CASSIDY: A Continual Adaptive System for Secure and Interpretable Drift-Aware IoT Anomaly Detection
Abstract
: The proliferation of Internet of Things (IoT) devices has intensified cybersecurity threats, including DDoS, ransomware, scanning, injection, and backdoor attacks. Existing intrusion detection systems (IDS) for IoT suffer from critical limitations: they are computationally heavy and not suitable for edge deployment, need labeled attack data which is scarce and rapidly outdated, are not adaptable to changing network traffic (concept drift), or provide no interpretability—and no prior work to date addresses all four deficiencies at once. We propose CASSIDY (Continual Adaptive System for Secure and Interpretable Drift-aware IoT Anomaly Detection), a unified framework that bridges this gap. CASSIDY is based on a lightweight dense autoencoder with only 2359 trainable parameters (9.21 KB footprint), and is trained only on normal traffic, removing the need for attack labels. The framework integrates three key innovations: (1) conformal prediction, to provide statistically-valid uncertainty quantification with a bounded false-positive rate; (2) test-time adaptation (TTA) that fine-tunes only the decoder on unlabeled live traffic, enabling rapid recovery from concept drift without full retraining; and (3) intrinsic feature deviation attribution that provides human-readable explanations at inference speed without external libraries. Extensive evaluation on the TON_IoT dataset demonstrates strong detection performance (area under the receiver operating characteristic curve (AUC) 0.9911, F1 0.9741, near-perfect recall) and full recovery from simulated concept drift (AUC rises from 0.5026 post-drift to 0.9926 after TTA). CASSIDY achieves sub-millisecond inference (0.0055 ms/sample), making it suitable for resource-constrained IoT gateways and lightweight edge deployment scenarios. Overall, CASSIDY demonstrates that a minimal, carefully designed model can achieve competitive performance with substantially larger models while maintaining a lightweight footprint., establishing it as a promising solution for practical IoT edge security.