MSD-CDRL: A generic fusion detection framework for logic covert attack towards cyber-physical system security

Abstract

Cyber-physical systems (CPSs) enable the integrated design of computing, communication, and physical systems, making the system more reliable, efficient, and collaborative in real time, with important and widespread applications. However, they have serious vulnerabilities to logic covert attacks (LCAs), while few existing approaches focus on LCAs. This paper developed a generic fusion detection framework that combines a mean standard deviation (MSD) module and a constrained deep reinforcement learning (CDRL) approach for CPSs. The MSD module is used to extract the fluctuation and trend characteristics of sensor measurements. Meanwhile, we use the CPS model in the DRL training process, which reduces the computational complexity and speeds up the convergence of the DRL. By establishing the physical platform and co-simulation system, the superior performance of MSD-CDRL has been demonstrated compared with three state-of-the-art methods (composite deep learning, observed Petri Nets, and DRL). Experimental results indicated that the ability of MSD-CDRL in detection accuracy has been increased significantly and the detection efficiency is 60 % higher than the existing verification methods.