Sensor attack detection using history based pairwise inconsistency

This paper focuses on the security issues of Cyber-Physical Systems with multiple sensors measuring the same physical variables. We use an abstract sensor model, and each sensor provides the controller with an interval that contains the true value. Some of the sensors may be subject to malicious attacks and provide the wrong measurements, thereby misleading the controller into performing an unsafe action. Although there are several existing methods for detecting sensor attacks in the presence of transient sensor faults, they treat all sensors' faults and attacks in the same way, and may not work well when an attacker has sufficient ability to cover up for different sensors at different times, e.g., stealth attacks. To address this problem, we propose a pairwise inconsistency based algorithm to enhance attack detection capability. The main idea is to build different fault models for different sensors, add a virtual sensor to utilize the fusion intervals and historical measurements, and use pairwise inconsistencies between real and virtual sensors to identify attacks. Finally, we validate the performance of the algorithm on real measurement data obtained from the LEGO EV3 ground vehicle, the results show that the proposed method outperforms state-of-the-art algorithms.