Design of Stealthy Attacks for Multi Rate Cyber-Physical Systems Based on Q-Learning
Shiyu Zhang, Quanhai Wang · 2025
This paper presents a Q-learning-based approach for designing stealthy attacks for multi-rate cyber-physical systems (CPS). Unlike traditional model-based attack strategies, which rely on known system dynamics, our approach is data-driven, utilizing real-time trajectory data to iteratively learn optimal attack policies. This is particularly advantageous in scenarios where system dynamics are partially known or entirely unknown. The proposed method uses a recursive formulation based on the Bellman equation to optimize the attack strategy, minimizing system performance degradation while ensuring stealthiness. By learning the kernel matrix from trajectory data, the system's response to attacks can be evaluated, allowing for effective attack design without requiring explicit system models. A simulation example is used to validate the effectiveness of the proposed scheme.