Adaptive utility-aware event-triggered reinforcement learning for hybrid attack scheduling against remote state estimation.

Abstract

Remote state estimation plays an important role in connected vehicle platoons, industrial automation systems, and other networked Cyber-Physical Systems (CPSs), where reliable state information is essential for monitoring, feedback control, and decision-making. However, due to the openness and unreliability of wireless communication links, remote estimation systems are vulnerable to eavesdropping and Denial-of-Service (DoS) attacks, which may cause information leakage, packet loss, and estimation performance degradation. This issue becomes more critical in time-varying wireless environments, where channel conditions and attack opportunities evolve dynamically over time, making conventional static or periodic attack models insufficient for characterizing practical security risks. To address this problem, this paper investigates a utility-aware event-triggered reinforcement learning framework for hybrid attack scheduling against remote state estimation over time-varying wireless channels. The attacker can select among eavesdropping, DoS, and silence actions to balance estimation disruption, information acquisition, and attack resource consumption. The hybrid attack scheduling problem is formulated as a partially observable Markov decision process (POMDP), and a utility-aware event-triggered mechanism is designed to activate attack decisions only when the estimated attack utility is sufficiently significant. At the triggered decision instants, a proximal policy optimization (PPO) algorithm is employed to learn an adaptive hybrid attack mode selection policy. The structural properties of the resulting policy are also analyzed theoretically, showing that the optimal belief-space policy has a piecewise constant structure and that the proposed adaptive threshold preserves a monotone triggering property with respect to the attack utility indicator. Simulation results in a connected vehicle platoon scenario demonstrate that, compared with several benchmark methods, the proposed method achieves a better trade-off among remote estimation degradation, attacker-side information acquisition, and energy consumption. These results indicate that the proposed event-triggered reinforcement learning framework can improve the adaptability and resource efficiency of hybrid attack scheduling under time-varying wireless channels. The study also provides useful insights for security vulnerability assessment, resilient estimation design, and defense strategy development for practical remote estimation systems.