE dataset demonstrates that the proposed ReEDNet achieves better performance in both quantitative and qualitative evaluations. Our research demonstrates that a simple CNN-based design can achieve better performance without estimating the complex parameter, making it computationally efficient.

References 1. He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. In: IEEE Conference on Computer Vision and Pattern Recognition (2009) 2. Zhu, Q., Mai, J., Shao, L.: A fast single image haze removal algorithm using color attenuation prior. IEEE Trans. Image Process. 24(11), 3522–3533 (2015) 3. Bhaumik, G., Verma, M., Govil, MC., Vipparthi, SK.: ExtriDeNet: an intensive feature extrication deep network for hand gesture recognition. Visual Comput 1–14 (2021) 4. Bhaumik, G., Verma, M., Govil, MC., Vipparthi, SK.: HyFiNet: hybrid feature attention network for hand gesture recognition. Multimedia Tools Appl. 1–20 (2022)

258

K. Keshaw et al.

5. Bhaumik, G., Verma, M., Govil, MC., Vipparthi, SK.: CrossFeat: multi-scale cross feature aggregation network for hand gesture recognition. In: 2020 IEEE 15th International Conference on Industrial and Information Systems (ICIIS), pp. 274–279. IEEE (2020) 6. Cai, B., Xu, X., Jia, K., Qing, C., Tao, D.: DehazeNet: an end-to-end system for single image haze removal. IEEE Trans. Image Process. 25(11), 5187–5198 (2016) 7. Ren, W., Liu, S., Zhang, H., Pan, J., Cao, X., Yang, M.-H.: Single image dehazing via multiscale convolutional neural networks. In: Computer Vision ECCV. Lecture Notes in Computer Science, vol. 9906, pp. 154–169. Springer (2016) 8. Li, B., Peng, X., Wang, Z., Xu, J., Feng, D.: AODNet: all-in-one dehazing network. In: IEEE International Conference on Computer Vision, pp. 4780–4788 (2017) 9. McCartney, E.J.: Optics of the atmosphere: scattering by molecules and particles. p. 421. John Wiley and Sons, Inc., New York (1976) 10. Maas, A.L., Hannun, A.Y., Ng, A.Y.: Rectifier nonlinearities improve neural network acoustic models. In: Proc. icml, vol. 30, p. 3 (2013) 11. Johnson, J., Alahi, A., Fei-Fei, L.: Perceptual losses for real-time style transfer and superresolution. In: Computer Vision-ECCV. Lecture Notes in Computer Science, vol. 9906, pp. 694–711. Springer (2016) 12. Kingma, D.P., Ba, J.: Adam: a method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014) 13. Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE Trans. Image Process. 28(1), 492–505 (2018) 14. Li, B., et al.: RESIDE: a benchmark for single image dehazing. arXiv preprint arXiv:1712.04143 (2017) 15. Ancuti, C.O., Ancuti, C., Timofte, R., De Vleeschouwer, C.: I-HAZE: a dehazing benchmark with real hazy and haze-free indoor images. ArXiv e-prints (2018) 16. Ancuti, C.O., Ancuti, C., Timofte, R., De Vleeschouwer, C.: O-HAZE: a dehazing benchmark with real hazy and haze-free outdoor images. ArXiv e-prints (2018)

Chapter 23

Detection of Flood Events from Satellite Images Using Deep Learning Anushree Rambhad, Dhirendra Pratap Singh, and Jaytrilok Choudhary

Abstract Floods are one of the deadliest and most devastating natural disasters, killing thousands of people and destroying homes, infrastructure, and agricultural areas. To address this, a flood management technique that can detect and respond to floods in real time must be developed and implemented. Flooded areas to begin relief efforts as soon as possible. The current imaging methods, which rely on traditional methods and satellites, have shown limited accuracy and slow response times, thus rendering them unreliable. This article presents a semantic segmentation technique for autonomously detecting floods. Dataset is taken ETCI flood detection. A transfer learning Unet model is employed in the suggested strategy. Our suggested approach is evaluated using several metrics precision, recall, and mean Intersection over Union (MIoU). The mobilenetv2 + Unet technique significantly outperforms vanilla Unet with a MIoU segmentation rate of 76.62%, a precision of 93.49%, and a recall rate of 79.44%.

23.1 Introduction Every year, natural disasters kill 60,000 individuals, accounting for 0.1% of all deaths globally [1]. The most prevalent natural calamity on the earth is flooding. If floods are not recognized quickly and there is a lack of precise and fast technology that can automatically detect the start of flooding in a given location, then lives will be lost. This emphasizes the importance of utilizing advanced digital technologies to swiftly and correctly identify flood-affected areas so that rescue efforts can begin immediately [2]. Flood identification using satellite images is critical for understanding present water cycle changes as a result of heavy rainfall. In the last decade, deep learning has evolved in the field of image processing. Deep convolutional nets have made great progress in image and frame processing. Miou et al. [1] discussed many semantic segmentation models that have been developed in recent years. The technique of assigning a label to each pixel in an image so that pixels with the same label have A. Rambhad (B) · D. P. Singh · J. Choudhary Maulana Azad National Institute of Technology, Bhopal, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_23

259

260

A. Rambhad et al.

comparable qualities is known as semantic segmentation [1]. In this article, semantic segmentation model with the transfer learning has been compared with vanilla Unet architecture. The remainder of the paper is formatted as follows. The second section discusses related work. The architectural methodology is discussed in Sect. 23.3. The methodology part is divided into five subsections. Section 23.4 presents the model comparison and analysis. Finally, Sect. 23.5 concludes the paper’s findings.

23.2 Related Work In the literature, various methods for identifying floods in satellite images have been developed. Gedik et al. [4] developed thresholds for one or more spectral bands, and threshold methods are widely utilized. These algorithms are simple to use; however, they are prone to noise, which leads to inaccurate classifications. Ciecholewski et al. [5] used an image-processing approach to investigate river channel segmentation using Watershed segmentation. The basic disadvantage of employing the watershed segmentation technique is that river boundaries are uneven and jagged. Baydargil et al. [6] proposed detecting roadside water flow. The images were captured by a roadside security camera. The author used the SegNet architecture. Collecting CCTV images is difficult. CCTV cameras frequently suffer from infrared glare. SVM, MLP, deep CNN (Convolutional Neural Network) models, and domain adaption methods in combination with these models were compared by Islam et al. [7]. The authors claim that combining a deep CNN with a semi-supervised domain adaption technique beats previous models. Chakraborty et al. [8] developed automated seed and range selection method clustering methodologies for Flood Extent Extraction. According to the authors, the proposed clustering methodology is more accurate than the mean shift and LPQ strategies for collecting flood-affected zones. Pai et al. [9] mentioned the usage of deep learning algorithms like Unet to separate rivers and land efficiently. Transfer Unet is used to compare the findings. Experiments demonstrate that transfer Unet architecture outperforms vanilla Unet design on SAR images. Additional CNN models, however, were not studied in this study. Vineethet al. [10] created two parallel modules flood detection and flood depth detection to minimize processing time. The flood detection module uses MobileNetV2 as the backbone network. Mask RCNN (region with Convolutional Neural Network) and VGG16 (Visual Geometry Group) are used to detect the depth of flood. Basnyat et al. [11] created a flood detection and notification prototype in Ellicott City, Maryland The authors used semantic segmentation and multimodal data fusions image segmentation to detect floods. Furthermore, this design was only tested on a small set of labeled data, and other image segmentation models were not evaluated. Sadleret al. [12] proposed a method for estimating flood severity using environmental variables such as water level, rainfall, wind, and other factors. Poisson regression and Random forest models were employed by the authors. The system created by Khalaf et al. [13] automates flood data processing by leveraging current breakthroughs in the Internet of Things (IoT) and machine learning. According to the authors, ensemble learning using the Long-Short Term

23 Detection of Flood Events from Satellite Images …

261

Memory model and random forest outperformed individual models. Examining some of the literature’s early attempts, it is evident that resolving the accuracy-complexity trade-off is one of the most difficult challenges. Recent research has found that image segmentation using a machine learning algorithm can produce better results than traditional statistical methods. This article compared multiple CNN-Unet architectures by altering Unet’s training parameters and determining which design gives the best results in the shortest amount of time.

23.3 Methodology Workflow has five major steps such as image acquisition, pre-processing of the images, creating semantic segmentation model, training model, and testing model. The initial step is to gather data from approximately 33,000 training and 10,000 testing data from [14]. The images are pre-processed in RGB format. A semantic segmentation model is then trained using Unet transfer learning. Finally, MIoU, recall, and precision are used to evaluate the model’s correctness. The next section of this article will go through each section in depth.

23.3.1 Image Acquisition This section outlines how data is gathered for an experiment. The dataset is made up of synthetic aperture radar (SAR) images taken over variety of locations around the world, including Nebraska, Alabama, Bangladesh, Red River North, and Florence, Italy. SAR satellites have an active sensor, and they can image at any time of day or night. There are 33K images in the training dataset and 10K images in the testing dataset. The VV and VH postfixes indicate the polarization condition of broadcast and received radar signals. Vertically polarized transmitted radar and vertically polarized received radar are abbreviated as VV. Vertically polarized transmitted radar and horizontally polarized received radar are abbreviated as VH. The images have already been pre-processed and scaled to integer values between 0 and 255. Fig. 23.1 presents VV and VH images of the Bangladesh region. Because it is often difficult for the untrained eye to recognize water in these grayscale images, it is usual to combine them into a more instructive color space. We combine the two images and then use a combination of the first two to build a new grayscale image. As a result, in this type of image, the water in the final three-channel RGB images is blue, making it easy to see. This is how a new image is created: Ratio = 1 −

VVImage VHImage

Figure 23.2 presents RGB image of Nebraska region.

(23.1)

262

A. Rambhad et al.

Fig. 23.1 VV and VH sentinel-2 images from Nebraska region

Fig. 23.2 A RGB image of Nebraska region

A flood mask is included in the dataset for each image in the training set, as well as a water body mask for all training and validation images. “Not a flood region” and “flood region,” respectively, are represented by the binary integers 0 and 1. The RGB image, water mask, and flood mask of Nebraska are shown in Fig. 23.3.

23.3.2 Preprocessing of the Images To increase the quality of the input images and prepare them for further processing, image pre-processing is required. The method includes downloading raw files, storing them in a database, decreasing noise, and executing rectification. As a result,

23 Detection of Flood Events from Satellite Images …

263

Fig. 23.3 The RGB image that was processed, as well as the flood mask and water mask images that were provided for Nebraska region

the current work used pre-processing to eliminate noise from the obtained images in the following data collection. This article proposes the loading of images into memory, normalizes them from 0–255 to 0–1, and then applies transformation operations. As a result, there is less overfitting and overall higher validation performance. Because the same transformation must be applied to both the input image and the mask, applying transformations to segmentation tasks is slightly more difficult than applying transformations to classification tasks.

23.3.3 Creating Semantic Segmentation Model Unet is a deep neural network-based pixel-wise semantic segmentation framework [15]. A Unet is an image segmentation convolutional neural network model. Unet has proven to be particularly useful when the output is the same size as the input and the output demands the same level of spatial precision. The semantic segmentation network is developed by fine-tuning the pixel classification layer. The photos are shuffled every epoch, and the minimum batch size is set to 64. The image is captured and down-sampled using a sequence of stride two convolutions, with the grid size reducing each time, when convolutional neural networks are used to classify images. An upsampling mechanism must be in place to increase the grid size so that the output image is the same size as or larger than the input. The left side of the U in a Unet is formed by the downsampling/encoder path, while the right side is formed by the up-decoder path. This is performed in the upsampling/decoder pipeline by a sequence of transposed convolutions, each of which adds pixels between and around the existing pixels. The approach essentially reverses the downsampling process. A pre-trained model has been trained to tackle high computational cost of training models like ResNet34 (Residual Neural Network) [16], MobileNetV2 [17], VGG19 [18], and DPN (Deep path network) [19]. A pre-trained model saves time and money by avoiding the need to reinvent from scratch. By using pre-trained models that have been previously trained on large datasets, researchers can immediately take the

264

A. Rambhad et al.

Fig. 23.4 ResNet + Unet architecture [20]

weights and architecture obtained and apply the learning to a new problem statement. “Transfer learning” is the term for this. Transfer learning is a machine learning technique in which a model trained on one task is reused on a related task as an optimization for efficiently predicting the second task. Classic deep learning models such as VGGNet19, ResNet34, mobilnetv2, and DPN68 model weights acquired from the ImageNet dataset are widely fine-tuned in flood detection applications to reduce modeling and training time and increase detection performance. As a result, Unet and transfer learning models have been applied in this paper to improve detection accuracy and efficiency (Fig. 23.4).

23.3.4 Training Model For the encoder/downsampling phase of the Unet, a pre-trained model can be employed (the left half of the U). VGG19, ResNet34, DPN68, and MobileNetV2 architectures were employed in this work because they are quite effective, as well as faster to train, and require less memory. Decoder The Unet learner will dynamically create the decoder half of the Unet model when given an encoder design. When a pretrained encoder is used, the time it takes to train an image prediction model utilizing Unet is greatly decreased. This architecture now recognizes the kind of characteris-

23 Detection of Flood Events from Satellite Images …

265

tics that must be identified and enhanced. The main goal of semantic segmentation’s loss function is to distinguish between flood and non-flooded pixels. As a result, the transfer learning Unet network employs the binary cross-entropy loss function.

23.3.5 Testing Model This section shows the testing parameters that were used in the experiments. In this article 75% of the image, a dataset was used to train the semantic segmentation network, while 25% was used to test it. It was put to the test with test images that were not used during the training period. During the testing, 10,000 images were used, each with a resolution of 256 * 256 pixels. More testing images are given to confirm the model’s ability to generalize. Examples of semantic segmentation findings for a range of test images are included in this article. The result section also includes images of the ground truth and flood detected by various CNN-Unet models. A variety of metrics, including MIoU, recall, and precision, are used to evaluate the suggested approach’s performance.

23.4 Results and Analysis The performance of the semantic segmentation model stage is critical for the enhancement of a flood detection system. As a result, multiple CNN-based semantic segmentation models were tested for segmenting flood region images. This experiment is tested on Google colab pro. Google colab pro has K80, T4, and P100 GPU support, and also, it provides 32 GB RAM. Table 23.1 compares the precision, recall, and MIoU of vanilla Unet and transfer learning Unet. Transfer learning with Unet architecture has increased the performance of vanilla Unet architecture for a small geographically localized dataset. Five CNN Unet variants were tested for this task: Resnet34-Unet, MobileNetV2-Unet, VGG-Unet, DPN68-Unet, and Unet with vanilla encoder. One of the study’s primary contributions is the investigation of various backbones. For each dataset’s flooded image, segmentation of the flood region images is performed using CNN models and the conventional model. The goal was to test the generalization ability of well-known pre-trained CNN models for the automatic segmentation of flood zone images developed by the computer vision community. As shown in Table 23.1, the suggested MobileNetV2 + Unet model performed better in terms of accuracy 0.9249, recall of −0.7944, and MIoU-0.7463 in flood detection than the vanilla Unet model precision of −0.9040, recall of 0.75920, and MIoU-0.6960. The DPN68 + Unet model (precision 0.9238, recall −0.7929, and MIoU-0.7440) and the ResNet + Unet model (precision 0.9275, recall −0.7985 and MIoU-0.7516) both had a significantly higher detection rate. We discovered that the proposed transfer learning Unet model has a strong ability to accurately

266

A. Rambhad et al.

Table 23.1 The result shows the precision, recall, and MIoU obtained by applying vanilla Unet and proposed transfer learning Unet models Model name Precision Recall MIoU Unet VGG19 + Unet ResNet34 + Unet DPN68 + Unet MobileNetV2 + Unet

0.9040 0.9113 0.9275 0.9238 0.9249

0.75920 0.77119 0.7985 0.7927 0.7944

0.6960 0.71740 0.7516 0.7440 0.7463

Fig. 23.5 Training accuracies of different graph. a Training graph of vanilla Unet. b Training graph of VGG19 + Unet. c Training graph of ResNet34 + Unet. d Training graph DPN68 + Unet. e Training graph of MobileNetV2 + Unet

Fig. 23.6 Example of testing image a RGB image. b Flood detection by vanilla Unet. c Flood detection by VGG19 + Unet. d Flood detection by ResNet34 + Unet. e Flood detection by DPN68 + Unet. f Flood detection by MobileNetV2 + Unet

identify floods in a given experiment. Furthermore, due to the usage of depth-wise separable convolutions, MobileNetV2-Unet’s training time is less than Resnet34Unit’s. The results reveal that the CNN-based model outperforms vanilla Unet models in datasets. According to experimental data, MobileNetV2-Unet could be a candidate for semantic segmentation of anatomic regions in a fully automated end-to-end flood detection system. Figure 23.5 depicts the training recall, accuracy, and MIoU of five different models. Figures 23.6 and 23.7 depict flood detection using multiple models. VGG + Unet was unable to detect flood in the testing image; however, Renet34 + Unet and MobileNetV2 + Unet correctly detected flood.

23 Detection of Flood Events from Satellite Images …

267

Fig. 23.7 Example of testing image a RGB image. b Flood detection by vanilla Unet. c Flood detection by VGG19 + Unet. d Flood detection by ResNet34 + Unet. e Flood detection by DPN68 + Unet. f Flood detection by MobileNetV2 + Unet

23.5 Conclusion A flood detection method based on transfer learning Unet is proposed in this study. The vanilla Unet model is compared to five deep learning models + Unet. Dataset consists of four public flood detection datasets: Alabama, Bangladesh, Nebraska, and Florence. The suggested model achieves a large performance improvement by utilizing transfer learning based on multi-fold knowledge and a mixed weighted loss function. The mobilenetv2 + Unet technique achieves 76.62% segmentation MIoU, 93.49% precision, and 79.44% recall. Thus, this experiment offers a flood detection solution for practical applications, ranging from raw gathered images to segmented flood-detected images. It is possible to detect floods in complex scenarios while also improving processing speed.

References 1. Miao, Z., Kun, F., Sun, H., Sun, X., Yan, M.: Automatic water-body segmentation from highresolution satellite images via deep networks. IEEE Geosci. Remote Sens. Lett. 15(4), 602–606 (2018) 2. Talal, M., Panthakkan, A., Mukhtar, H., Mansoor, W., Almansoori, S., Al Ahmad, H.: Detection of water-bodies using semantic segmentation. In: 2018 International Conference on Signal Processing and Information Security (ICSPIS), pp. 1–4. IEEE (2018) 3. AHassan, A., Mahmood, A.: Efficient deep learning model for text classification based on recurrent and convolutional layers. In: 2017 16th IEEE International Conference on Machine Learning and Applications (ICMLA), pp. 1108–1113. IEEE (2017). McFeeters, S.K.: Using the normalized difference water index (NDWI) within a geographic information system to detect swimming pools for mosquito abatement: a practical approach. Remote Sens. 5(7), 3544–3561 (2013) 4. Senaras, C., Gedik, E., Yardimci, Y.: A novel dynamic thresholding and categorizing approach to extract water objects from VHR satellite images. In: 2014 IEEE Geoscience and Remote Sensing Symposium, pp. 4934–4937. IEEE (2014) 5. Ciecholewski, M.: River channel segmentation in polarimetric SAR images: watershed transform combined with average contrast maximisation. Expert Syst. Appl. 82, 196–215 (2017)

268

A. Rambhad et al.

6. Baydargil, H.B., Park, J., Shin, H.-S., Park, K.: Water flow detection using deep convolutional encoder-decoder architecture. In: 2018 18th International Conference on Control, Automation and Systems (ICCAS), pp. 841–843. IEEE (2018) 7. Islam, K.A., Uddin, M.S., Kwan, C., Li, J.: Flood detection using multi-modal and multitemporal images: a comparative study. Remote Sens. 12(15), 2455 (2020) 8. Chakraborty, A., Chakraborty, D.: Computerized seed and range selection method for flood extent extraction in SAR image using iterative region growing. J. Indian Soc. Remote Sens. 47(4), 563–571 (2019) 9. Pai, M.M.M., Mehrotra, V., Aiyar, S., Verma, U., Pai, R.M.: Automatic segmentation of river and land in SAR images: a deep learning approach. In: 2019 IEEE Second International Conference on Artificial Intelligence and Knowledge Engineering (AIKE), pp. 15–20. IEEE (2019) 10. Vineeth, V., Neeba, E.A.: Flood detection using deep learning. In: 2021 International Conference on Advances in Computing and Communications (ICACC), pp. 1–5. IEEE (2021) 11. Basnyat, B., Roy, N., Gangopadhyay, A.: Flood detection using semantic segmentation and multimodal data fusion. In: 2021 IEEE International Conference on Pervasive Computing and Communications Workshops and other Affiliated Events (PerCom Workshops), pp. 135–140. IEEE (2021) 12. Sadler, J.M., Goodall, J.L., Morsy, M.M., Spencer, K.: Modeling urban coastal flood severity from crowd-sourced flood reports using Poisson regression and Random Forest. J. Hydrol. 559, 43–55 (2018) 13. Khalaf, M., Alaskar, H., Hussain, A.J., Baker, T., Maamar, Z., Buyya, R., Liatsis, P., Khan, W., Tawfik, H., Al-Jumeily, D.: IoT-enabled flood severity prediction via ensemble machine learning models. IEEE Access 8, 70375–70386 (2020) 14. https://nasa-impact.github.io/etci2021/ 15. Ronneberger, O., Fischer, P., Brox, T.: U-Net: convolutional networks for biomedical image segmentation. In: Navab, N., Hornegger, J., Wells, W.M., Frangi, A.F. (eds.) MICCAI 2015. LNCS, vol. 9351, pp. 234–241. Springer, Cham (2015) 16. He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770–778 (2016) 17. Andrew, G., Zhu M.: Efficient convolutional neural networks for mobile vision applications (2017) 18. Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556 (2014) 19. Chen, Y., Li, J., Xiao, H., Jin, X., Yan, S., Feng, J.: Dual path networks. Adv. Neural Inf. Process. Syst. 30 (2017) 20. https://forums.fast.ai/t/u-net-with-resnet34-backbone/85744/5 21. Naik, D., Jaidhar, C.D.: Image segmentation using encoder-decoder architecture and region consistency activation. In: 2016 11th International Conference on Industrial and Information Systems (ICIIS), pp. 724–729. IEEE (2016) 22. Du, Z., Yang, J., Huang, W., Ou, C.: Training SegNet for cropland classification of high resolution remote sensing images. In: AGILE Conference (2018) 23. Badrinarayanan, V., Kendall, A., Cipolla, R.: Segnet: a deep convolutional encoder-decoder architecture for image segmentation. IEEE Trans. Pattern Anal. Mach. Intell. 39(12), 2481– 2495 (2017) 24. AlMaazmi, A.: Water bodies extraction from high resolution satellite images using water indices and optimal threshold. In: Image and Signal Processing for Remote Sensing XXII, vol. 10004, p. 100041J. International Society for Optics and Photonics (2016)

Chapter 24

Development and Implementation of an Efficient Deep Residual Network for ECG Classification Rishabh Arya, Ujjawal Agrawal, Ananya Singh, Eshaan Gupta, and Priya Ranjan Muduli Abstract In this paper, we propose an efficient deep residual network-based method to classify electrocardiogram signals using Web/edge devices with improved accuracy and CPU latency. The method classifies different irregular heartbeats according to the Association for the Advancement of Medical Instrumentation standard. We designed, trained, and validated an improved deep residual network using the MITBIH dataset. The dataset comprises five distinct classes: normal beat, supraventricular premature beat, premature ventricular contraction, the fusion of ventricular and normal beat, and unclassifiable beat. After training and validating our model, we incorporate quantization techniques using the TensorFlow library to reduce the size of the model. Due to a great reduction in model size using the techniques of quantization, we develop a Web application for online ECG monitoring. The model size is optimized to reduce it by more than 90% of the original size. The results obtained after the quantization technique in our designed convolutional neural network demonstrate an acceptable classification performance. The proposed method seems suitable for deployment in smartphones and Web applications.

R. Arya · U. Agrawal · A. Singh · E. Gupta · P. R. Muduli (B) Department of Electronics Engineering, Indian Institute of Technology (BHU), Varanasi, Uttar Pradesh 221005, India e-mail: [email protected] R. Arya e-mail: [email protected] U. Agrawal e-mail: [email protected] A. Singh e-mail: [email protected] E. Gupta e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_24

269

270

R. Arya et al.

24.1 Introduction In recent years, the human population affected by cardiovascular disease has increased tremendously. There are several issues with manually analyzing electrocardiograms (ECG) signals, such as there may be similarities to other data, difficulty in identifying and differentiating different types of waveforms, and signal patterning. For human beings, this task is time-taking, and the probability of errors is high. To solve this problem of ECG signal analysis, we can use state-of-the-art machine learning techniques to accurately detect these vital signals. There is a lot of research done in the ECG analysis field, for example, Wang et al. [1] propose a method to identify ECG signals for short-term signals. This paper has used a principal component analysis network. Lu et al. used resource extraction and balancing through the random sampler algorithm to classify ECG signals [2]. Real-time ECG signal processing is implemented in the paper of Raj et al. [3] and Varatharajan et al. [4], and support vector machine algorithm is used by them for pattern recognition. In [5], Zihlmann et al. propose the use of two deep neural networks (DNNs) architectures for the ECG classification by assessing the PhysioNet/CinC Challenge 2017’s atrial fibrillation classification dataset. In [6], Hannui et al. implemented a DNN to classify 12 heart rate classes using a dataset having 91,232 single-lead ECGs from 53,549 patients. Denoising plays a crucial role in the classification of ECG signals. In this regard, Muduli et al. propose several filtering methods to mitigate the effect of noise before classification [7, 8]. A wearable ECG monitoring system is proposed by Xia et al. [9]. Data is sent to a computer using Bluetooth 4.2, where regression (softmax) is used to classify ECG beats. A long short-term memory (LSTM) neural network is implemented in the paper of Gao et al. [10], which uses the timing features in ECG signals. A convolutional neural network (CNN)-based algorithm is proposed in the paper of Liu et al. [11], which classifies the MIT-BIH arrhythmia dataset [12]. Li et al. [13] proposed a ResNet-based deep learning technique to classify arrhythmia. The work employed limited lead ECG signals. A CNN and an extreme learning machine (ELM)-based model are proposed in [14] to detect arrhythmia. Hammad et al. [15] presented an end-to-end method based on CNN with the focal loss for myocardial infarction (MI). In this method, they used only one stage for MI detection from the input ECG signals without using any machine learning stages. They obtained good accuracy. A deep residual convolutional neural network-based arrhythmia classification technique is proposed by using an overlapping segmentation method and discrete wavelet transform (DWT)-based denoising [16]. Despite the availability of numerous ECG classification tools, a few methods can be deployed on an edge device or Webserver for real-time applications. Computational complexity is one of the major concerns. In this paper, we present a Web ECG signal classification system capable of classifying different arrhythmia according to the ANSI/AAMI EC57 standard [17]. The proposed system is capable of classifying normal beat, supraventricular premature beat, premature ventricular contraction, a fusion of ventricular and normal beat, and unclassifiable beat. The proposed model is trained and then tested using the arrhythmia dataset (MIT-BIH), which contains

24 Development and Implementation of an Efficient Deep Residual …

271

109,449 ECG signal samples of different classes. After training and validation, the model was quantized, where the post-training quantization technique was used, using the TensorFlow Lite (TFL) conversion method. This technique drastically reduces the model size, enabling the deployment of our model on edge devices, smartphones, and Web applications.

24.2 Methodology This section presents the methods used for the development of the CNN model to classify of ECG signals. The step-by-step development of Web application and deployment process is explained in this section. Figure 24.1 provides a general and top-level view of the proposed method in this paper. The dataset is preprocessed before it is used for training purposes. The model designed is trained on the dataset, and then, it is tested in the testing dataset or validation dataset. If the results are not satisfactory, changes are made in the model architecture to increase its accuracy. Once the satisfactory result is obtained, we go to the next step of quantization. And finally, a Web application is made on our quantized neural network. Figure 24.1 demonstrates the overview of the proposed method. It consists of the following main sections: (a) CNN training, (b) model quantization using TensorFlow Lite, which aims to optimize the model and reduce its size, and (c) deployment on edge/Web applications for the final classification of ECG signal.

24.2.1 Neural Network Training and Architecture Herein, we use an improved CNN architecture for the classification of ECG beat types using the real-world database. Two different types of blocks have been used in

Fig. 24.1 Overall workflow of the proposed method

272

R. Arya et al.

model architecture: (1) identity block and (2) convolutional block. Figures 24.2 and 24.3 illustrate the network architecture of identity block and convolutional block. The identity block consists of three convolutional layers, two batch normalization, two rectified linear activation function or ReLU activation layers, and a residual skip connection, and after that at the end of the block ReLU, nonlinearity is also used. The convolutional block is similar to the identity block and consists of three convolutional layers, two batch normalization, two ReLU activation layers, and a residual skip connection with one convolutional and batch normalization layer followed by the block ReLU. The ResNet model architecture employed in this paper is shown in Fig. 24.4. The model architecture begins with zero padding to the given input shape. Further, one convolutional layer with batch normalization and ReLU activation is used and one max pooling layer with stride 2 to reduce the input dimensions. Next, one convolutional block with 3 different filters of size 128, 128, and 256 and two identity blocks with 3 different filters of size 128, 128, and 256, all with stride 1 is used. One convolutional block with filters of size 128, 128, and 512, and three identity blocks with filters of size 128, 128, and 512, all with stride 2. Another convolutional block with filters of size 256, 256, and 1024, and five identity blocks with filters of size 256, 256, and 1024, all with stride 2. Further, a convolutional block with filters of size 512, 512, and 2048 and two identity blocks with filters of size 512, 512, and 2048, all with stride 2, average pooling layer with stride 2. The output layer is being flattened out, and two dense layers with 256, and 128 units and ReLU activation are used. Finally, to have a five-class classification, one dense layer with 5 units and softmax activation is used.

24.2.2 Quantized Neural Networks After the model has been trained, it is extracted using built-in functions. Generally, the size of the model is very large; in this case, the size is 204 MB (214,673,248 bytes), which is a large size for any edge device to be deployed. Hence, we need to reduce our model size before we can deploy it on the mobile application. Quantization is one of the best ways to reduce the model size, so that it can be run on edge devices. Quantization is the process of converting data (like weights and activations of the convolution neural network) from FP32/FP64 (floating point 32 bits or 64 bits) to a smaller precision like INT8 (integer 8 bit). Quantized neural networks (QNNs) use low precision weights and activations. Whereas in precision, QNN with fewer bits needs deeper and wider network architectures compared to the networks that use more precise operators, while requiring less complex arithmetic and fewer bits by weight. There are two ways to perform quantization, such as the first one is post-training quantization, and the second is quantization aware training. For this application, post-training quantization was performed. It essentially reduces the model size and improves CPU latency with acceptable accuracy. These techniques can be performed on a trained TensorFlow model and applied during TFL conversion. After quantization, the size of the model reduces drastically to 17.6 MB

24 Development and Implementation of an Efficient Deep Residual … Fig. 24.2 Identity block

Fig. 24.3 Convolutional block

273

274

R. Arya et al.

Fig. 24.4 ResNet model architecture

(18,475,888 bytes). There is a 91.37% of size reduction. This light model can be integrated with edge devices to execute real-world signal classification tasks.

24.2.3 Deployment Once we achieve a model with reduced size and acceptable accuracy, we can deploy it for ECG diagnosis via a mobile application or a Web application using our model. The Web application interface can be observed according to Fig. 24.5. The home page as shown in Fig. 24.5 has two routes: (1) register—for new users, and (2) login for existing users. The registration page takes in a user’s personal information like name, age, gender, profile picture, email id, and password. It also has a sign-up with google feature for users to directly create their authorization credentials by signing up with their google accounts.

Fig. 24.5 Home page, registration page, and login page

24 Development and Implementation of an Efficient Deep Residual …

275

Figure 24.6 shows that an email verification link is sent to the email id provided by the user during registration as soon as the ‘register’ button is hit, which authenticates the user and directs us to the login page. The login page provides two ways for users to log in: (1) using the email and password, (2) sign-up with a google account. From the user’s profile page, one can update their profile, i.e., their name, age, or profile picture can be edited, submit their ECG report, or log out. The Web application framework Express.js is used to provide a robust set of features for Web and mobile applications. MongoDB, which is a NoSQL database program, is employed in this paper. Some of the important Javascript libraries used are mongoose, nodemailer, multer, passport, and GoogleStrategy. Coming to the main part, how users are going to benefit from the Website. We are here dealing with a simplified version, wherein the user is going to input a CSV file, and our model is going to predict the heartbeat type. We need to put together the following two things: (1) Python code loads the model, gets user input from the Web form, makes predictions, and returns the result. (2) HTML templates allow users to enter their data and display the results. Flask takes the POST request, extracts the input, runs the model, and eventually renders the main.html file when the user fills out the form on the page and clicks the submit button. After that, the user can check the outcome and send more data to be processed. Flask extracts data from a form that acts like a dictionary. You can get the data you need from the form by referencing it by name. We preprocessed the data in the same way as we preprocessed the training data while we were training the model. The model generates a [1 × 5] vector where each of the 5 indices has a fraction showing the probability for each of the 5 classes. Then, we pick the index with the maximum fraction and display the label corresponding to the index. Eventually, the function yields the main.html template and generates the result as depicted in Fig. 24.7. The first query takes about 10 s to get executed because it involves loading the model and preparing the normalizer for normalizing the data. After that, each query

Fig. 24.6 Email id verification and user’s profile

276

R. Arya et al.

Fig. 24.7 Online forms rendered by the GET and the POST query

gets executed in less than a second. That is, if the same user uploads multiple files, then each query except the first one gets executed at a very high speed.

24.3 Dataset Description and Analysis In this work, we employed the MIT-BIH dataset, which contains 109,449 samples from 47 individuals, with a sampling frequency of 125 Hz and about 48 h of recording, the longest beats have 260 values [10]. The samples are fixed to a dimension of 187 values. The data collected consists of samples of ECG signals that had a recording time of 1400 ms. The number of samples is divided into five classes, namely class 0 with 90,589, class 1 with 2779, class 2 with 7236, class 3 with 803, and class 4 with 8039 samples. The given dataset is segmented into the training data and the testing data. The training dataset has total of 87,554 samples, and the testing dataset has 21,892 samples. Each class is defined as follows. Class 0 (Normal beat): That would be a sign of ECG from a person with heart health under normal and healthy conditions. Class 1 (Supraventricular): Supraventricular is a fast and irregular type of heartbeat that affects the upper chamber of the heart. Class 2 (Premature ventricular contraction): Premature ventricular contraction refers to the state when extra heartbeats begin in one of the lower chambers. Class 3 (Fusion of a ventricular and normal beat): Ventricular fusion beats happen whenever the natural rhythm of our heart and the impulse from a pacemaker coincide. Class 4 (Unclassifiable beat): Beats have not been associated with any rating.

24.4 Numerical Experiments Model performance can be evaluated via metrics like AUC score, F1 score, confusion matrix, classification score, precision score, and accuracy score. While training our ResNet model, area under the curve or AUC score has been used as metric. Similarly, the ROC curve is a graphical plot for classifying problems in various classification thresholds. ROC is a probability curve, and AUC represents the degree or measure of separability. We have considered other metrics like precision, F1 score, and recall

24 Development and Implementation of an Efficient Deep Residual …

277

to compare the performance of our model. As mentioned earlier, the database is split up into train, validation, and test sets. The AUC scores achieved on the validation, and testing sets are 99.83 and 99.76, respectively. Similarly, the F1 scores on the validation and testing sets are 98.30, and 98.13 respectively. The confusion matrices shown in Figs. 24.8 and 24.9 demonstrate the per class precision value on validation and test sets. Fig. 24.8 Confusion matrix obtained on validation data

Fig. 24.9 Confusion matrix obtained on testing data

278

R. Arya et al.

Table 24.1 Comparison of average accuracy

Work

Approach

Average accuracy (%)

Proposed method

Deep residual networks

98.38

Acharya et al. [21]

Augmentation + CNN

93.50

Martis et al. [18]

DWT + SVM

93.80

Li et al. [19]

DWT + Random forest

94.60

Kachuee et al. [20]

CNN

93.40

Table 24.2 Comparison of various metrics with different models Work

Approach

Accuracy

AUC

Precision

Proposed method

Deep residual networks

98.18

99.76

98.17

Kachuee et al. [20]

CNN

93.40

–

–

Acharya et al. [21]

CNN

93.50

–

–

Tables 24.1 and 24.2 demonstrate a performance study of various state-of-the-art methods. The best performance of the proposed method is highlighted in bold. A majority of the existing methods are too complex to be deployed on hardware platforms. However, the proposed method is lightweight, and it demonstrates an acceptable accuracy for ECG classification.

24.5 Conclusion and Future Scope This paper presents an improved deep residual network architecture and a Web application with improved CPU latency. The proposed method can classify 5 different types of heartbeats, where the first class is normal and the others arrhythmias. Furthermore, this paper presents a way to reduce the model size so that it can be implemented in the Web application. We provide a step-by-step methodology to make a Web application using our model. In future, one can deploy the proposed model on single board computers and edge devices, including a Raspberry Pi or other relevant embedded systems. The method can be extended to develop the Internet of medical things that can be deployed in the medical centers and the hospitals to make ECG classification faster and more accurately.

24 Development and Implementation of an Efficient Deep Residual …

279

References 1. Wang, D., Si, Y., Yang, W., Zhang, G., Liu, T.: A novel heart rate robust method for short-term electrocardiogram biometric identification. Appl. Sci. 9(1), 201 (2019) 2. Lu, W., Hou, H., Chu, J.: Feature fusion for imbalanced ECG data analysis. Biomed. Signal Process. Control 41, 152–160 (2018) 3. Raj, S., Ray, K.C.: ECG signal analysis using DCT-Based DOST and PSO optimized SVM. IEEE Trans. Instrum. Meas. 66(3), 470–478 (2017) 4. Varatharajan, R., Manogaran, G., Priyan, M.: A big data classification approach using LDA with an enhanced SVM method for ECG signals in cloud computing. Multimedia Tools Appl. 77(8), 10195–10215 (2018) 5. Zihlmann, M., Perekrestenko, D., Tschannen, M.: Convolutional recurrent neural networks for electrocardiogram classification. In: 2017 Computing in Cardiology (CinC), pp. 1–4 (2017) 6. Hannun, A.Y., Rajpurkar, P., Haghpanahi, M., Tison, G.H., Bourn, C., Turakhia, M.P., Ng, A.Y.: Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nat. Med. 25(1), 65 (2019) 7. Muduli, P.R., Mukherjee, A.: A robust estimator-based nonlinear filtering approach to piecewise biosignal reconstruction. IEEE Trans. Instrum. Measur. 69(2), 362–370 (2020) 8. Muduli, P.R., Mukherjee, A.: A moreau envelope-based nonlinear filtering approach to denoising physiological signals. IEEE Trans. Instrum. Measur. 69(4), 1041–1050 (2020) 9. Xia, Y., Zhang, H., Xu, L., Gao, Z., Zhang, H., Liu, H., Li, S.: An automatic cardiac arrhythmia classification system with wearable electrocardiogram. IEEE Access 6, 16529–16538 (2018) 10. Gao, J., Zhang, H., Lu, P., Wang, Z.: An effective LSTM recurrent network to detect arrhythmia on imbalanced ECG dataset. J. Healthc. Eng. (2019) 11. Liu, J., Song, S., Sun, G., Fu, Y.: Classification of ECG arrhythmia using CNN, SVM and LDA. In: International Conference on Artificial Intelligence and Security, pp.191–201. Springer (2019) 12. Goldberger, A.L., Amaral, L.A., Glass, L., Hausdorff, J.M., Ivanov, P.C., Mark, R.G., Mietus, J.E., Moody, G.B., Peng, C.-K., Stanley, H.E.: Physiobank, physiotoolkit, and physionet: components of a new research resource for complex physiologic signals. Circulation 101(23), 215–220 (2000) 13. Li, Z., Zhou, D., Wan, L., Li, J., Mou, W.: Heartbeat classification using deep residual convolutional neural network from 2-lead electrocardiogram. J. Electrocardiol 58, 105–112 (2020) 14. Zhou, S., Tan, B.: Electrocardiogram soft computing using hybrid deep learning CNN-ELM. Appl. Soft Comput. 86, 105778 (2020) 15. Hammad, M., Alkinani, MH., Gupta, BB., Abd El-Latif, AA.: Myocardial infarction detection based on deep neural network on imbalanced data. Multimedia Syst. 1–13 (2021) 16. Li, Y., Qian, R., Li, K.: Inter-patient arrhythmia classification with improved deep residual convolutional neural network, Comput. Methods Programs Biomed. 214(106582), 0169–2607 (2022) 17. The Advancement of Medical Instrumentation, A., et al.: Testing and reporting performance results of cardiac rhythm and ST segment measurement algorithms. In: ANSI/AAMI EC38 (1998) 18. Martis, R.J., Acharya, U.R., Lim, C.M., Mandana, K., Ray, A.K., Chakraborty, C.: Application of higher order cumulant features for cardiac health diagnosis using ECG signals. Int. J. Neural Syst. 23(04), 1350014 (2013) 19. Li, T., Zhou, M.: ECG classification using wavelet packet entropy and random forests. Entropy 18(8), 285 (2016)

280

R. Arya et al.

20. Kachuee, M., Fazeli, S., Sarrafzadeh, M.: ECG heartbeat classification: a deep transferable representation. In: 2018 IEEE International Conference on Healthcare Informatics (ICHI), pp. 443–444 (2018) 21. Acharya, U.R., Oh, S.L., Hagiwara, Y., Tan, J.H., Adam, M., Gertych, A., San Tan, R.: A deep convolutional neural network model to classify heartbeats. Comput. Biol. Med. 89, 389–396 (2017)

Chapter 25

Study of Class Incremental Learning Strategies for Intrusion Detection System Parvati Bhurani, Satyendra Singh Chouhan, and Namita Mittal

Abstract In today’s digital world, the Internet connects many apps and gadgets. This advancement is beneficial in many aspects, but on the other hand, it provides more chances for attackers to explore and exploit vulnerabilities. There is a rapid increase in cyberattack; due to this, there is a need for a self-adaptable defense mechanism system that evolves continually. The classic machine learning-based model fails to meet this condition since it is incapable of continuous learning. Specifically, neural network (NN)-based model suffers from catastrophic forgetting (CF). In this work, we implemented continual learning (CL) algorithm on the network dataset for a class incremental learning scenario. We have implemented three CL approaches, namely regularization (EWC, SI), rehearsal (replay, GEM), and hybrid on the benchmark dataset. Our findings suggest that rehearsal-based methods have comparable performance.

25.1 Introduction A cyberattack is an attempt to invade a computer system, multiple computers, or a network infrastructure with the intent to cause some harm. Cybercriminals launch cyberattacks to disrupt, disable, or gain unauthorized access to someone else computer or network [1]. Due to technological advancements, cyberattacks are increasing. WannaCry attack [2] is one of the worst ransomware attacks that happened in 2017, caused a $4 Billion loss and affected 150 countries. The attack causes a threat to confidentiality, integrity, and availability (CIA) of the information system. Intrusion Detection Systems (IDS) are robust solutions to provide security to the network system. It monitors traffic and devices connected to the network, and P. Bhurani (B) · S. S. Chouhan · N. Mittal Malaviya National Institute of Technology, Jaipur 302017, Rajasthan, India e-mail: [email protected] S. S. Chouhan e-mail: [email protected] N. Mittal e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_25

281

282

P. Bhurani et al.

after analyzing, it will detect an intrusion or privacy violations. Once any suspicious activity is detected, the IDS will trigger the alert. The key objective of any IDS is to maintain the CIA of the information system. There have been numerous attempts to build machine learning (ML)-based IDS Systems [3]. As more devices become connected, a vast amount of data will be generated and transferred over the network. More devices and more data lead to more chances of attack. As a result, developing an ML-based security system that protects the network from security threats or cyberattacks becomes more difficult. As the network threat data is not stationary thus, it becomes complex to train the model on such kind of data; due to this complexity, we must have a system that is self-adaptable. Continual learning (CL) seems to be a promising solution to meet such requirements. Van de Ven in [4] discussed three scenarios for continual learning, namely Task Incremental, Domain Incremental, and Class Incremental. In our work, we have used a class incremental learning scenario, which is challenging and realistic. Class incremental learning represents the real-world problem situation closely. In this scenario, new classes or concepts come over time, and the model has to adopt the new set of classes without forgetting the learning of existing classes or concepts. While making inference, the task information is not provided to the model explicitly, which makes it challenging. Incremental IDS is a challenging problem as the new network intrusion variants are continually evolving, and the issue of catastrophic forgetting (CF) is also there [5]. Catastrophic forgetting means the loss of knowledge learned from past tasks whenever a static model tries to learn a new task. As a result, past learned tasks will be underperformed. This phenomenon does not align with biological learning. Consider a model which incrementally learns the different classes of dogs, and the model tends to forget the already learned class of dog; it remembers only the recently learned class. To mitigate the issue of CF, several approaches are mentioned in [6]; these approaches are Regularization, Rehearsal, and Parameter Isolation. By using a regularization term, keeping a small subset of exemplars from previous tasks, and providing task-appropriate model parameters, the approaches above attempt to preserve the learned knowledge of earlier tasks. In real life, intrusion data is gathered in stages as intrusions appear over time rather than all at once. Due to high storage and computational overhead, it is not feasible to have a model which needs to be trained from scratch whenever a new attack class is encountered. In one recent paper, the author used the class incremental learning (CIL) and domain incremental learning (DIL) setup. Elastic weight consolidation (EWC) and Gradient Episodic Memory(GEM) on the network dataset have been evaluated [7]. They reported the impact of task order execution on model performance as the dataset was imbalanced. In our work, we apply the framework of CL to make IDS continually evolving. We conducted the experimental analysis of various CL approaches, namely regularization, replay, and hybrid for CIL setup on the CICIDS2018 dataset. We observed that some form of rehearsal helps in mitigating the CF issue. The detailed observations are reported in the result section.

25 Study of Class Incremental Learning Strategies …

283

25.2 Literature Review In this paper, we applied various CL algorithms to build an intrusion detection system that is self-adaptable. Also, the aim is to reduce the retraining overhead. According to the literature, significant advances in CL have been focused on the computer vision domain. Very little attention has been given to Network Intrusion Detection System based on CL. This section presents relevant literature on continual learning first; later in this section, we discuss the work done specifically in the network security domain. Kirkpatrick et al. [8] proposed Elastic Weight Consolidation(EWC), in which they used Fisher Information Matrix(FIM) to decide the importance of task parameters. It is one of the pioneering works in the field of continual learning. Liu et al. [9] extended the work proposed in [8] and introduced the diagonal FIM to mitigate the CF issue. The parameter space is rotated to have a better approximation of the FIM. Amalapuram et al. [10] addressed the impact of task order execution due to the presence of class imbalance (CI) on network data. They use the algorithmic level method, which preserves the original data distribution to deal with the CI issue. Author in [11] proposed the concept of progressive network. This network will retain the pool of pre-trained networks throughout the training process, and it learns the lateral connections to extract the useful features to learn the new task. The size of the network grows with the number of tasks. Mallaya et al. [12] proposed iterative pruning and network retraining to handle the issue of catastrophic forgetting. With the help of network pruning, the size of the network will remain fixed. There will be a change in the network connectivity due to pruning, which may impact the model’s performance. So to regain the accuracy, the network is retrained for a smaller number of epochs. The proposed approach uses explicit connections for every task, thus limiting the total number of tasks. Authors also showed that the order of tasks during training is important, impacting the overall model performance. Therefore, it is advantageous to add more challenging tasks at the initial model training stage. Yasir et al. [13] proposed hybrid approach for incremental learning. For classification of attack, they used Bidirectional Long short-term memory, which is the extension of LSTM. While for anomaly detection, semi-supervised approach has been used. To distinguish between anomaly instances with normal instances, novelty detector is utilized. For this, Local Outlier Factor (LOF) has been used. In [7], author proposed the framework of CL and implemented the class incremental learning(CIL) and domain incremental learning(DIL) setup. They found that task order execution has a greater impact in CIL setup due to an imbalanced dataset than in the DIL setup. They evaluated Elastic weight consolidation (EWC) and Gradient Episodic Memory (GEM). Authors in [14] proposed Variational AutoEncoder for Anomaly detection. This work comes under the category of generative replay (GR), which uses synthetic samples for replay. In this work, generative capability of the VAE is used for synthetic samples. Doshi et al. in [15] proposed online anomaly detection for surveillance videos. The transfer learning mechanism has been used for feature extraction,

284

P. Bhurani et al.

Fig. 25.1 Overview of experimental methodology

which reduces overall training complexity. Also, the statistical framework is proposed for sequential anomaly detection, which makes continual learning possible without CF.

25.3 Experimental Methodology The experimental methodology is given in Fig. 25.1. There are two sections, namely Data Preprocessing and Continual Learning Strategies.

25.3.1 Data Preprocessing Preparing data for any learning algorithm is critical; one needs to handle data efficiently to improve the model’s overall performance. Feature engineering is such technique that makes data suitable for processing. • Feature Engineering: It has been observed in the benchmark dataset that some features are unnecessary as they are redundant and should be removed during preprocessing. Also, some features hold infinity and nan value which are replaced by 0. In the original dataset, there are 80 features; after feature engineering, only 70 features will be there. This step makes the dataset appropriate for the algorithm. The removed features are Bwd PSH Flags, Bwd URG Flags, Fwd Byts/b Avg, Fwd Pkts/b Avg, Fwd Blk Rate Avg, Bwd Byts/b Avg, Bwd Pkts/b Avg, Bwd Blk Rate Avg , and Timestamp. • Normalization: It is a feature scaling technique used to change the numeric column values to a common scale without disturbing differences in the range of values. Normalization can be of min-max normalization, mean normalization, and standardization. In this work, we have used min-max normalization.

25 Study of Class Incremental Learning Strategies …

x_norm =

285

x − min(x) max(x) − min(x) + 

(25.1)

Equation 25.1 brings all numeric columns in the range of 0 and 1. Here, a small number  is used to avoid dividing by zero. • Task Split: We divided the preprocessed dataset into four tasks, and each task consisted of two or more classes. The task order has been created by keeping benign samples in different tasks. As we can introduce new classes with each task and no classes are repeated, this setting is known as Class Incremental Learning.

25.3.2 Continual Learning Strategies Due to catastrophic forgetting while learning the task continually, learning efficiency diminishes. There are two categories of algorithms; one does not handle the CF as in Eq. 25.2, whereas another type of algorithm employs some regularization constraint to deal with CF as in Eq. 25.3. Based on this categorization, we can have belowmentioned objective functions: • Without CF avoidance: No additional constraint is applied to mitigate CF. The model keeps on learning the new task without worrying about the performance of the past task. It is simply fine-tuning the existing model whenever a new data set enters the system. L(Dt , θ ) =

 1 ( f θ (x), y), |Dt | (x,y)∈D

(25.2)

t

Here, x and y are feature vector and ground truth label,  is a loss function, f θ (x) is a predictor function, Dt is the dataset for tth task. • With CF avoidance: An additional term is added as a constraint for CF avoidance. L(Dt , θ, φ) =

 1 ( f θ (x), y) + R(φ) |Dt | (x,y)∈D

(25.3)

t

Here, R(φ) is the constraint provided by applied CF avoidance approach. Now, we discuss the CL strategies implemented in this paper. Baseline strategies We evaluated the model by two baseline methods, namely naive and cumulative. 1. Naive: It is the most straightforward strategy and does not employ any constraint to mitigate catastrophic forgetting as in Eq. 25.2. It serves the soft lower bound on performance and fine-tunes the model incrementally. It is equivalent to serial transfer learning.

286

P. Bhurani et al.

2. Cumulative: This approach adapts the data whenever it comes and trains the model on both previous and current data. It serves the soft upper-bound performance, inefficient in terms of storage requirements. It is equivalent to the rehearsal method with unlimited storage capacity or sequential transfer learning on a continually expanding dataset. Regularization Strategies. Regularization-based methods are mentioned below. 1. EWC [8]: This approach computes importance as a diagonal approximation of the FIM of each weight at the end of training on current experience. Parameters of high importance are constrained during updates, and changes to these parameters are penalized while training the task. In EWC, performance degrades as the number of the task grows. L(θ ) = Lcurrent (θ ) +

λ 2

i

∗ Fi (θi − θprevious,i )2

(25.4)

where i is the label for each parameter, Lcurrent (θ ) is loss specific to current task, ∗ θprevious parameter after learning the previous task, λ computes the importance of new task w.r.t old task, F is the FIM. 2. SI [16]: The regularization penalty is similar to EWC. It suggests an efficient way to compute parameter importance. Parameter importance is calculated during training by approximating the effect on loss and gradient update. Modified loss function is used in [16] to reduce the large changes in important parameter θk while learning the new task. L∗new = Lnew + μ ∗



ψknew (θk∗ − θk )2

(25.5)

k

where ψknew is parameter-wise regularization strength. μ is a hyperparameter which tradeoff between old and new task memories. As opposed to EWC, it computes the parameter during training. Both EWC and SI come under the weight regularization category. Rehearsal Strategies Rehearsal-based approaches are given below. 1. Experience Replay [17]: Replay-based approaches keep a limited set of exemplars from previous tasks. It provides excellent performance at a modest computational cost. Overfitting to a subset of stored samples may be a matter of concern. Constrained optimization seems to be a promising solution for better transfer in both directions. 2. GEM [18]: It avoids forgetting and encourages positive backward transfer. This approach makes use of episodic memory Ms to store the examples from the seen task ‘s’. The learning model has a total memory budget of M units. The

25 Study of Class Incremental Learning Strategies …

287

gradient will be updated on the current task only when all the stored examples are maintained. This method tries to minimize the gap between human and machine learning by using the empirical Risk Minimization principle. L( f θ , Ms ) =

 1 ( f θ (x, s), y) |Ms | (x,y,s,)∈M

(25.6)

s

While observing the triplet (x, y, s), the objective is to solve the following problem: min  ( f θ (x, t), y) θ

  subjected to  ( f θ , Ms ) ≤  f θt−1 , Ms for all s < t Here, f θt−1 is the state of predictor after learning task t − 1. Hybrid Strategies. These approaches make use of various continual learning strategies to improve performance further. We combined both EWC and experience replay approaches to make our model better.

25.4 Result and Analysis This section presents the experimental results. We have used Avalanche framework [19] of continual learning based on PyTorch. We executed a continual learning algorithm on Multilayer Perceptron (MLP) in the first set of experiments. In the second set of experiments, we have used Convolutional Neural Network (CNN). For the evaluation purpose, we have used a dataset (CICIDS2018) prepared by The Canadian Institute for Cybersecurity (CIC).1 It is a multi-class attack dataset. The CICFlowmeter extracted the traffic features. It is one of the recent network attack datasets and covers a wide range of attacks. The dataset comprises 15 classes (1 benign and 14 attack classes) collected over ten days. The overall dataset comprises ten csv files. We concatenated these files during experimentation. In Table 25.1, it is visible that there is an imbalance in the dataset. In our experiments, we removed four classes with samples less than 10k and only kept 11 classes for processing. Performance Evaluation Measures To evaluate the performance of implemented continual learning strategies, we have used various measures, namely accuracy, precision, recall, F-score, and forgetting. The confusion matrix shows the correct and incorrect predictions made by the model. The standard performance measures such as accuracy, precision, recall, and F-score use the confusion metrics for computation. 1

https://www.unb.ca/cic/datasets/ids-2018.html.

288

P. Bhurani et al.

Table 25.1 Description of CICIDS2018 dataset S. No. Description 1 2 3 4 5 6

Total number of classes Total number of samples Total features Percentage of benign class Percentage of attack class Features after preprocessing

Statistics 15 16,233,005 80 83 17 70

The other vital measures related to continual machine learning are also considered Experience Forgetting and stream forgetting. Experience Forgetting is computed for every experience separately. It describes the loss detected over accuracy for a particular experience. Stream Forgetting is computed over all observed experiences during training. It is the average difference between the accuracy result obtained after first training on an experience and the accuracy result obtained on the same experience at the end of successive experiences.

25.4.1 Results While experimenting, we created different task orders (PIDs) by placing benign samples in different tasks for execution. Each PID consists of a set of experiences encountered in increments. One such PID is {{0, 1, 2}, {3, 4, 5}, {6, 7, 8}, {9, 10}}. Here, 0 represents benign class, and others correspond to attack classes. • It has shown in Tables 25.2 and 25.3 that Replay, GEM, and Hybrid strategies are superior to Naive, EWC, and SI. This is because these algorithms are using a buffer to keep the subset of samples from learned experiences. • Both experimental setups support that ordering of class impacts model performance. • In Fig. 25.2a, we can see that stream forgetting is more in the case of Naive and regularization approaches as compared to rehearsal and hybrid approaches. • Size of buffer impacts the performance of the algorithm as shown in Fig. 25.2b. Overall, in both experiments, we found that some form of rehearsal is beneficial to make the model better. It has been shown in Tables 25.2 and 25.3 that both rehearsal and hybrid approaches have accuracy metrics close to the cumulative approach, which gives a soft upper-bound performance.

25 Study of Class Incremental Learning Strategies …

289

Table 25.2 Comparison of performance metrics for various continual learning strategies on MLP architecture Performance PID Naive EWC SI Replay GEM Hybrid Cumulative metrics Accuracy

F1-score

Stream forgetting

PID0 PID1 PID2 PID3 PID0 PID1 PID2 PID3 PID0

0.18 0.4 0.43 0.41 0.07 0.31 0.37 0.31 0.55

0.18 0.4 0.46 0.41 0.06 0.31 0.42 0.31 0.55

0.25 0.4 0.48 0.42 0.19 0.31 0.45 0.34 0.47

0.74 0.92 0.66 0.43 0.77 0.95 0.63 0.33 0.1

0.81 0.4 0.68 0.43 0.87 0.31 0.64 0.34 0.07

0.74 0.93 0.67 0.43 0.77 0.96 0.63 0.33 0.09

0.95 0.96 0.7 0.44 0.97 0.98 0.65 0.33 0

PID1 PID2 PID3

0.55 0.57 0.25

0.52 0.37 0.25

0.54 0.37 0.25

0.03 0.09 0.06

0.44 0.19 0.13

0.02 0.09 0.06

0 0.09 0.06

Bold significance are the best results among EWC, SI, replay, GEM and hybrid Table 25.3 Comparison of performance metrics for various continual learning strategies on CNN architecture Performance PID Naive EWC SI Replay GEM Hybrid Cumulative metrics Accuracy

F1-score

Stream forgetting

PID0 PID1 PID2 PID3 PID0 PID1 PID2 PID3 PID0

0.19 0.40 0.45 0.41 0.08 0.31 0.40 0.31 0.54

0.19 0.40 0.48 0.41 0.08 0.31 0.46 0.31 0.54

0.25 0.40 0.45 0.41 0.18 0.31 0.41 0.31 0.46

0.75 0.93 0.67 0.43 0.78 0.96 0.64 0.33 0.09

0.81 0.41 0.67 0.41 0.87 0.31 0.64 0.32 0.07

0.72 0.93 0.66 0.43 0.74 0.96 0.63 0.33 0.10

0.95 0.96 0.70 0.43 0.97 0.98 0.65 0.33 0.00

PID1 PID2 PID3

0.50 0.37 0.25

0.51 0.38 0.25

0.50 0.37 0.27

0.02 0.09 0.06

0.42 0.22 0.19

0.02 0.10 0.06

0.01 0.09 0.06

Bold significance are the best results among EWC, SI, replay, GEM and hybrid

25.5 Conclusion We used a highly evolving network attack dataset to conduct an experimental analysis of various continual learning approaches. We observe that regularization approaches are more prone to forgetting than methods that rely on episodic memory. It is good

290

P. Bhurani et al.

(a) Stream Forgetting Illustration

(b) Buffer Size Impact

Fig. 25.2 a Stream forgetting. b Buffer size impact

to keep some form of rehearsal to deal with catastrophic forgetting. We observed that size of the buffer impacted the performance. In the future, we would like to investigate the optimal strategies to maintain the buffer space to keep the exemplars. Other directions of future research would be the design of a systematic approach to formulating an efficient class ordering for execution.

References 1. Cyber attack statistics data and trends kernel description. https://parachute.cloud/2022-cyberattack-statistics-data-and-trends. Accessed 29 Apr 2022 2. WannaCry ransomware attack kernel description. https://en.wikipedia.org/wiki/WannaCry_ ransomware_attack 3. Patgiri, R., Varshney, U., Akutota, T., Kunde, R.: An investigation on intrusion detection system using machine learning. In: 2018 IEEE Symposium Series on Computational Intelligence (SSCI), pp. 1684–1691. IEEE (2018) 4. Van de Ven, G.M., Tolias, A.S.: Three scenarios for continual learning. arXiv preprint arXiv:1904.07734 (2019) 5. McCloskey, M., Cohen, N.J.: Catastrophic interference in connectionist networks: the sequential learning problem. In: Psychology of Learning and Motivation, vol. 24, pp. 109–165. Elsevier (1989) 6. Delange, M., Aljundi, R., Masana, M., Parisot, S., Jia, X., Leonardis, A., Slabaugh, G., Tuytelaars, T.: A continual learning survey: defying forgetting in classification tasks. IEEE Trans. Pattern Anal. Mach. Intell. (2021) 7. Amalapuram, S.K., Tadwai, A., Vinta, R., Channappayya, S.S., Tamma, B.R.: Continual learning for anomaly based network intrusion detection. In: 2022 14th International Conference on COMmunication Systems & NETworkS (COMSNETS), pp. 497–505. IEEE (2022) 8. Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J., Desjardins, G., Rusu, A.A., Milan, K., Quan, J., Ramalho, T., Grabska-Barwinska, A., et al.: Overcoming catastrophic forgetting in neural networks. Proc. Natl. Acad. Sci. 114(13), 3521–3526 (2017) 9. Liu, X., Masana, M., Herranz, L., Van de Weijer, J., Lopez, A.M., Bagdanov, A.D.: Rotate your networks: better weight consolidation and less catastrophic forgetting. In: 2018 24th International Conference on Pattern Recognition (ICPR), pp. 2262–2268. IEEE (2018)

25 Study of Class Incremental Learning Strategies …

291

10. Amalapuram, S.K., Reddy, T.T., Channappayya, S.S., Tamma, B.R.: On handling class imbalance in continual learning based network intrusion detection systems. In: The First International Conference on AI-ML-Systems, pp. 1–7 (2021) 11. Rusu, A.A., Rabinowitz, N.C., Desjardins, G., Soyer, H., Kirkpatrick, J., Kavukcuoglu, K., Pascanu, R., Hadsell, R.: Progressive neural networks. arXiv preprint arXiv:1606.04671 (2016) 12. Mallya, A., Lazebnik, S.: Packnet: adding multiple tasks to a single network by iterative pruning. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 7765–7773 (2018) 13. Farrukh, Y.A., Khan, I.: An autonomous self-incremental learning approach for detection of cyber attacks on unmanned aerial vehicles (UAVs). arXiv preprint arXiv:2112.11219 (2021) 14. Wiewel, F., Yang, B.: Continual learning for anomaly detection with variational autoencoder. In: ICASSP 2019—2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp. 3837–3841. IEEE (2019) 15. Doshi, K., Yilmaz, Y.: Continual learning for anomaly detection in surveillance videos. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pp. 254–255 (2020) 16. Zenke, F., Poole, B., Ganguli, S.: Continual learning through synaptic intelligence. In: International Conference on Machine Learning, pp. 3987–3995. PMLR (2017) 17. Rolnick, D., Ahuja, A., Schwarz, J., Lillicrap, T., Wayne, G.: Experience replay for continual learning. Advances in Neural Information Processing Systems, vol. 32 (2019) 18. Lopez-Paz, D., Ranzato, M.: Gradient episodic memory for continual learning. Advances in Neural Information Processing Systems, vol. 30 (2017) 19. Lomonaco, V., Pellegrini, L., Cossu, A., Carta, A., Graffieti, G., Hayes, T.L., De Lange, M., Masana, M., Pomponi, J., Van de Ven, G.M. et al.: Avalanche: an end-to-end library for continual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 3600–3610 (2021)

Chapter 26

Classification of High-Priority Tweets for Effective Rescue Operations During Natural Disaster Combining Twitter’s Textual and Non-textual Features E. Arathi and S. Sasikala Abstract Disaster management is highly responsible for managing the evacuation and deploying rescue teams to mitigate the loss of life and property. Twitter assists in getting emergency aid requests from victims in crisis regions for quicker response by the rescue team. However, obtaining crucial information from the impacted zones on time is a challenge. This paper suggests a technique for categorizing help requests of the micro-blog text that require immediate attention by blending textual and nontextual attributes. This work aims to implement a model that gives a combined classification of textual features using the random forest (RF) classifier model and nontextual features (reply and retweet) value by applying the Fisher score to extract high-priority request tweets shared during a crisis event. Our experiment determines the ideal values for α and β statistical variables for type I error and type II error of classification for combining the two types of features using the standard Natural Hazard Standard dataset of tweets collected during various catastrophe events. The results show the model’s efficiency, with an accuracy of 90.52% and F1-score of 93.68% at an ideal value of alpha as 0.8% and beta as 0.2% indicating improved classification of combining textual and non-textual elements of the tweets.

26.1 Introduction Tweeting is the act of sharing one’s ideas or opinions with one’s followers. You can publish text messages up to 280 characters long, as well as links, images, and videos. Figure 26.1 illustrates a tweet’s text and non-textual characteristics. When a user’s followers view a tweet, they have the option to like (heart), retweet, or reply. Additionally, extra tasks, such as a hashtag or a mention, can be processed by utilizing the unique symbols described below in a tweet.

E. Arathi (B) · S. Sasikala Department of Computer Science, IDE, University of Madras, Chepauk, Chennai, India e-mail: [email protected] © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2023 V. Bhateja et al. (eds.), Intelligent Data Engineering and Analytics, Smart Innovation, Systems and Technologies 327, https://doi.org/10.1007/978-981-19-7524-0_26

293

294

E. Arathi and S. Sasikala

Fig. 26.1 Non-textual features of Twitter

• Hashtag (#): This is a term that is commonly used when referring to keywords in a tweet. When a user clicks on the link preceded with a “#” symbol, search results for the relevant term are displayed. • Mention (@) is a character that is used for the tweet notification, sent to the mentioned user by putting the user’s name after @, and you can refer to a specific person. Mention is most commonly used to start a conversation or a questionand-answer (Q&A) between two or more users, with the content of the exchange being made public. The widespread adoption of social networking sites such as Twitter and Facebook has resulted in a plethora of new applications, extending the role of social media beyond its original domain but is not limited to, analysis of health and detection of disease [1], quantification of contentious information [2], and disaster management [3, 4]. Due to damaged infrastructure, natural disasters typically disrupt routine communication, resulting in information outflow [5]. Retweeting is an activity that allows us to propagate tweet from another user’s timeline to our own, with the result that individuals who follow us will see the associated tweet as well. Retweet is a Twitter function symbolized by a little arrowsshaped icon; users can tap or click this button to retweet or repost a tweet. This repost function allows user to share a tweet with all of their followers quickly and effortlessly. It is possible to retweet one’s own or someone else’s tweets [6]. Responding to tweets is an interesting strategy to maintain a friendly tone and engage in conversation on Twitter. There is a reply represented using dialog symbol, clicking this icon, allows the user to mention the reply with the “@username” in the message for a specific user. Otherwise, responses are public, which means they are accessible on your Twitter profile. In this paper, we utilize random forest (RF) classifier with GloVe vectors as model to effectively classify the prioritized information from the tweet text of a natural disaster. The proposed classification engine involves two phases. First phase involves preprocessing of texts, annotations to prioritize, vectorization using GloVe word vector, RF classification with its performance metrics. The classification is made effective using a pre-trained word vectors that includes the global vectors for word representation (GloVe). This pre-training captures the semantic meaning from the input tweets. Finally, extraction is performed to increase the accuracy of the

26 Classification of High-Priority Tweets for Effective Rescue Operations …

295

model. The second phase incorporates both textual and non-textual features prediction using combined feature random forest (CFRF) classification approach. This method arrives to combined classification result using the categorization result of the previous phase with weights of non-textual characteristics (reply, retweet) determined by Fisher score. The structure of this paper is as follows: The second section contains similar works that describe the theory and concept employed in this work. Section 26.3 displays the proposed method of integrating the two parts and their respective functions. Section 26.4 describes the experimental design and validation approach for evaluating the proposed combining strategy. Section 26.5 contains a summary of presented work, a conclusion, and a request for further work. This study can be further extended for other features derived from the combination of both features which can improvise the classification of the tweet text.

26.2 Related Works Apart from text, [7–11] a tweet contains a plethora of the attributes or features that can be used for classification on the social media tweets categorization which gains immediate response, such as the user’s language, the time the tweet was generated, the person who shared the tweet, entities, a like indicator, the number of users who liked the tweet, the geo-location (longitude and latitude) or location of the tweets posted, and the tagging system’s unique id. A Hurricane Sandy report [12] shows that more people communicate through social media. Requesting help, seeking information on transportation, shelter, and food, while trying to get better communication via family/friends in and out of the disaster region. This makes it more beneficial to manage a natural disaster by means of the huge information flow across social media. Although [13] the use of social networks appears to be appealing, the majority of applications still lack features, and they are inoperable. Retweet is a technique for disseminating information via a process that increases the popularity of the tweet. The [14] amount of retweets indicates the popularity of a tweet, which can be utilized to better the process of detecting sentiment in a tweet. In [15], employed feeding a self-exciting point processes model without features, merely on the amount of retweets over time. This paper [16] investigates the predictive power of Twitter response measures such as “likes,” “replies,” and “retweets” for classifying tweets into three classes: low, medium, and high response using four ML algorithms with the F1-score of 0.655. In [17], it presents a grouping of textual and non-textual features to boost sentiment prediction performance. Naive Bayes classifier was used which resulted in the F1-score as 0.838, while alpha and beta are 0.6 and 0.4, respectively. In [18], non-textual elements were used for classification, accomplished by the RF classifier [18] which gave accuracy of 81.3%. The authors of [19] investigated tweets posted after multiple hurricane-related crises in the United States. Experiments were conducted utilizing the bag-of-words—RF classifier model

296

E. Arathi and S. Sasikala

for multi-class classification, which includes classes such as impacted individual, infrastructure and utilities damage, caution, and advice, etc., with an 83.7% accuracy rate. The studies [20–22] concentrated on approaches for prioritizing and classifying tweets based on event type and location. The authors of [23] used datasets from earthquakes, floods, hurricanes, and forest fires to train a random forest classifier to automatically classify eyewitness reports as direct eyewitness, indirect eyewitness, or vulnerable direct eyewitness. The earthquake datasets from Nepal, Macedonia, and Kerala were used to test low-supervision and transfer learning-based algorithms for detecting urgent tweets in [24]. Their proposed strategy has been proved to outperform current standard procedures. This research [25] examines Twitter posts during a flood-related tragedy and proposes an algorithm to detect victims requesting assistance, which categorizes tweets as high priority or low priority. Also Markov model predicts the user location of high-priority tweets based on the historical locations of the users. The categorization accuracy of the system was 81%, while the location prediction accuracy was 87%. All of these initiatives, however, concentrate on disaster-related tweets as opposed to subcategories of tweets such as infrastructure damage, human harm, and resource demand and availability. In addition, disaster-related data, events, and locations are categorized primarily. The majority of the works concentrate on textual and syntactic characteristics. Aside from textual features, other Twitter features can help to increase the importance of information propagated. Information can also be prioritized based on the event’s situational awareness. To overcome these limitations, a model is proposed that combines both types of Twitter features to prioritize help requests for effective disaster relief operations.

26.3 Methodology The proposed method entails extracting and classifying high-priority tweets with GloVe pre-trained feature vectors and random forest classifier, and the weights of the tweet’s non-textual features are combined to disseminate the tweets. Micro-blog text requesting assistance with food, housing, medical assistance, or medications is classified as high priority, whereas tweets containing general information are classified as low priority. Manual annotation is necessary, as tweets include no information about the class. Manual annotation of class information is done on preprocessed tweets. This works are segregated into two phases represented in Fig. 26.2.

26 Classification of High-Priority Tweets for Effective Rescue Operations …

297

Fig. 26.2 Proposed methodology

26.3.1 Preprocessing The preprocessing steps for cleaning tweets are case folding, punctuation removal, stop words removal [26], stemming [27], and tokenization [28]. These measures are taken to eliminate words and symbols that are not helpful in the learning process. Data that has been preprocessed is later used in feature extraction for model training.

298

E. Arathi and S. Sasikala

Algorithm preprocess_tweets(dara) Step 1: fil_data