Secure system identification

Mehrdad Showkatbakhsh, Paulo Tabuada, Suhas N. Diggavi

Abstract: This work is concerned with the identification of linear time-invariant systems in the presence of an adversarial agent that attacks sensor measurements. The attacker is omniscient and we impose no restrictions (statistical or otherwise) on how the adversary alters the sensor measurements. We work in a noisy scenario where, in addition to the attacks, the sensor measurements are also affected by additive noise. Given a bound on the number of attacked sensors, and under a certain observability condition, we show that we can still construct a model that is useful for stabilization. Furthermore, we show that this model is closely related to the original system.