On the Stealth of Unbounded Attacks Under Non-Negative-Kernel Feedback

Abstract: The stealth of false data injection attacks (FDIAs) against feedback sensors in linear time-varying (LTV) control systems is investigated. In that regard, the following notions of stealth are pursued: For some finite $ε> 0$, i) an FDIA is deemed $ε$-stealthy if the deviation it produces in the signal that is monitored by the anomaly detector remains $ε$-bounded for all time, and ii) the $ε$-stealthy FDIA is further classified as untraceable if the bounded deviation dissipates over time (asymptotically). For LTV systems that contain a chain of $q \geq 1$ integrators and feedback controllers with non-negative impulse-response kernels, it is proved that polynomial (in time) FDIA signals of degree $a$ - growing unbounded over time - will remain i) $ε$-stealthy, for some finite $ε> 0$, if $a \leq q$, and ii) untraceable, if $a < q$. These results are obtained using the theory of linear Volterra integral equations.