Conformance-Driven Anomaly Detection for Cyber-Physical Systems

The rapid integration of learning-enabled components into safety-critical systems has created a rising demand for robust frameworks to ensure operational reliability and safety. Current approaches, which focus on enhancing robustness during training or implementing self-monitoring mechanisms, remain vulnerable to adversarial manipulation of system inputs or internal data flows, exposing risks of malicious exploitation. To address this gap, we propose a novel third-party monitoring framework that independently evaluates cyber-physical systems safety by statistically analyzing real-time behavior against training behaviors recorded in historically safe operational contexts. Our method employs stochastic conformance testing to account for environment and system stochasticity, enabling comparisons between observed behaviors and validated behaviors. We demonstrate our proposed pipeline using a cartpole example in which reinforcement learning serves as the system controller.