WIP: False-Data Injection Against Physics-Constrained Neural Networks for Trajectory Planning in Autonomous Racing

Abstract

Prediction of competitors' motion is critical for high-speed autonomous racing, yet its robustness under adversarial manipulation remains unclear. We analyze black-box false- data-injection (FDI) attacks that corrupt perceived motion histories, inducing ego mispre diction and enabling adversarial overtaking. Our results reveal that high-speed trajectory prediction systems remain vulnerable, exposing safety risks in multi-agent autonomous driving