Black-Box Reward Attacks Against Deep Reinforcement Learning Based on Successor Representation

Although the deep reinforcement learning (DRL) technology has been widely adopted in various fields, it become an important research hotspot to study vulnerability of DRL for improving robustness agents. The adversarial attack methods based on white-box models, where adversary can access all information victims, have intensively investigated. However, most practical situations, cannot obtain internal victim’s neural network. Furthermore, reward-based attacks, agent perform anomaly detection perturbed rewards detect whether attacked. In this paper, we propose a black-box method with corrupted rewards, which employs exploration mechanisms improve effectiveness attacking builds network advance learn successor representation (SR) each state. Then, determine timing attacks and generate imperceptible perturbations values SR. Experimental results show that algorithm SR proposed paper effectively agents fewer samples.