EDGE INTELLIGENCE : deep learning-enabled edge computing

PRELIMS.pdf Acknowledgements Author biography Shajulin Benedict CH001.pdf Chapter Edge intelligence 1.1 Edge computing 1.1.1 Objectives of edge computing 1.2 History of edge computing 1.3 Edge intelligence 1.3.1 Machine learning-based edge intelligence 1.3.2 Deep learning-based edge intelligence 1.3.3 Hype on edge intelligence 1.4 Advantages: edge intelligence 1.4.1 Low latency 1.4.2 Enhanced security 1.4.3 Bandwidth efficiency 1.5 Challenges: edge intelligence 1.6 Applications 1.6.1 Retail sector 1.6.2 Agriculture sector 1.6.3 Healthcare sector 1.6.4 Education sector 1.6.5 Industrial sector 1.6.6 Finance sector 1.6.7 Smart city sector 1.6.8 Transportation sector 1.6.9 Gaming sector 1.6.10 Personal sector 1.7 The need for this book 1.8 Potential readers 1.9 Organization of the book References CH002.pdf Chapter Edge computing architectures 2.1 Internet of Things 2.1.1 Subdomains of the IoT 2.1.2 Service-level protocols 2.1.3 Communication-level protocols 2.2 IoT-enabled components 2.2.1 Sensors 2.2.2 Edge devices 2.2.3 Fog nodes 2.2.4 Cloud nodes 2.3 Learning techniques: machine learning 2.3.1 Machine learning: the history 2.3.2 Algorithmic views 2.4 Learning techniques: deep learning 2.5 Edge-AI architectures 2.5.1 Generic architecture 2.5.2 Layered architecture 2.5.3 Service oriented architecture 2.5.4 Interoperability architecture 2.5.5 Metric-oriented edge architecture 2.5.6 Open/closed architecture 2.5.7 Large/micro/macro-level architecture 2.5.8 Security-specific architecture 2.6 Conclusion References CH003.pdf Chapter Edge OS and programming models 3.1 Operating systems 3.1.1 Hardware components 3.1.2 Functions of edge-based operating systems 3.2 Objectives of OS in edge devices (Edge OS) 3.3 Taxonomy of OS: toward edge 3.4 Edge OS: examples 3.4.1 QNX neutrino OS 3.4.2 FreeRTOS 3.4.3 Integrity OS 3.4.4 ThreadX 3.4.5 PikeOS 3.4.6 Chibi OS 3.4.7 Raspbian OS 3.4.8 TinyOS 3.4.9 Zephyr OS 3.4.10 Contiki-NG OS 3.4.11 mbed OS 3.4.12 eCoS OS 3.5 Process states in edge-OS 3.6 FreeRTOS: OS for embedded devices 3.6.1 FreeRTOS: ESP 32 3.6.2 Parallel tasks 3.7 Conclusion References CH004.pdf Chapter Edge intelligence: learning techniques 4.1 Edge intelligence: a need 4.2 Federated learning 4.2.1 Federated learning architecture 4.2.2 Federated learning approaches 4.2.3 Federated learning challenges 4.3 DNN splitting 4.3.1 DNN splitting: architectural concepts 4.3.2 Types of split computing 4.3.3 Dynamic versus static 4.3.4 Advantages of split computing 4.3.5 Split computing: tools 4.4 Transfer learning 4.4.1 Traditional view 4.4.2 Transfer learning: pictorial representation 4.4.3 Transfer learning: stages 4.4.4 Transfer learning: types 4.4.5 Transfer learning: TensorFlow implementation 4.4.6 Edge-enabled transfer learning 4.5 Gossip learning 4.6 Conclusion References CH005.pdf Chapter Inference/prediction techniques 5.1 Learning stages 5.2 Inference knowledge levels 5.3 Distributed inferences 5.3.1 Cooperative inference technique 5.3.2 Collaborative inference technique 5.3.3 Real-time inference technique 5.3.4 Metric-oriented distributed inferences 5.3.5 Scheduling resources/tasks 5.4 Distributed inferences: implementation strategies 5.4.1 Lightweight implementations 5.4.2 FPGA-based implementation 5.4.3 Client–server implementation 5.5 Interactive versus batched inferences 5.6 Partitioning distributed inferences 5.6.1 DNN inference partitioning 5.6.2 Real-time partitioning 5.6.3 Horizontal partitioning 5.6.4 Vertical partitioning 5.6.5 Stochastic partitioning 5.7 Accelerating distributed inferences: strategies 5.7.1 Hardware and software levels 5.7.2 Knowledge distillation 5.7.3 Model parallelism approach 5.7.4 Adaptive inference approach 5.7.5 Model compression 5.7.6 Caching inferences 5.7.7 Offloading inferences 5.8 Conclusion References CH006.pdf Chapter Edge resources and accelerators 6.1 Edge resources: basics 6.1.1 Microprocessors 6.1.2 Microcontrollers 6.1.3 Differences 6.2 Micro-controller-level devices: an edge? 6.2.1 An edge? 6.2.2 Importance of Arduino 6.2.3 Arduino families 6.2.4 Arduino UNO board 6.2.5 Arduino programming 6.3 Micro-processor-level devices: general purpose 6.3.1 Raspberry Pi: an example 6.4 GPUs, TPUs, and FPGAs: special purpose 6.5 SoC, SoM, system-on-board 6.6 Edge accelerators 6.6.1 Low-power 6.6.2 Compute-intensive 6.6.3 Memory-intensive 6.6.4 Improved bandwidth 6.6.5 Extensible AI framework 6.6.6 Real-time assistance 6.7 Commercial edge accelerators 6.7.1 Intel-based edge accelerators 6.7.2 NVIDIA-based edge accelerators 6.7.3 Samsung-based accelerators 6.8 Examples and use-cases 6.8.1 Accelerators for CNNs 6.8.2 Accelerators for audio 6.9 Conclusion References CH007.pdf Chapter Performance analysis of edge-enabled applications 7.1 Performance concerns 7.2 Model-specific performance concerns 7.3 Architecture-specific performance concerns 7.3.1 Performance concerns: hardware-dependent 7.3.2 Performance concerns: software-dependent 7.3.3 Performance concerns: integration-related 7.4 Algorithm-specific performance concerns 7.4.1 Compiler-specific refinements 7.4.2 Tuning applications 7.4.3 Selective packages 7.4.4 Programming models 7.5 Data-specific performance concerns 7.6 Performance monitoring: a need 7.7 Performance monitoring: metrics 7.8 Energy-efficiency methods 7.8.1 Energy measurements 7.8.2 Energy efficiency methods 7.9 Carbon efficiency methods 7.10 Workload scheduling and performance impacts 7.11 Performance monitoring tools 7.11.1 Tracing approach 7.11.2 Sandboxing approach 7.11.3 Modeling approach 7.11.4 Simulation approach 7.11.5 Online approach 7.12 Cloud/fog/edge-level performance monitoring 7.13 Conclusion References CH008.pdf Chapter Security in edge-AI systems 8.1 Existing security challenges 8.2 Security attacks in edge-AI 8.2.1 Physical attacks 8.2.2 DDoS attacks 8.2.3 Side channel attack 8.2.4 Malware-injection attack 8.2.5 Authentication/authorization attack 8.2.6 Jamming attack 8.2.7 Forgery attack 8.2.8 Passive security attacks 8.3 Data-specific vulnerabilities 8.3.1 Data integrity-related issues 8.3.2 Data confidentiality-related issues 8.3.3 Data availability-related issues 8.3.4 Configuration mismatches 8.4 Security architectures in edge-AI 8.4.1 Security architecture: a layered perspective 8.4.2 Security architectures: user-centric 8.4.3 Security architectures: device-centric 8.4.4 Security architectures: application-specific 8.4.5 Deployment-specific security architectures 8.5 Preventing security breaches: strategies 8.5.1 Protocol-specific assistance 8.5.2 Reliable routing protocols 8.5.3 Combination of lightweight and computationally-intensive implementations 8.5.4 Firmware-level updates 8.5.5 Prior security testing 8.5.6 Encryption approaches 8.5.7 Newer technologies 8.6 Tools and solutions 8.6.1 FireEye 8.6.2 PaloAlto networks 8.6.3 Symantec solutions 8.6.4 Security: federated systems 8.7 Conclusion References CH009.pdf Chapter Frameworks: edge-AI platforms 9.1 Essential characteristics 9.2 Types of framework 9.3 Resource-allocation frameworks 9.3.1 Real-time resource allocation framework 9.3.2 Cooperative resource allocation framework 9.3.3 Collaborative resource allocation framework 9.3.4 QoS-aware resource allocation framework 9.3.5 Optimal resource allocation framework 9.3.6 End-to-end resource allocation framework 9.4 Cloud-specific frameworks 9.4.1 Microsoft Azure-based frameworks 9.4.2 Google-based edge-AI frameworks 9.4.3 Amazon edge 9.5 Application-specific frameworks 9.5.1 Edge-AI in agriculture 9.6 Distributed federated learning frameworks 9.6.1 Federated learning 9.6.2 SplitFed 9.6.3 FedAdapt 9.6.4 FedLesScan 9.6.5 FATE framework 9.6.6 OpenFL framework 9.6.7 Tensorflow federated 9.6.8 FLAME tool 9.7 Conclusion References CH010.pdf Chapter Orchestration platforms: computing continuum 10.1 Orchestration and integration 10.1.1 Characteristics 10.1.2 Approaches 10.2 Algorithmic/application orchestration 10.3 Workload orchestration 10.4 Hierarchical versus non-hierarchical orchestration 10.5 Adaptiveness in orchestration 10.6 Automation in orchestration 10.7 Metric-oriented orchestration 10.8 Orchestration frameworks 10.8.1 Oakestra framework 10.8.2 ORCH framework 10.8.3 Node-RED framework 10.8.4 KubeEdge framework 10.8.5 Tiny-MLOps framework 10.8.6 Pipeline framework 10.9 Integration platforms 10.10 Conclusion References CH011.pdf Chapter Edge-AI applications 11.1 Applications 11.2 Edge-AI for healthcare 11.2.1 Remote healthcare monitoring and prediction 11.2.2 Personalized medicine 11.2.3 Stress release systems 11.3 Industrial applications using edge-AI 11.3.1 Predictive maintenance 11.3.2 Digital twin 11.3.3 Connected factories 11.3.4 Smart tools 11.3.5 Optimized productivity 11.3.6 Interactive production 11.4 Edge-AI for agriculture 11.4.1 Vegetation-related agriculture 11.4.2 Animal-related farming 11.4.3 Use-case: elephant emotion detection framework using deep learning 11.5 Edge-AI for forensics 11.5.1 Forensic applications 11.5.2 Edge-AI requirements 11.5.3 Edge-AI forensics: stages 11.6 Edge-AI for mobility/logistics 11.6.1 Smart mobility 11.6.2 Shared mobility intelligence 11.6.3 Logistics planning 11.6.4 Product placement 11.7 Conclusion References CH012.pdf Chapter Business opportunities using edge-AI 12.1 Digital business 12.2 Economic impacting factors 12.2.1 Cloud/edge-specific costs 12.2.2 Memory-related cost improvements 12.3 Business opportunities: edge-AI platforms 12.3.1 Edge AI software: business opportunities 12.3.2 Edge AI hardware: business opportunities 12.4 Cost models 12.4.1 Cloud-based cost models 12.4.2 Serverless functions: costs 12.5 Economic simulators for FaaS implementation 12.5.1 Request-making component 12.5.2 Queues 12.5.3 Function instances 12.5.4 Monitoring server 12.6 Conclusion References CH013.pdf Chapter Challenges and future directions 13.1 Edge-AI challenges 13.1.1 Energy consumption issues 13.1.2 Resource constraints 13.1.3 Performance efficiency 13.1.4 Secured learning 13.1.5 Legal and standards