Stealthy Sensor Attacks Against Direct Data-Driven Controllers

Abstract

This paper investigates the vulnerability of discrete-time linear time-invariant systems to stealthy sensor attacks during the learning phase. In particular, we demonstrate that a {data-driven} adversary, without access to the system model, can inject attacks that mislead the operator into learning an {unstable} state-feedback controller. We further analyze attacks that degrade the performance of data-driven ${H}_2$ controllers, while ensuring that the operator can always compute a feasible controller. Potential mitigation strategies are also discussed. Numerical examples illustrate the effectiveness of the proposed attacks and underscore the importance of accounting for adversarial manipulations in data-driven controller design.