G-DCAP: Stealthy Adversarial Attacks on Multi-Sensor IoT Systems

Recent advances in machine learning have greatly enhanced anomaly detection in complex, correlated data streams, such as IoT sensor networks and network traffic. These models identify anomalies by detecting violations of domain constraints and interdependencies between data streams. Consequently, adversarial attacks targeting these models must carefully adhere to these constraints to remain undetected. In this work, we introduce Graph-based Domain Constraint Adhering Perturbations (G-DCAP), a novel method for generating adversarial perturbations in multivariate time series data while preserving domain constraints. G-DCAP leverages a graph-based approach to dynamically learn and enforce constraints, propagating perturbations within correlated IoT sensor networks to maximize stealth. We validate this approach across four datasets and four anomaly detection models, demonstrating that up to 92% of attacks evade detection with complete sensor visibility and 83% with limited visibility, compared to traditional attacks. This study is the first to explore domain-constrained adversarial attacks in multi-sensor IoT environments, highlighting potential security risks in realworld deployments.