Model Optimization Methods for Efficient and Edge AI E-Book
126,99 €
Mehr erfahren.
- Herausgeber: John Wiley & Sons
- Kategorie: Wissenschaft und neue Technologien
- Sprache: Englisch
Comprehensive overview of the fledgling domain of federated learning (FL), explaining emerging FL methods, architectural approaches, enabling frameworks, and applications
Model Optimization Methods for Efficient and Edge AI explores AI model engineering, evaluation, refinement, optimization, and deployment across multiple cloud environments (public, private, edge, and hybrid). It presents key applications of the AI paradigm, including computer vision (CV) and Natural Language Processing (NLP), explaining the nitty-gritty of federated learning (FL) and how the FL method is helping to fulfill AI model optimization needs. The book also describes tools that vendors have created, including FL frameworks and platforms such as PySyft, Tensor Flow Federated (TFF), FATE (Federated AI Technology Enabler), Tensor/IO, and more.
The first part of the text covers popular AI and ML methods, platforms, and applications, describing leading AI frameworks and libraries in order to clearly articulate how these tools can help with visualizing and implementing highly flexible AI models quickly. The second part focuses on federated learning, discussing its basic concepts, applications, platforms, and its potential in edge systems (such as IoT).
Other topics covered include:
- Building AI models that are destined to solve several problems, with a focus on widely articulated classification, regression, association, clustering, and other prediction problems
- Generating actionable insights through a variety of AI algorithms, platforms, parallel processing, and other enablers
- Compressing AI models so that computational, memory, storage, and network requirements can be substantially reduced
- Addressing crucial issues such as data confidentiality, data access rights, data protection, and access to heterogeneous data
- Overcoming cyberattacks on mission-critical software systems by leveraging federated learning
Written in an accessible manner and containing a helpful mix of both theoretical concepts and practical applications, Model Optimization Methods for Efficient and Edge AI is an essential reference on the subject for graduate and postgraduate students, researchers, IT professionals, and business leaders.
Sie lesen das E-Book in den Legimi-Apps auf:
Seitenzahl: 853
Veröffentlichungsjahr: 2024
Ähnliche
Table of Contents
Cover
Table of Contents
Title Page
Copyright
About the Editors
List of Contributors
1 Fundamentals of Edge AI and Federated Learning
1.1 Introduction
1.2 Concepts and Fundamentals of Edge AI
1.3 Concepts and Fundamentals of FL
1.4 Combining FL and Edge AI
1.5 Background
1.6 Applications of Edge AI and FL
1.7 Challenges, Future Research Directions, and Solutions
1.8 Conclusion
References
2 AI Applications – Computer Vision and Natural Language Processing
2.1 Introduction
2.2 Artificial Intelligence: 2023’s Top 18 Applications
2.3 AI in Computer Visions
2.4 AI in Natural Language Processing
2.5 Conclusion
2.6 AI’s Opportunities for Computer Vision and Natural Language Processing
References
3 An Overview of AI Platforms, Frameworks, Libraries, and Processors
3.1 Introduction: Artificial Intelligence Platforms, Frameworks, Libraries, and Processors: The Building Blocks of AI Development
3.2 Edge AI
References
4 Model Optimization Techniques for Edge Devices
4.1 Overview of Model Optimization Techniques
4.2 Deep Dive of Predeployment Model Optimization Techniques
4.3 Deep Dive of Deployment-Time Model Optimization Techniques
4.4 Deep Dive of Post-Deployment Model Optimization Techniques
4.5 Summary
References
5 AI Model Optimization Techniques
5.1 Introduction
5.2 Pruning
5.3 Quantization
5.4 Model Distillation
5.5 Layer Fusion
5.6 Parallelization
5.7 Hardware Acceleration
5.8 Transfer Learning
5.9 Neural Architecture Search
5.10 Pragmatic Optimization
5.11 Conclusion
References
6 Federated Learning: Introduction, Evolution, Working, Advantages, and Its Application in Various Domains
6.1 Introduction to Machine Learning and Federated Learning
6.2 Evolution of Federated Learning
6.3 How Federated Learning Works
6.4 Some Scholarly Work Related to Federated Learning
6.5 Distinct Advantages of Federated Learning
6.6 Applications of Federated Machine Learning
6.7 Conclusion
References
Note
7 Application Domains of Federated Learning
7.1 Introduction
7.2 Healthcare
7.3 Finance and Banking
7.4 E-Commerce and Recommender Systems
7.5 Conclusion
References
8 Advanced Architectures and Innovative Platforms for Federated Learning: A Comprehensive Exploration
8.1 Introduction
8.2 Basics of Federated Learning
8.3 Advanced Architectures for Federated Learning
8.4 Innovative Platforms for Federated Learning
8.5 Security and Privacy in Federated Learning
8.6 Optimization Techniques for Federated Learning
8.7 Real-World Applications and Case Studies
8.8 Challenges and Future Directions
8.9 Conclusion
References
9 Federated Learning: Bridging Data Privacy and AI Advancements
9.1 Introduction
9.2 Horizontal
9.3 Vertical
9.4 Federated Transfer Learning
9.5 Optimization Algorithms in Federated Learning
9.6 Federated Learning Applications
9.7 Conclusion
References
10 Securing Edge Learning: The Convergence of Block Chain and Edge Intelligence
10.1 Introduction
10.2 Fundamentals of Block Chain
10.3 Edge Intelligence
10.4 Securing Edge Learning with Block Chain
10.5 Existing Research and Case Studies
10.6 Framework for Securing Edge Learning with Block Chain
10.7 Evaluation and Future Directions
10.8 Conclusion
References
11 Training on Edge
11.1 Introduction
11.2 The Field of Training on the Edge
11.3 Incremental Training on Edge
11.4 Data at the Edge
11.5 Distributed Edge Training
11.6 Use Cases
11.7 Conclusion
References
12 Architectural Patterns for the Design of Federated Learning Systems
12.1 Introduction
12.2 Federated Learning Fundamentals
12.3 Architectural Frameworks
12.4 Design Patterns in Federated Learning
12.5 Applications of Federated Learning
12.6 Technologies and Methodologies
12.7 Implementing Federated Learning Systems
12.8 Emerging Trends and Future Scopes
12.9 Conclusion
References
13 Federated Learning for Intelligent IoT Systems: Background, Frameworks, and Optimization Techniques
13.1 Introduction
13.2 Basics of Federated Learning
13.3 IoT Architectures and Their Need for Federated Learning
13.4 Intelligent Federated Learning Frameworks Suitable for IoT
13.5 Optimization Techniques for Federated ITIoT Systems
13.6 Security and Privacy Considerations
13.7 Conclusion
Acknowledgment
References
14 Enhancing Cybersecurity Through Federated Learning: A Critical Evaluation of Strategies and Implications
14.1 Introduction
14.2 Literature Review
14.3 Theoretical Framework
14.4 Methodology
14.5 Empirical Study
14.6 Conclusion
References
15 Blockchain for Securing Federated Learning Systems: Enhancing Privacy and Trust
15.1 Introduction
15.2 Literature Review
15.3 Fundamentals of Blockchain Technology
15.4 Federated Learning: A Primer
15.5 Blockchain Integration with Federated Learning Systems
15.6 Enhancing Privacy in Federated Learning Through Blockchain
15.7 Trust and Security in Federated Learning via Blockchain
15.8 Use Cases of Blockchain-Enabled Federated Learning
15.9 Challenges and Limitations
15.10 Future Prospects and Research Directions
15.11 Conclusion
References
16 Blockchain-Enabled Secure Federated Learning Systems for Advancing Privacy and Trust in Decentralized AI
16.1 Introduction
16.2 Fundamentals of Federated Learning
16.3 Practical Significance of Federated Learning
16.4 Privacy and Security Issues in Federated Learning
16.5 Practical Implementation of Federated Learning
16.6 Case Studies
16.7 Conclusion
16.8 Future Scope
References
17 An Edge Artificial Intelligence Federated Recommender System for Virtual Classrooms
17.1 Introduction
17.2 Work Related To
17.3 Federated Recommender-Proposed Framework
17.4 Final Thoughts and Future Work
References
18 Federated Learning in Smart Cities
18.1 Introduction
18.2 Background
18.3 Application and Scenario
18.4 FL in Industrial Application
18.5 FL in the Healthcare System
18.6 FL in Medicine Recommendation
18.7 FL in Drug Discovery
18.8 FL in Fault Tolerance
18.9 Future Aspects and Research Benefits
18.10 Conclusion
18.11 Parameters’ Declaration
References
Index
End User License Agreement
List of Tables
Chapter 1
Table 1.1 FL and Edge AI challenges, future research directions, and solutio...
Chapter 3
Table 3.1 AI frameworks and their inference solutions.
Table 3.2 GPU Specifications and their importance.
Table 3.3 CPU Specifications and their importance.
Chapter 4
Table 4.1 Comparison overview of post-training quantization and quantization...
Chapter 5
Table 5.1 Summary of optimization techniques.
Table 5.2 Description for each subsection and reference.
Chapter 8
Table 8.1 Comparison of Centralized and Federated Learning.
Table 8.2 Overview of Open-source Platforms for Federated Learning.
Table 8.3 Security Mechanisms in Federated Learning.
Table 8.4 Optimization Techniques and Their Impact.
Table 8.5 Real-World Federated Learning Applications Across Industries.
Chapter 13
Table 13.1 Comparison of federated learning platforms for ITIoT.
Chapter 16
Table 16.1 Difference in practical and fundamental federated learning.
List of Illustrations
Chapter 1
Figure 1.1 Edge AI architecture.
Figure 1.2 Taxonomy of edge AI applications.
Figure 1.3 FL architecture.
Figure 1.4 Taxonomy of FL applications.
Chapter 2
Figure 2.1 Overview of AI—computer vision and natural language processing.
Figure 2.2 Three subcategories of artificial intelligence.
Chapter 3
Figure 3.1 Overview of the layers of Edge AI [1].
Figure 3.2 MLIR workflow from frameworks to hardware [31].
Chapter 4
Figure 4.1 Model optimization techniques classified into predeployment, depl...
Figure 4.2 Quantization-aware training and post-training quantization overvi...
Figure 4.3 Knowledge Distillation Overview [36].
Figure 4.4 Dynamic batching in triton inference server.
Figure 4.5 Model monitoring deployed using Cnvrg.io in VMware Tanzu MLOPs pi...
Chapter 5
Figure 5.1 Computational optimization techniques.
Figure 5.2 Efficiency optimization techniques.
Figure 5.3 Pruning: an efficiency optimization technique that can be used to...
Figure 5.4 Quantization reduces the precision of weights to improve inferenc...
Figure 5.5 Model distillation transfers knowledge from a large model to a sm...
Figure 5.6 Layer fusion merges neural network layers to improve efficiency o...
Chapter 6
Figure 6.1 Federated learning architecture.
Figure 6.2 Working of federated learning in five steps.
Figure 6.3 Federated learning in hospital sector.
Figure 6.4 Security of private data during training in Federated Learning.
Figure 6.5 Federated learning for providing more diverse data.
Figure 6.6 Applications of Federated Learning.
Chapter 7
Figure 7.1 Clinical decision support systems.
Figure 7.2 Federated learning in financial sector.
Figure 7.3 Fraud Detection and Prevention with Federated Learning.
Figure 7.4 E-commerce and Recommender Systems.
Chapter 10
Figure 10.1 Block chain technology.
Figure 10.2 Cryptocurrency.
Figure 10.3 Distributed ledger technology.
Figure 10.4 Consensus mechanisms.
Figure 10.5 Smart contract.
Figure 10.6 51% attack.
Figure 10.7 Sybil attack.
Figure 10.8 Double spending.
Figure 10.9 Edge intelligence.
Figure 10.10 Edge learning.
Figure 10.11 Edge learning architecture.
Figure 10.12 Identity and access management.
Figure 10.13 Reputation management.
Figure 10.14 Federated learning.
Figure 10.15 Fault tolerance.
Figure 10.16 Byzantine faults.
Chapter 11
Figure 11.1 MLOps pipeline demonstrating various stages. After the new data ...
Figure 11.2 Different types of incremental learning.
Figure 11.3 Illustration for learning without forgetting. (a) Original m...
Figure 11.4 Incremental Few Shot Learning (FS-IL).
Chapter 13
Figure 13.1 Organization of paper.
Chapter 14
Figure 14.1 General framework of Federated Learning. (a) Initialization Phas...
Figure 14.2 Federated learning for cybersecurity.
Figure 14.3 Data aggregation methods.
Chapter 15
Figure 15.1 Blockchain model.
Figure 15.2 A typical blockchain structure for creating a chain in blockchai...
Figure 15.3 A typical blockchain structure (hardware).
Figure 15.4 Future of Federated Learning.
Chapter 16
Figure 16.1 Architecture of a blockchain-enabled secure federated learning s...
Figure 16.2 Global loss and accuracy.
Chapter 17
Figure 17.1
Figure 17.2 Engineering of the combined proposal framework.
Figure 17.3 A sample of information linked to a space.
Figure 17.4 Model for suggestion diagram.
Figure 17.5 The point of interaction of the gadget.
Chapter 18
Figure 18.1 A taxonomy tree of the chapter purposes. FL; Smart city; Industr...
Figure 18.2 A schematic of IoT smart city and metaverse smart city.
Figure 18.3 A schematic of IoT smart city and metaverse smart city.
Figure 18.4 (a) Application 1 based on FLIA definition, (b) scenario 1 based...
Figure 18.5 (a) Application 2 based on FHLS definition, (b) scenario 2 based...
Figure 18.6 (a) Application 3 based on FLMR definition, (b) scenario 3 based...
Figure 18.7 (a) Application 4 based on FLDC definition, (b) scenario 2 based...
Figure 18.8 (a) Application 5 based on FLFT definition, (b) scenario 2 based...
Figure 18.9 (a) Peer-to-Peer topology for FL, (b) task-based node classifica...
Figure 18.10 The three main steps of aggregating the global model based on F...
Figure 18.11 A neural network-based head cluster nomination.
Figure 18.12 Graph mapping by application nodes.
Figure 18.13 A hybrid hierarchal-based structure for implementing FL.
Figure 18.14 (a) Malicious role of FL for fault-tolerant, (b) constructive r...
Guide
Cover
Table of Contents
Series Page
Title Page
Copyright
About the Editors
List of Contributors
Begin Reading
Index
End User License Agreement
Pages
ii
iii
iv
xxi
xxii
xxiii
xxxiv
xxv
xxvi
xxvii
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
43
44
45
46
47
48
49
50
51
52
53
54
55
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
391
392
393
394
395
396
397
398
IEEE Press445 Hoes LanePiscataway, NJ 08854
IEEE Press Editorial BoardSarah Spurgeon, Editor-in-Chief
Moeness Amin
Jón Atli Benediktsson
Adam Drobot
James Duncan
Ekram Hossain
Brian Johnson
Hai Li
James Lyke
Joydeep Mitra
Desineni Subbaram Naidu
Tony Q. S. Quek
Behzad Razavi
Thomas Robertazzi
Diomidis Spinellis
Model Optimization Methods for Efficient and Edge AI
Federated Learning Architectures, Frameworks and Applications
Edited by
Pethuru Raj Chelliah
Chief Architect at the Edge AI division of Reliance Jio Platforms Ltd. (JPL), Bangalore, India
Amir Masoud Rahmani
Artificial intelligence faculty member at the National Yunlin University of Science and Technology, Taiwan
Robert Colby
Principal Engineer in IT Infrastructure responsible for Manufacturing Network Architecture and IoT Infrastructure at Intel Corporation
Gayathri Nagasubramanian
Assistant Professor, Department of Computer Science and Engineering, GITAM University, Bengaluru, India
Sunku Ranganath
Principal Product Manager for Edge Infrastructure Services at Equinix
Copyright © 2025 by The Institute of Electrical and Electronics Engineers, Inc. All rights reserved.
Published by John Wiley & Sons, Inc., Hoboken, New Jersey.Published simultaneously in Canada.
No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, scanning, or otherwise, except as permitted under Section 107 or 108 of the 1976 United States Copyright Act, without either the prior written permission of the Publisher, or authorization through payment of the appropriate per-copy fee to the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, (978) 750-8400, fax (978) 750-4470, or on the web at www.copyright.com. Requests to the Publisher for permission should be addressed to the Permissions Department, John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, (201) 748-6011, fax (201) 748-6008, or online at http://www.wiley.com/go/permission.
Trademarks: Wiley and the Wiley logo are trademarks or registered trademarks of John Wiley & Sons, Inc. and/or its affiliates in the United States and other countries and may not be used without written permission. All other trademarks are the property of their respective owners. John Wiley & Sons, Inc. is not associated with any product or vendor mentioned in this book.
Limit of Liability/Disclaimer of Warranty: While the publisher and author have used their best efforts in preparing this book, they make no representations or warranties with respect to the accuracy or completeness of the contents of this book and specifically disclaim any implied warranties of merchantability or fitness for a particular purpose. No warranty may be created or extended by sales representatives or written sales materials. The advice and strategies contained herein may not be suitable for your situation. You should consult with a professional where appropriate. Further, readers should be aware that websites listed in this work may have changed or disappeared between when this work was written and when it is read. Neither the publisher nor authors shall be liable for any loss of profit or any other commercial damages, including but not limited to special, incidental, consequential, or other damages.
For general information on our other products and services or for technical support, please contact our Customer Care Department within the United States at (800) 762-2974, outside the United States at (317) 572-3993 or fax (317) 572-4002.
Wiley also publishes its books in a variety of electronic formats. Some content that appears in print may not be available in electronic formats. For more information about Wiley products, visit our web site at www.wiley.com.
Library of Congress Cataloging-in-Publication Applied for:
Hardback ISBN: 9781394219216
Cover Design: WileyCover Image: © argus/Shutterstock
About the Editors
Pethuru Raj Chelliah, PhD, is the Chief Architect at the Edge AI division of Reliance Jio Platforms Ltd. (JPL) Bangalore, India. Previously, he worked at the IBM Global Cloud Centre of Excellence (CoE), Wipro Consulting Services (WCS), and Robert Bosch Corporate Research (CR). He has gained over 23 years of IT industry experience and 9 years of research experience. He completed his CSIR-sponsored PhD at Anna University, Chennai, and continued with UGC-sponsored postdoctoral research in the Department of Computer Science and Automation at the Indian Institute of Science (IISc), Bangalore. Following that, he was granted two international research fellowships (JSPS and JST) to work as a research scientist for 3.5 years in two leading Japanese universities.
He focuses on some of the emerging technologies like artificial intelligence (AI) model optimization techniques, Internet of Things use cases, cloud-native and edge-computing paradigms, reliability engineering practices, 6G communication and blockchain technologies, quantum cryptography algorithms, and multimodal generative AI methods.
Amir Masoud Rahmani, PhD, is an artificial intelligence faculty member at the National Yunlin University of Science and Technology, Taiwan. He has published over 240 journal papers (around 180 ISI-indexed papers, many in the best quartile journal, Q1) and more than 100 reference conference papers. His research interests are in the Internet of Things, cloud/fog computing, big data, and artificial intelligence.
Robert Colby is the Principal Engineer in IT Infrastructure, responsible for manufacturing network architecture, and IoT infrastructure at Intel Corporation. Rob joined Intel in 1999, and over the next 10 years drove a variety of network infrastructure and factory security architectures. In 2010, Rob shifted his focus to wireless sensor technologies (RFID, Bluetooth, 802.15.4, ultrasound, etc.) to standardize wireless telemetry and location sensor infrastructure. Over the past six years, Rob has been driving IoT infrastructure standardization, defining how market products land within Intel to solve business problems. His E2E IoT analytic architecture has enabled factories to accurately predict component failures weeks in advance. More recently, Rob has focused on connectivity modernization within Intel Factories and has shifted Intel from using “classic ethernet” to a software-defined networking strategy for industrial edge access. Rob has a long history of mentoring in repeatable innovation methodologies, running programs such as innovation “boot camps” and kick-starter projects. Rob has been an Intel Patent committee member for six years, is currently a voting member of the IoT and Fog Committee, and Rob now holds sixteen US patents.
Gayathri Nagasubramanian, PhD, is an assistant professor in the Department of Computer Science and Engineering at GITAM University, Bengaluru, India. She received her PhD in Information and Communication Engineering from Anna University, Chennai. She received her ME degree in Computer Science and Engineering and her, BTech degree in Information Technology from Thiagarajar College of Engineering, Madurai. She has several publications in journals and has edited many books. She holds 7 Indian patents, 3 Australian patents. Her research interests include Big Data Analytics, IoT, and networks, and she is actively involved as an IEEE WIE faculty advisor.
Sunku Ranganath is the Principal Product Manager for Edge Infrastructure Services, Equinix. He is a product management and solution leader in edge computing, private 5G, and IoT SaaS and PaaS services. He has expertise in defining technical product requirements, cross-functional requirements, product positioning, and developing Go-To and Get-To Market solutions with inhouse and public cloud-based (AWS, GCP) offerings. He has experience driving collaboration with CXOs and leadership teams in OEMs, SaaS/PaaS startups, ISVs, SIs, Telco vendors, and service providers, and establishing open-source alliances. He also has expertise in the architecture, design, integration, and performance analysis of IoT and edge services, edge infrastructure, private 5G, edge security, NFV, and microservices, as well as strong. His strong business acumen enables him to prioritize and map business requirements to solution requirements and software deliverables.
He has experience leading and working with architecture, engineering, marketing, sales, data science teams distributed across Europe, Asia, China, and Mexico, using agile and Scrum methodologies. He leads the solution architecture for building Edge and NFV reference architectures, PoCs, and demos with Kubernetes, Service Mesh, OpenStack, OVS, DPDK, Prometheus, Collectd, Influxdb, Kafka, etc.
He has demonstrated servant leadership as a maintainer and contributor to multiple open-source projects, leading cross-geographical teams of engineers and contract workers toward an established cadence of releases.
He is an avid innovator with multiple patent filings and he has presented at multiple industry and IEEE conferences. He has also authored multiple ETSI standards, industry publications, white papers, and solution briefs.
List of Contributors
Pavan Kumar Akkisetty
Lead AI Architect – LLM Pipelines & Cloud As as a Service Lead AI Architect
Intel Corporation
Hillsboro, Oregon
USA
S. Annamalai
School of Computer Science and Engineering
Jain (Deemed-to-be University)
Bangalore, Karnataka
India
Hanieh Mohammadi Arzanagh
Department of Computer Science and Engineering
Shahid Beheshti University
Tehran
Iran
M. Ashok Kumar
Department of Computer Science and Software Engineering
Skyline University Nigeria
Kano
Nigeria
Sundaravadivazhagan Balasubaramanian
Department of Information Technology
University of Technology and Applied Sciences
Al Mussanah
Sultanate of Oman
Balamurugan Baluswamy
Shiv Nadar University
Delhi-National Capital Region (NCR)
India
Neha Bhati
Department of Research & Development
AVN Innovations
Ajmer, Rajasthan
India
Likitha Chowdary Botta
School of Computer Science and Engineering
VIT-AP University
Inavolu
Andhra Pradesh
India
Madala Guru Brahmam
School of Computer Science Engineering & Information Systems
Vellore Institute of Technology
Vellore, Tamil Nadu
India
Sachin Chaudhary
School of Computer Science and Applications
IIMT University
Meerut, Uttar Pradesh
India
Balakrishnan Chinnaiyan
Department of Computer Science
CHRIST University
Bangalore, Karnataka
India
Avi Das
School of Computer Science and Engineering
VIT-AP University
Inavolu, Andhra Pradesh
India
Priyanka Gupta
Department of CSIT
Guru Ghasidas Central University
Bilaspur, Chhattisgarh
India
K. Hemanth Sai
Department of AI
Anurag University
Hyderabad, Telangana
India
Atefeh Hemmati
Department of Computer Engineering, Science and Research Branch
Islamic Azad University
Tehran
Iran
Anupriya Jain
School of Computer Applications MRIIRS Faridabad
Faridabad, Haryana
India
Mahalakshmi Jeyabalu
Department of Computer Science
CHRIST University
Bangalore, Karnataka
India
Naresh Kumar Kar
GITAM Deemed to be University
Hyderabad, Telangana
India
A. Suresh Kumar
School of Computer Science and Engineering
Jain (Deemed-to-be University)
Bangalore, Karnataka
India
Bhupendra Kumar
School of Computer Science and Applications
IIMT University
Meerut, Uttar Pradesh
India
M. Kumaresan
School of Computer Science and Engineering
Jain (Deemed-to-be University)
Bangalore, Karnataka
India
Seyedeh Yasaman Hosseini Mirmahaleh
Department of Electrical Engineering, Science and Technology
Lille University
Lille
France
A. Mallikarjuna Reddy
Department of AI
Anurag University
Hyderabad, Telangana
India
Aliyu Mohammed
Department of Management
Skyline University Nigeria
Kano
Nigeria
Rakhi Mutha
Department of IT
Amity University Rajasthan
Jaipur, Rajasthan
India
Yamini Nimmagadda
Network and Edge, Intel
Portland, OR
USA
Manoj Kumar Pandey
Department of CSE
Pranveer Singh Institute of Technology
Kanpur, Uttar Pradesh
India
and
Department of CSE
Chandigarh University
Mohali
India
Rajneesh Panwar
School of Computer Science and Applications
IIMT University
Meerut, Uttar Pradesh
India
Jayalakshmi Periyasamy
School of Computer Science Engineering & Information Systems
Vellore Institute of Technology
Vellore, Tamil Nadu
India
Amir Masoud Rahmani
Future Technology Research Center
National Yunlin University of Science and Technology
Douliou
Taiwan
Vijay Anand Rajasekaran
School of Computer Science Engineering & Information Systems
Vellore Institute of Technology
Vellore, Tamil Nadu
India
U. Rakesh
Department of Computational Intelligence
MRCET
Secunderabad
India
Partha Pratim Ray
Department of Computer Applications
Sikkim University
Gangtok, Sikkim
India
Mure Sai Jaideep Reddy
School of Computer Science and Engineering
VIT-AP University
Inavolu, Andhra Pradesh
India
T. Rupa Rani
Department of CSE
Narsimha Reddy Engineering College
Hyderabad, Telangana
India
N. Sangeetha
Department of Artificial Intelligence and Machine Learning
City Engineering College
Bangalore, Karnataka
India
Kewal Krishan Sharma
School of Computer Science and Applications
IIMT University
Meerut, Uttar Pradesh
India
Rattan Sharma
Delhi School of Business
New Delhi, Delhi
India
Seema Sharma
School of Computer Applications MRIIRS Faridabad
Faridabad, Haryana
India
Vikas Sharma
School of Computer Science and Applications
IIMT University
Meerut, Uttar Pradesh
India
M. Sirish Kumar
School of Computing
Mohan Babu University
Tirupati, Andhra Pradesh
India
V. Sivanantham
Department of Computer Science
Periyar University
Salem, Tamil Nadu
India
G. Sridhar Reddy
Department of AI
Anurag University
Hyderabad, Telangana
India
S. Sumanth
Department of Computer Science and Computer Applications
Government College for Women (Affiliated to Bengaluru North University)
Kolar, Karnataka
India
D. Sumathi
School of Computer Science and Engineering
VIT-AP University
Inavolu, Andhra Pradesh
India
G. Sunil Kumar
Department of CSE
Narsimha Reddy Engineering College
Telangana
India
Dyavarashetty Sunitha
Department of CSE
Narsimha Reddy Engineering College
Hyderabad, Telangana
India
Dommaraju Tejavarma
School of Computer Science and Engineering
Jain (Deemed-to-be University)
Bangalore, Karnataka
India
M. Trupthi
Department of AI
Anurag University
Hyderabad, Telangana
India
Gandhodi Harsha Vardhan
School of Computer Science and Engineering
Jain (Deemed-to-be University)
Bangalore, Karnataka
India
Tarun Kumar Vashishth
School of Computer Science and Applications
IIMT University
Meerut, Uttar Pradesh
India
G. Victor Daniel
Department of AI
Anurag University
Hyderabad, Telangana
India
Narayan Vyas
Department of Computer Science and Application
Vivekananda Global University
Jaipur, Rajasthan
India
Gayathry S. Warrier
Department of Computer Science
CHRIST University
Bangalore, Karnataka
India
Pawan Whig
Vivekananda Institute of Professional Studies-TC
New Delhi, Delhi, India
Nikhitha Yathiraju
University of The Cumberlands
Williamsburg, KY
USA
1Fundamentals of Edge AI and Federated Learning
Atefeh Hemmati1, Hanieh Mohammadi Arzanagh2, and Amir Masoud Rahmani3
1Department of Computer Engineering, Science and Research Branch, Islamic Azad University, Tehran, Iran
2Department of Computer Science and Engineering, Shahid Beheshti University, Tehran, Iran
3Future Technology Research Center, National Yunlin University of Science and Technology, Douliou, Taiwan
1.1 Introduction
Machine Learning (ML) and Artificial Intelligence (AI) have advanced unparalleled in recent years. These developments have changed many industries, and interactions between people and technology have experienced significant changes. Data processing in cloud environments was a primary focus of traditional AI and ML techniques. The need for real-time processing has expanded due to the growth of the Internet of Things (IoT). Because of this, a new era of technology known as Edge AI and Federated Learning (FL) has appeared [1, 2].
Edge AI gives edge devices and servers the ability to assess data and carry out AI operations close to the data generated. Autonomous vehicles, remote monitoring, and customized user experiences are some of the applications of this technology that stand out. Because data is not centralized and is handled on local devices, edge AI can help with security and privacy issues [3–5].
At the same time, FL has become a crucial strategy for protecting data privacy and distributing the work of developing ML models among several devices. Devices train ML models using local data using this technique, sharing only the model’s parameters with a centralized server. This ensures that consumers’ privacy is protected and that sensitive data remains on the devices [2, 4].
With the assistance of edge AI and FL, we can create applications and systems that are smarter and safer. This is a new claim in the field of AI. With this novel approach, we can utilize the sophisticated capabilities of AI in a connected world while enhancing privacy, security, and efficiency [2, 6, 7].
The concepts and foundations of Edge AI and FL are entirely clarified in this chapter. This chapter also examines the advantages and challenges of Edge AI and FL. We explore the benefits, which include decreased latency, better bandwidth consumption, improved privacy, and increased robustness. We also present the challenges, such as resource limitations on edge devices, communication costs, and the requirement to manage uniformly distributed data distributions in FL. We also discuss real-world applications where Edge AI and FL have proven effective, opening the door for revolutionary solutions across numerous industries. We also discuss FL and Edge AI’s challenges, future research directions, and open issues.
The remaining part of the chapter is organized as follows: In Section 1.2, the fundamental concepts of Edge AI are discussed, including the advantages of Edge AI and the challenges faced by Edge AI. Section 1.3 introduces the concepts and fundamentals of FL, including the advantages of FL and the challenges associated with FL. In Section 1.4, the chapter emphasizes the combined power of FL and Edge AI, highlighting the benefits of their integration. Section 1.5 offers background insights into the technological landscape and motivations behind exploring Edge AI and FL. Applications of Edge AI and FL are discussed in Section 1.6. Section 1.7 addresses challenges, future research directions, and potential solutions for integrating Edge AI and FL. Finally, Section 1.8 concludes the chapter.
1.2 Concepts and Fundamentals of Edge AI
In an era of data generation and IoT, Edge AI is emerging as a revolutionary paradigm that brings AI computing closer to the edge of the network, where data is generated. Unlike traditional AI architectures that rely on cloud servers for processing, Edge AI distributes computing tasks to local edge devices such as smartphones, smart sensors, and edge servers. This change in the deployment of AI brings significant benefits and exciting possibilities in various fields [3, 8].
1.2.1 Defining Edge AI
Figure 1.1 shows that Edge AI directly integrates AI and ML capabilities on edge devices. This allows devices to perform data processing and inference locally, reducing the need to transfer data to centralized cloud servers constantly. This decentralized operation reduces latency in terms of time and avoids dependence on a stable Internet connection [9, 10].
Figure 1.1 Edge AI architecture.
Edge devices play a vital role in the edge AI ecosystem. These devices are typically equipped with limited resources compared to powerful cloud servers but with sufficient computing capabilities to run lightweight AI models. In other words, these devices can perform AI operations with their limited resources [11]. There are many examples of edge devices, including smartphones, tablets, smart cameras, wearables, unmanned aerial vehicles (UAVs), and IoT sensors [12, 13]. These devices can collect, process, and interpret data without frequent communication with central servers and are ideally suited for situations where network bandwidth is limited, or data privacy is required. With these edge devices, the possibility of analyzing and using data locally increases to a certain extent, which leads to reduced response delays and improved user experience [5, 14, 15].
Edge AI has improved the process of directly deploying AI models to edge devices. This type of deployment results in significant latency reduction and increased privacy due to its local processing capabilities. Several different architectures have been used to implement edge AI, each of which covers its own specific needs and applications [16]:
Local Processing in IoT Devices
: IoT devices mostly have limited computing resources, which makes direct deployment of complex AI models challenging. This type of architecture is suitable for applications that require fast processing, such as sensor data analysis, anomaly detection, and simple classification tasks [
1
,
17
,
18
]. In this instance, a unique local processing technique suggested by Bebortta et al.
[19]
supports an enhanced IoT platform structure for smart buildings. The proposed approach helps decrease bandwidth at the nodes’ data collection level from a green computing standpoint. The researchers also discussed the most effective usage of sensors in wireless sensor networks (WSNs), who used the well-known queue model to calculate costs associated with non-Poisson and Poisson arrival of data packets at local processors. The experiments show that the suggested model successfully utilizes green computing standards. Therefore, this study offers a thing-centric, data-centric, and service-oriented IoT architecture within the framework.
Edge Servers and Gateways
: Edge servers and gateways are more powerful computing devices closer to the edge devices. They act as intermediaries between edge devices and cloud servers, performing initial AI processing before sending data to cloud servers for further analysis. These devices can accommodate more resource-intensive models and are suitable for applications such as video analytics, natural language processing (NLP), and data preprocessing [
2
,
20
]. As a real-world example, Rahmani et al.
[21]
utilized the ideal position of these gateways at the network’s edge to provide several more advanced solutions, including storage locally, real-time local data processing, and integrated data mining, for giving an intelligent electronic health care gateway in the procedure. They then suggested creating a geo-distributed intermediate layer of awareness between sensor nodes and the cloud to use fog computing for medical IoT devices. Their fog-assisted design might handle obstacles in omnipresent medical facilities, such as mobility, energy consumption, flexibility, and dependability issues, by focusing on part of the duties of the sensor network and a distant medical center. Also, we can mention that Li et al.
[22]
suggested edge content-centric networking (ECCN). This enabling strategy combined Software-Defined Networking (SDN) with content-centric networking into a structured framework. To separate the data and control planes of ECC and CCN, SDN technology was included in the hierarchical framework. To manage data transmission, an SDN framework was created. To assess the effectiveness of the ECCN system, two apps were also deployed in the testbed. Thorough computations and results of experiments from the experimentation applications show that ECCN surpassed the original structures.
Edge Computing
: The concept of edge computing expands the conceptual scope of edge computing by introducing a hierarchical architecture that spans from the edge to the cloud. In this setup, middle fog nodes are responsible for AI processing and decision-making, minimizing latency and network traffic. These nodes may be located in cellular base stations, access points, and local data centers. Edge computing is particularly suitable for applications that require real-time analytics in distributed environments such as smart cities and industrial IoT
[23]
. Chen et al.
[24]
addressed a distributed computing model called Mobile Edge Computing (MEC), which addresses energy and low latency requirements by integrating energy absorption technology into IoT devices. The research aimed to decrease system costs by proposing a hybrid power supply model for IoT devices and concurrently optimizing local computing, task transfer time, and edge computing decisions. Utilizing stochastic optimization theory, the study introduced an online dynamic algorithm named DTOME for task transfer in MEC with a hybrid energy supply. This algorithm facilitated task transfer decisions by balancing the system cost and queue stability. Simulation results affirmed the effectiveness of the DTOME approach, demonstrating significant improvements in reducing system costs compared to other comparative methods. Furthermore, Wu et al.
[25]
explored a multiuser MEC system within the context of the Internet of Vehicles, designed explicitly for handling computing tasks near vehicles. Compared to previous research that primarily focused on minimizing task transfer costs assuming complete channel estimation, this study addressed the issue of incomplete channel estimation caused by the dynamic movement of vehicles, which had been overlooked in prior work. The central objective was to decrease computation, communication delays, and energy consumption while considering the challenges of incomplete channel estimation. The study initially introduced a system cost metric that combined delay and energy consumption to achieve this goal. The optimization problem was then formulated to minimize this cost. The research further proposed an innovative approach combining deep reinforcement learning with Lagrange coefficients to reduce the overall system cost concurrently. Simulation results demonstrated the superiority of this approach over traditional methods in terms of performance.
Edge Multilayer Architecture
: This architecture uses a multilayer approach with different levels of processing at other edge nodes. Low-level edge devices perform basic data preprocessing, feature extraction, and analysis tasks. Mid-level edge devices handle more complex tasks, including AI and model inference. Finally, the next level could be a local data center or cloud server that performs advanced analytics, model training, and long-term storage. This architecture optimizes resources and hides latency by distributing tasks at different levels [
14
,
26
]. Robles-Enciso and Skarmeta
[27]
proposed an innovative multilayer enhancement to reinforcement learning (ML-RL) methods that enabled edge agents to reach an upper-level agent with supplementary information to improve efficiency in challenging and unpredictable scenarios. They initially proposed an RL approach to tackling the Task Assignment Problem (TAP) at the edge layer. Before task allocation, they assessed the potential balance between energy consumption and execution time. The scalability of each technique was subsequently validated through multiple simulations involving varying device counts.
Edge-Distributed AI
: In this architecture, several edge devices work together to process and analyze data, and each one contributes to AI by executing a part of the model or performing specific tasks. Communication and device coordination are essential for efficient model integration and optimal decision-making. Distributed edge AI is particularly suitable for rapid data processing scenarios, such as autonomous vehicles and robotics [
4
,
28
]. In this order, Mwase et al.
[29]
described the cloud-to-thing continuum and offered a framework to make AI possible in totally edge-based applications. They also provided ways to deal with the interaction difficulties brought on by totally edge-based situations’ scattered nature. These methodologies’ performance enhancements demonstrated in cutting-edge research were provided, along with suggestions for future research approaches. The information was offered to encourage comprehension and, as a result, the involvement of diverse researchers in tackling this difficulty. Moustafa
[30]
provided an innovative IoT experimentation design that could be utilized to assess safety features based on AI. To provide dynamic experimentation networks that enabled communication across edge, fog, and cloud layers, the platform named NSX vCloud NFV was used to ease the execution of Software-Defined Networks (SDNs), Network Function Virtualization (NFV), and Service Orchestration (SO). Real-world routine and assault scenarios were run while the architecture was deployed to gather datasets with labels.
Hybrid Cloud-Edge Architectures
: Hybrid cloud-edge architectures combine the advantages of cloud and edge computing and use AI models on both cloud and edge devices. This approach depends on access to computing resources and data and allows for flexible processing based on different conditions. This architecture suits applications that balance real-time processing and more detailed analysis. For example, real-time monitoring may be done at the edge, while complex analysis of historical data is done in the cloud
[31]
. Some researchers use this architecture: Celesti et al.
[32]
focused on the problems that needed to be solved for the Cloud-to-Edge platform to provide privacy, reliability, genuineness, and nonrepudiation. Additionally, they examined a real-world case study while considering the Message-Oriented Middleware (MOM) architectural model. The overall efficiencies of the entire platform were unaffected by the additional safety capabilities, according to the experiments on a genuine testbed. Also, Lei et al.
[33]
considered that geographically dispersed edge servers had performance that varied over time and presented a dynamic offloading approach built on a framework of probabilistic evolutionary games. Researchers undertook practical case studies based on a real-world dataset of cloud-edge resource placements to evaluate their suggested framework. The results demonstrated that their proposed method outperformed conventional ones in terms of several measures.
Choosing the right Edge AI architecture depends on factors such as computing resources [
2
,
15
,
17
], latency constraints
[28]
, communication bandwidth [
14
,
15
], and application requirements
[34]
. As AI at the edge continues to evolve, various architectures have been introduced that enable AI models to be deployed closer to data sources. These choices enable improved performance, reduced network load, increased privacy, and the realization of scheduled AI applications across environments.
1.2.2 Advantages of Edge AI
Edge AI offers several advantages that will change AI. These advantages derive from the basic idea of processing data locally on edge devices and result in several benefits [35, 36]:
Reduction of Latency
: Edge AI greatly decreases the time it takes to transport data to centralized servers by processing data locally on edge devices. This low-latency operation is vital for real-time applications like self-driving cars, video analytics, and augmented reality, where quick answers are required for a smooth user experience and safety. Edge AI, for example, analyses ambient data in augmented reality glasses to send timely digital information to the user’s vision, resulting in a smooth and responsive augmented experience
[28]
.
Advancing in Privacy and Security
: Edge’s AI solves privacy issues by decreasing the need for data sent to other servers. Personal data is kept on-site, lowering the risk of data breaches and illegal access. This benefit is especially essential when sensitive or secret data must be processed. Home security cameras, for example, may identify possible risks such as unlawful access or suspicious activity in the area using Edge AI, eliminating the need to transfer recordings containing sensitive information to external servers [
37
,
38
].
Bandwidth Efficiency
: Edge AI minimizes network bandwidth by processing data on edge devices, reducing the data transferred to the cloud for analysis. This is especially useful in areas with restricted connections or high data transmission costs. Field sensors collect information on soil moisture, temperature, and crop health in agriculture. Edge AI analyses this data locally to offer farmers precise irrigation and fertilization suggestions, conserving resources and lowering data transfer costs [
14
,
15
,
34
].
Network Reliability
: Edge AI apps can function even if the Internet connection is unreliable or interrupted. Edge devices rely less on continual Internet connectivity since they can process data locally, ensuring ongoing performance and flexibility. Edge AI, for example, allows continuous monitoring and control of remote offshore wind farms by processing data locally on the turbines. It preserves dependability in the face of unreliable networks, optimizes data transfer, and boosts overall efficiency by recognizing abnormalities and taking prompt steps
[39]
.
Real-Time Decision-Making
: Edge AI allows for immediate decision-making without needing external servers. This is especially useful for applications requiring immediate reactions, such as predictive maintenance in industrial settings or healthcare monitoring. Edge AI, for example, analyses live cameras at junctions in an intelligent traffic management system to detect and respond to traffic infractions or emergencies in real time. This technology improves traffic flow efficiency and safety [
17
,
28
].
Cost Efficiency
: Edge AI can help cut the expenses of transporting and storing massive volumes of data on cloud servers. Organizations may optimize cloud utilization and decrease operating costs by processing data locally. Retail establishments, for example, employ Edge AI to manage real-time inventory, monitor stock inventories, and identify customer patterns. This gadget works locally, helps with refilling operations, and optimizes shelf space
[26]
.
Scalability and Distributed Processing
: Edge AI enables distributed processing across several edge devices, allowing more efficient data-intensive operations administration. This scalability is functional when centralized processing of vast data is impossible. Smart grids, for example, utilize edge AI to monitor and optimize energy use across an extensive network of connected devices. This technology enables more effective energy distribution across devices while avoiding overwhelming a centralized server, ensuring energy is transferred optimally and without needless stress
[2]
.
1.2.3 Challenges of Edge AI
Although edge AI offers significant benefits, it also faces challenges that need to be carefully addressed:
Limitations and Resource Optimization
: Edge devices frequently have limits owing to computer power, memory, and battery life. AI models must be carefully optimized to balance efficiency and effectiveness and assure more optimal resources throughout day-to-day operations [
11
,
40
].
Complex Model Management
: Managing sophisticated models on edge devices presents significant hurdles. To preserve the AI ecosystem’s standards and accuracy, synchronous model updating, version control, and model change adaptation are required
[15]
.
Navigation Heterogeneity
: The increasing number of edge devices with varying hardware features, operating systems, and communication protocols have resulted in navigation heterogeneity. Imposing coherence and standards over this broad span creates unity, integration, and compatibility [
2
,
17
].
Edge-Cloud Balancing
: The ability to balance edge-driven AI with cloud computing is essential. Because not all jobs can be completed on edge devices, accurately determining which tasks should be performed on edge devices and which should be offloaded to cloud processing is critical. This complicated balance demands accurate task understanding and effective computing resource allocation while considering the dynamics of environment traversal and constant changes [
7
,
20
].
Consequently, by utilizing edge devices’ capabilities, edge AI will radically alter AI deployment and bring intelligence closer to the data source. The benefits provided by this technology in terms of latency reduction, privacy protection, and bandwidth efficiency make it a promising invention with broad applications in areas such as healthcare, manufacturing, transportation, and smart cities. It is critical to solve problems and optimize AI models.
1.2.4 Taxonomy of Edge AI Applications
Edge AI applications have affected various fields. In the context of IoT, smart devices have taken center stage, encompassing wearables like smartwatches, fitness trackers, and health monitors that empower individuals to monitor their well-being seamlessly. Home automation, featuring smart thermostats, cameras, and appliances, fosters a more connected and convenient lifestyle. Industrial sensors cater to manufacturing monitoring, enhancing operational oversight. The automotive industry leverages edge AI for self-driving cars and advanced driver assistance systems, ensuring safer and smarter mobility. Healthcare benefits from remote patient monitoring, medical imaging interpretation (MRIs, X-rays, CT scans), and personalized medicine. Agriculture embraces precision farming and livestock management for optimized resource allocation.
Computer vision thrives with applications like object detection and image classification. NLP drives speech recognition for voice commands and language translation, while sentiment analysis and named entity recognition refine textual insights.
Anomaly detection and predictive maintenance enhance sectors such as industrial equipment monitoring, infrastructure health assessment, healthcare tracking, energy management, and supply chain optimization. As edge AI continues to evolve, its transformative impact on diverse domains becomes increasingly evident. Figure 1.2 provides a taxonomy of Edge AI applications.
1.3 Concepts and Fundamentals of FL
FL networks have developed as a technique that enables collaborative learning in ML without requiring centralized data aggregation in an era where data privacy and security have become vital objectives. This decentralized learning paradigm enables numerous edge devices to train a shared model jointly while storing and keeping their data locally. FL offers a novel approach to utilizing the collective intelligence of dispersed devices while maintaining data privacy and control [11, 16, 41].
1.3.1 Defining FL
FL, or Federated ML, is a decentralized ML approach that allows edge devices to collectively train a shared model while concurrently keeping their data locally, as seen in Figure 1.3. Unlike classical ML, which collects and processes data on a centralized cloud server, FL models are trained directly on edge devices. Instead of delivering raw data, model updates are also sent to a central server or peer devices [41, 42].
Figure 1.2 Taxonomy of edge AI applications.
The key steps involved in the FL process are as follows [16, 43]:
Initialization
: An initial global model is generated randomly or by pretrained weights on a central server.
Figure 1.3 FL architecture.
Distribution
: The initial global model is delivered to participating edge devices, forming a dispersed network of devices.
Local Training
: The global model is trained locally at each edge device without exchanging raw data with the central server or other devices.
Model Update
: Following local training, each edge device provides a model update incorporating training-related modifications into the global model.
Aggregation
: All device model changes are aggregated to a central server using various approaches, such as averaging or weighted averaging.
Global Model Update
: The central server aggregates the bulk model updates to provide a more accurate global model.
Iteration
: This procedure is done numerous times, allowing edge devices to build the global model collectively utilizing their different datasets.
1.3.2 Advantages of FL
FL offers a range of advantages, including [41]:
Data Privacy and Security
: FL solves privacy concerns since raw data is never sent beyond the edge devices, lowering the danger of exposing sensitive data. This method keeps data decentralized and confidential, enhancing confidence and compliance with data protection requirements. In a medical research context, for example, FL allows hospitals to collaboratively train a model to forecast illnesses without revealing individual patient data, allowing them to comply with privacy requirements [
37
,
42
].
Decentralization
: FL utilizes the computational capacity of several devices by spreading the training process among edge devices, making it scalable and appropriate for big data sets. FL enables multiple urban sensors in a smart city setting to train a model for optimal traffic flow. The approach is effortlessly scalable to suit the intricacies of an extensive and complex metropolitan environment by utilizing the processing capacity of these dispersed devices [
5
,
16
].
Low Communication Cost
: FL minimizes communication costs over standard centralized ML systems requiring raw data delivered to a central server since only model changes are communicated. Consider a retail chain that employs FL to enhance customer referrals. Instead of sending vast volumes of client purchase data to a central server, only refined model updates are communicated, lowering communication costs while ensuring data privacy
[26]
.
Adaptation to Heterogeneous Data
: FL is appropriate for instances where edge devices have heterogeneous data distributions, allowing the model to be trained on various data sources. FL allows several fields with diverse soil and climate variables to prepare a crop yield prediction model in agriculture jointly. This enables the model to account for various data distributions, resulting in more accurate and adaptable predictions [
2
,
44
].
Disturbance Resistance
: FL is more resistant to network disruptions and connectivity interruptions due to model training on edge devices. This capability is beneficial in unstable circumstances or with intermittent Internet connections. FL’s toughness shows through in industrial automation. FL is used in manufacturing facilities to train quality control models on edge devices, ensuring that production is not hampered by network outages, which is crucial for maintaining constant product quality
[43]
.
Local Improvement
: Each edge device can improve locally by training the global model on its device. This enables network operators to update their models depending on their needs, resulting in a more ideal user experience. FL lets each cell tower in a telecommunications network enhance its coverage prediction model locally. This personalization improves network performance in specific geographic areas, assuring continuous connectivity and customer pleasure
[17]
.
Model Flexibility
: FL enables network operators to construct customized models depending on the demands and interactions of edge devices. This allows the models to adjust and respond to real-world changes dynamically. FL reacts to changing situations, such as environmental monitoring. Weather stations work together to train a model that predicts storms, update and fine-tune parameters based on real-time data from numerous sensors, and produce accurate and fast weather predictions
[34]
.
Reduction of Training Time
: Because of local data usage in edge devices, FL can cut model training time. This enables operators to train more up-to-date models in less time and enhance performance more quickly. FL speeds fraud detection model changes in the banking industry. Banks collaborate to train the model utilizing data from dispersed branches, allowing faster adaptability to developing fraud trends and reducing possible losses [
4
,
43
].
1.3.3 Challenges of FL
Despite the FL process’s vast commitment, it also presents problems that must be carefully considered:
Communication Efficiency
: Ensuring efficient communication between edge devices and the central server is critical. Managing possible communication delays and faults is essential for maintaining optimal FL performance. For example, if a cellular network with restricted capacity is used for FL, the communication overhead may be significant, resulting in delays and inefficiencies [
14
,
23
].
Privacy and Security
: Protecting user privacy is a significant priority in FL. Sharing model updates may expose sensitive information about individual participants’ data. For example, while designing a tailored recommendation system, user activity data from numerous devices may be collected. Updates to the sharing mechanism may mistakenly reveal users’ preferences and behaviors
[4]
.
Heterogeneity of Models
: Edge devices’ computational capabilities necessitate adaptive model architectures and adaptive learning rates for various hardware restrictions.
Quality and Accuracy of Models
: Because of the decentralized structure and diversity of data, model changes may result in a decline in model quality and accuracy. To increase the accuracy of models in FL, quality standards and techniques must be established.
Using Local Resources
: The emphasis on local training at the edge devices may result in insufficient use of aggregated data across the FL network. Balancing local training and global model updates is complex and requires careful consideration.
Management Complexity
: Managing numerous devices, executing updates, and balancing local training and update collection can be complex and challenging in a more extensive FL network.
Optimization of Energy Consumption
: The most efficient use of energy resources in edge devices for training models and collecting updates necessitates consideration of the energy elements of this process.
1.3.4 Taxonomy of FL Applications
FL is an innovative form of ML that uses several data sources while protecting the privacy and security of that data. Device-centric and server-centric data sources can be divided into two categories. Data originating from mobile devices such as smartphones, tablets, and IoT gadgets; wearables such as smartwatches and fitness trackers; and edge devices such as cameras, sensors, and edge servers, are all included under the device-centric category. Data centers, including centralized data centers, cloud servers, distributed servers, and edge servers, are the data sources for the server-centric type.
Many different industries can benefit from this method, including healthcare and medical for clinical data analysis, medical imaging diagnosis, and drug discovery; finance and banking for fraud detection, credit scoring, and risk assessment; smart cities and IoT for traffic management, energy consumption prediction, and environmental monitoring; manufacturing and industrial for quality control, predictive maintenance, and process optimization; and retail and e-commerce.
With a focus on methods like secure aggregation, Differential Privacy, and Homomorphic Encryption, the methodology also emphasizes security and privacy. It navigates ethical issues like bias reduction, fairness, responsibility, and openness to ensure responsible and effective implementation. Figure 1.4 presents an FL taxonomy.
1.4 Combining FL and Edge AI
The combination of FL and Edge AI technologies is a novel and sophisticated strategy in the field of ML and AI that dramatically increases the performance, security, and privacy of AI devices and IoT. ML models are trained on local devices and regularly synced to each other utilizing FL approaches in this combination. The local models are updated, and their training results are communicated to the central hub at the end of each cycle. Model combination approaches are then used to turn local models into a new combined model with improved accuracy and predictive ability [2, 5, 7]. These two approaches offer complementary benefits that can help achieve unique real-world benefits:
Using Local Data for Global Insight
: The use of federated training approaches is one solution for ensuring privacy and maximizing the usage of AI models in networks of dispersed edge devices. This method enables AI models to be trained without relying on raw data. This method defines the models’ concentration on data processing in the immediate vicinity or at the data source. This improves data privacy while also addressing privacy concerns and legal mandates. Edge AI, conversely, is concerned with processing data at or near the place to reduce latency and enable real-time decision-making. Organizations may use the large quantity of local data accessible on edge devices by integrating FL with edge AI. This information may be utilized to train and develop appropriate AI models for each device and environment. This method protects sensitive local information and allows devices to make intelligent real-time judgments based on local data. This combination provides for developing customized, adaptable, and efficient AI systems [
45
,
46
].
Figure 1.4 Taxonomy of FL applications.
Collaborative Learning Across Edge Devices
: FL focuses on teaching collaborative paradigms across various edge devices. A global and representative representation of the whole network is produced by spreading data amongst devices. This combined feature complements Edge’s AI. Devices in various areas generate local knowledge, improving the system’s accuracy and dependability. In a smart city, for example, edge devices of multiple regions may train models for traffic control, energy optimization, and public safety while sharing their local expertise. This cooperative learning strategy enhances model accuracy and generalization while making them more adaptable to dynamic situations
[47]
.
Reducing Communication Costs
: Managing the communication burden is one of the issues in FL and edge AI, especially in resource-constrained contexts. Federated training avoids sending raw data to servers and keeps models updated as little as possible during aggregation. On the other hand, Edge AI is concerned with locally processing data to lessen dependency on cloud communications. When these two strategies are combined, data sharing becomes more efficient. Edge devices process initial input and train models locally, transmitting only necessary updates to the central server. This method considerably minimizes the communication burden, saves bandwidth, and speeds up the model update process. AI systems gain from faster convergence, lower latency, and edge resource optimization in this manner [
26
,
40
].
Enhanced Privacy and Security
