Directed Poisoning Attacks on FRIT in Adaptive Cruise Control

Recent advances in connected-vehicle technologies have enabled the large-scale collection of driving data, facilitating the deployment of data-driven control schemes. Although these methods offer advantages by eliminating the need for explicit modeling, they also introduce vulnerabilities due to their reliance on stored data. This study investigates a class of targeted data poisoning attacks on fictitious reference iterative tuning, a widely used data-driven controller tuning approach. We present a method that allows an adversary to influence closed-loop dynamics by manipulating the training data so that the resulting controller behavior matches a maliciously defined reference response. This strategy differs from conventional poisoning attacks, which aim only to the degrade control performance. Instead, it enables deliberate alteration of control characteristics such as overshoot and convergence time. The proposed attack is formulated as a constrained optimization problem under bounded tampering signals. Through a numerical study involving adaptive cruise control with stop functionality, we show that minor data modifications, indistinguishable from sensor noise, can cause significant degradation in control behavior. These findings highlight the need for robust security mechanisms in data-driven control implementation.