Data-Driven Attack Detection for Linear Systems

This letter studies the attack detection problem in a data-driven and model-free setting, for deterministic systems with linear time-invariant dynamics. Differently from existing that leverage knowledge of system dynamics to derive security bounds monitoring schemes, we treat cases where dynamics, as well strategy location, are unknown. We fundamental limitations function only observed data without estimating (in fact, no assumption is made on identifiability system). In particular, (i) informativity length observation window, (ii) provide characterization undetectable attacks, (iii) construct monitor. Surprisingly, our results show while requires larger window attain capability, once attained it shares same model-based monitoring.