Secure Set-Based State Estimation for Linear Systems under Adversarial Attacks on Sensors

When a strategic adversary can attack multiple sensors of a system and freely choose a different set of sensors at different times, how can we ensure that the state estimate remains uncorrupted by the attacker? The existing literature addressing this problem mandates that the adversary can only corrupt less than half of the total number of sensors. This limitation is fundamental to all point-based secure state estimators because of their dependence on algorithms that rely on majority voting among sensors. However, in reality, an adversary with ample resources may not be limited to attacking less than half of the total number of sensors. This paper avoids the above-mentioned fundamental limitation by proposing a set-based approach that allows attacks on all but one sensor at any given time. We guarantee that the true state is always contained in the estimated set, which is represented by a collection of constrained zonotopes, provided that the system is bounded-input-bounded-state stable and redundantly observable via every combination of sensor subsets with size equal to the number of uncompromised sensors. Additionally, we show that the estimated set is secure and stable irrespective of the attack signals if the process and measurement noises are bounded. To detect the set of attacked sensors at each time, we propose a simple attack detection technique. However, we acknowledge that intelligently designed stealthy attacks may not be detected and, in the worst-case scenario, could even result in exponential growth in the algorithm’s complexity. We alleviate this shortcoming by presenting a range of strategies that offer different levels of trade-offs between estimation performance and complexity. To illustrate the efficacy of our approach, we apply it to a vertically interconnected mechanical system that models a three-story building structure. Our results demonstrate that the proposed set-based method provides a robust and secure state estimation method that can handle a greater number of attacked sensors than existing point-based estimators.