Less Is More: A Stealthy and Efficient Adversarial Attack Method for DRL-Based Autonomous Driving Policies

Existing research has demonstrated that autonomous driving policies based on deep reinforcement learning (DRL) are vulnerable to adversarial attacks, which poses challenges for the practical deployment of these policies. Designing effective adversarial attacks is a crucial prerequisite for building robust driving policies. In view of this, we propose a novel adversarial attack method, which can attack the DRL-based autonomous driving agents in a stealthy and efficient manner. This method models the attack as a mixed-integer optimization problem that aims to maximize the safety violations (e.g., collisions) of the agents while minimizing the number of attack steps. Then, a DRL-based adversary is devised in this method to solve the problem to automatically learn the optimal attack policy without domain knowledge. To further enhance the adversarys learning capability, this method incorporates attack-related information into its observations to provide more decisionmaking context and employ a trajectory clipping technique to enhance sample quality. Extensive evaluation results reveal that our method achieves a remarkable 105% enhancement in attack efficiency compared to existing methods.