Adversarial Vulnerability Assessment of GNN-Based Anomaly Detection in Smart Home Environments

Abstract

This thesis investigates the adversarial robustness of Graph Neural Network (GNN)-based anomaly detection systems in smart-home environments. Recent anomaly detection ap￾proaches increasingly leverage graph-based learning to model temporal and relational depen￾dencies between heterogeneous Internet of Things (IoT) devices. While such models achieve promising detection performance under normal operating conditions, their resilience against adversarial manipulation remains insufficiently understood. The study focuses on two rep￾resentative GNN-based anomaly detection models, Graph Deviation Network (GDN) and MTAD-GAT, evaluated using the IoT Garage smart-home dataset. A preprocessing pipeline was developed to transform multivariate sensor streams into graph-structured temporal data suitable for anomaly detection. The robustness of both models was assessed under multiple adversarial scenarios corresponding to black-box, grey-box, and white-box threat settings. Several perturbation strategies, including random, query-based, explainability-guided, and gradient-based attacks, were applied under different perturbation budgets. Experimental results demonstrate that adversarial perturbations can significantly degrade anomaly detection performance, even when modifications to sensor measurements remain limited. The findings further reveal differences in robustness between GDN and MTAD GAT, highlighting the influence of architectural design and graph-based dependency mod eling on adversarial vulnerability. Performance degradation generally increases with per turbation magnitude, although attack effectiveness varies across threat models and attack techniques.