검색 상세

Multi-Domain Deep Learning Approaches for Integrated Security and Sensing in Smart Environments

초록/요약

The rapid proliferation of interconnected cyber-physical systems, including communication networks, smart homes, and intelligent mobility infrastructures, has created a growing demand for adaptive, secure, and resource-efficient deep learning approaches. Despite notable progress in artificial intelligence, achieving high detection and recognition performance under real-world constraints – limited annotated data, rapidly evolving environments, and embedded-level computational budgets – remains a persistent challenge. This dissertation addresses these challenges through three complementary research contributions that collectively enhance the robustness, scalability, and efficiency of deep learning across heterogeneous domains. The first contribution, Optimisation of Recurrent Neural Networks for High-Performance Intrusion Detection in Network Traffic, introduces the Optimised Recurrent Neural Network (O-RNN), designed to strengthen temporal sequence modelling for intrusion detection in IoT and conventional networks. By restructuring LSTM operations, refining sequence shaping, and applying a hybrid optimisation strategy, O-RNN reduces vanishing gradient effects and computational overhead, achieving stable and competitive detection accuracy across IoT-23, CICIDS2017, and KDD99 datasets. The second contribution, Adaptive Transfer Learning with Ensemble Networks for Activity Recognition in Smart Homes (ATLEN), presents a domain-adaptive ensemble framework that enables robust human activity recognition despite sensor heterogeneity and scarce labelled data. Incorporating TeLU-enhanced recurrent networks and adaptive domain alignment, ATLEN achieves 98.7% accuracy and significantly improves cross-domain generalisation, demonstrating effective transfer-based intelligence for dynamic smart home environments. The final contribution, STEP-SAFE: Smart TinyML and Few-Shot Enabled Pedestrian and Cyclist Safety Framework for Intelligent Transportation Systems, extends lightweight adaptive modelling to safety-critical urban mobility. Integrating TinyML-compatible architectures with few-shot meta-learning, STEP-SAFE recognises vulnerable road user (VRU) behaviours with only a few labelled samples per class. Through INT8 quantisation, sparsity-aware pruning, and a risk-aware alert mechanism, STEP-SAFE achieves efficient on-device inference (latency under 64 ms) while maintaining 85.2% accuracy and 0.85 mean average precision under constrained intelligent mobility settings. Together, these contributions advance deep learning for security, adaptability, and resource efficiency across intrusion detection, smart home activity recognition, and intelligent mobility safety. The unified perspective offered by this dissertation demonstrates that targeted model optimisation, adaptive transfer strategies, and efficient learning paradigms can transform deep neural architectures into scalable, context-aware solutions suitable for next-generation intelligent cyber-physical ecosystems.

more

목차

I. Introduction 1
1.1 Motivation and Objectives 2
1.2 Research Contributions 3
1.2.1 Optimized Recurrent Neural Networks for Network Security Domain 5
1.2.2 Activity Recognition with Ensemble Networks for Ambient Sensing Domain 5
1.2.3 TinyML and Few-Shot Enabled Safety Framework for Mobility Sensing Domain 7
1.3 Organization of the Dissertation 7
II. Background 9
2.1 Deep Learning in Network Intrusion Detection 9
2.1.1 Evolution of Learning based Intrusion Detection 10
2.1.2 Deep Learning Paradigm for IDS 10
2.1.3 Hybrid, Autoencoder, and Graph based Models 11
2.1.4 Emergence of Transformer and Attention based IDS 12
2.1.5 Feature Representation and Dataset Challenges 12
2.1.6 Optimization, Adaptability, and Real-Time Constraints 13
2.2 Advancements in Human Activity Recognition for Smart Homes 14
2.2.1 From Conventional Methods to Deep Learning 14
2.2.2 Hybrid and Attention driven Architectures 15
2.2.3 Domain Adaptation and Transfer Learning 15
2.2.4 Ensemble and Meta learning Approaches 16
2.2.5 Graph and Context-Aware Representations 17
2.2.6 Challenges in Data, Privacy, and Efficiency 18
2.3 Deep Learning for Pedestrian and Cyclist Safety in Mobility Systems 18
2.3.1 Evolution of Vision-Based Safety Detection 19
2.3.2 Multimodal and Spatio-Temporal Learning Approaches 20
2.3.3 Attention, Transformer, and Graph-Based Models 20
2.3.4 Few-Shot and Meta learning for Data Efficient Safety Systems 21
2.3.5 TinyML and Edge-Aware Safety Frameworks 22
2.3.6 Interpretability and Ethical Considerations 22
2.4 Challenges and Research Gaps 23
III. Optimization of Recurrent Neural Networks for High-Performance Intrusion Detection in Network Traffic 25
3.1 Introduction 25
3.2 Problem Statement and Research Contributions 28
3.2.1 Problem Statement 28
3.3 Proposed Methodology 31
3.3.1 O-RNN Overview 31
3.3.2 Adaptive LSTM Layers in O-RNN 32
3.3.3 Hybrid Optimization Strategy 34
3.3.4 Performance Impact 35
3.4 Experimental Setup and Results 36
3.4.1 Experimental Design and Evaluation Protocol 37
3.4.2 Datasets and Descriptive Overview 38
3.4.3 Hyperparameter Tuning and Model Configuration 39
3.4.4 Performance Metrics Comparison 40
3.4.5 Performance Analysis: Insights and Observations 43
3.4.6 Comparative Study with State-of-the-Art Models 45
3.5 Summary and Key Findings 46
IV. Adaptive Transfer Learning with Ensemble Networks for Activity Recognition in Smart Homes 47
4.1 Introduction 47
4.2 Problem Statement and Research Contributions 50
4.3 Proposed Methodology 52
4.3.1 Phase I: Ensemble Based Feature Extraction 53
4.3.2 Phase II: Adaptive Transfer Learning 56
4.3.3 Positioning ATLEN Relative to SDL and MDL 57
4.4 Experimental Setup and Results 58
4.4.1 Smart Home Data Collection: van Kasteren Dataset 58
4.4.2 Generating Structured Activity Sets and Settings 59
4.4.3 Evaluation Metrics 60
4.4.4 Feature Extraction and Cross-Domain Adaptation Performance 61
4.4.5 Ensemble RNN Comparison 62
4.4.6 Evaluation of ATLEN in Cross-Domain HAR 64
4.4.7 Classification Performance of ATLEN in Cross-Domain HAR 64
4.4.8 Training and Validation Performance Analysis 65
4.5 Summary and Key Findings 66
V. Smart TinyML and Few-Shot Enabled Pedestrian and Cyclist Safety Framework for Intelligent Transportation System 68
5.1 Introduction 68
5.2 Problem Statement and Research Contributions 71
5.2.1 Problem Formulation 71
5.3 Proposed Methodology 73
5.3.1 Embedding-Based Classification via Few-Shot Episodic Meta-Learning 73
5.3.2 TinyML Optimized Inference and Risk-Aware Alert Generation 75
5.4 Experimental Setup and Results 78
5.4.1 Dataset and Task Design 78
5.4.2 Model Architecture, Training Protocol, and Deployment Simulation 79
5.4.3 Evaluation Metrics 80
5.4.4 Results and Discussion 85
5.5 Summary and Key Findings 86
VI. Conclusion and Future Work 87
6.1 Conclusion 87
6.2 Future Work 88
References 90

more
반출 Meta View 목록