Antagonistic Attack and Effective Defense of Autonomous Vehicle

This paper focuses on the adversarial attack of autonomous vehicle. Based on robust optimization theory, an integrated defense architecture is proposed that includes dynamic course adversarial training (DCAT) and Bayesian uncertainty detector (BUD). The defense strategies are divided into three categories: model enhancement, input purification, and runtime monitoring, and their implementation methods and applicable stages are clarified. The results show that the DCAT+BUD method has high robust accuracy in white box, black box and physical attack scenarios, and effectively improves the safety of the auto drive system.