검색 상세

LSTM-AE–Backed Behavioural Intent Verification for Matter Hubs : Deviation Scoring and Graded Access Enforcement

초록/요약

The Matter standard has become a significant basis for secure and interoperable smart- home communication by providing unified device commissioning, encrypted commu- nication, and Access Control List （ACL）-based authorisation across devices from dif- ferent vendors. While these mechanisms offer strong protection against unauthorised access, they generally assume that once a user or device has been authenticated, subsequent behaviour will remain legitimate. In real smart-home environments, this assumption can be overly limiting. An attacker may obtain valid credentials through device compromise, token theft, misconfigured applications, or misused automation rules and may then issue commands that satisfy protocol and ACL requirements while still appearing suspicious in context. This thesis addresses this limitation by proposing a behaviour-aware access-control framework for Matter hubs. The framework extends conventional authentication and authorisation with a lightweight behavioural verification layer that evaluates whether an incoming command aligns with learned household activity patterns. It consists of three closely connected components. First, a Long Short-Term Memory Autoencoder （LSTM-AE） is trained in an unsupervised manner to capture the normal behavioural structure from historical smart-home event streams. Second, a Behavioural Intent Verification （BIV） module transforms reconstruction error into a normalised devia- tion score that indicates how atypical a behavioural window is relative to previously observed household activity. Third, a Behaviour-Aware Policy Engine （BPE） converts that score into graded enforcement actions such as Accept, Accept+Log, Confirm, and MFA/Block. In this manner, the system evaluates whether a command is authorised and whether it is behaviourally plausible. To evaluate the proposed framework, the study uses real smart-home traces drawn primarily from the CASAS Aruba dataset, with additional cross-household testing on the CASAS Kyoto dataset. Device and sensor events are converted into Matter-like event streams and grouped into fixed temporal windows for behavioural modelling. Several behaviour-inspired anomaly scenarios are then introduced, including location- inconsistent motion patterns, nighttime unlock attempts without presence signals, rapid toggle-spikes, and persistent light-always-on behaviour. The experimental re- sults show strong anomaly-detection performance, with mean ROC–AUC values of 0.9968 on Aruba and 0.9925 on Kyoto. Scenario-level analysis also reveals near-perfect discrimination in several cases, while ablation studies indicate that the compact model remains robust across different window sizes and configuration settings. Besides anomaly detection, this thesis also explores the security–usability impli- cations of using behavioural scores for access control decisions. The Behaviour-Aware Policy Engine supports threshold-based enforcement, allowing the system to balance anomaly coverage against unnecessary disruption of benign activity. With approx- imately 55K parameters and a lightweight inference profile, the proposed model is compact enough to support further evaluation for deployment on resource-constrained matter hubs. Overall, this thesis shows that behavioural modelling can act as a practical defence- in-depth layer for Matter-enabled smart homes. By extending access control beyond identity and permission to include behavioural context, the proposed framework sup- ports more adaptive, interpretable, and risk-aware enforcement in modern smart-home environments. Keywords: Matter, smart-home security, behavioural intent verification, behaviour- aware policy engine, LSTM autoencoder, anomaly detection, access control.

more

목차

Chapter 1 Introduction 1
1.1 Motivation 1
1.2 Problem Statement 2
1.3 Research Objectives 3
1.4 Scope of Study 3
1.5 Contributions 4
1.6 Thesis Organization 5
Chapter 2 Background, Related Work and Threat Model 6
2.1 Matter Architecture and Access Control 6
2.2 Limitations of Credential-Based Enforcement 7
2.3 Related Work 7
2.3.1 Behavioural Anomaly Detection in Smart Homes 7
2.3.2 Context-Aware and Risk-Aware Access Control 8
2.4 Research Gap 8
2.5 Threat Model 10
2.5.1 System Model and Assumptions 10
2.5.2 Adversary Model 10
2.5.3 Security and Usability Objectives 11
Chapter 3 System Design and Methodology 12
3.1 Architecture Overview 12
3.2 Behaviour Representation and Windowing 13
3.3 LSTM Autoencoder for Behaviour Modelling 14
3.4 Behavioural Intent Verification 14
3.5 Behaviour-Aware Policy Engine 15
3.6 Deployment Rationale 16
3.7 Methodology 16
3.7.1 Dataset Selection 16
3.7.2 Sensor Subset and Behavioural Scope 17
3.7.3 Preprocessing and Window Construction 17
3.7.4 Model Configuration and Training 18
3.7.5 Baseline Methods 18
3.7.6 Behaviour-Inspired Anomaly Scenarios 19
3.7.7 Evaluation Metrics 19
Chapter 4 Experimental Results and Analysis 21
4.1 BPE Decisions on Benign Behaviou 21
4.2 Detection of Behavioural Anomalies 22
4.3 Comparison with Baselines 23
4.4 Robustness to Noise and Sensor Failure 24
4.5 Security–Usability Trade-of 25
4.6 End-to-End Perspective 26
4.7 Ablation Across Configurations 26
4.8 Cross-Household Generalization 27
Chapter 5 Discussion and Conclusion 28
5.1 Practical Deployment Considerations 28
5.2 Why the Design Matters 28
5.3 Limitations 30
5.4 Future Directions 30
5.5 Conclusion 31
Bibliography 33

more
반출 Meta View 목록