Optimal Attacks on Reinforcement Learning Policies

Alessio Russo, Alexandre Proutiere

Control policies, trained using the Deep Reinforcement Learning, have been recently shown to be vulnerable to adversarial attacks introducing even very small perturbations to the p…