On Zero-Dynamics Stealthy Attacks with Learned State Space Models

Zero-dynamics stealthy attacks are a subset of false data injection attacks (FDIAs) that can be catastrophic as they are designed to be undetectable by traditional residual based anomaly detectors. The attacker's ability to successfully attack a system relies heavily on their capacity to learn an accurate state space model. Utilizing a grey box approach to system identification, we show that even when the attacker is able to learn a state space model close enough to have a high probability of a successful attack, making small improvements to the system's anomaly detector causes the probability of a successful attack to drop drastically. Finally, we study the trade-offs between making the system less susceptible to zero-dynamics attacks and maintaining its controllability, by increasing the sampling time of the system, thus providing the attacker fewer samples to learn a state space model.