Abstract
Reinforcement learning (RL) algorithms learn and explore nearly any state any number of times in their environment, but minute adversarial attacks cripple these agents. In this work, we define our threat model against RL agents as such: Adversarial agents introduce small permutations to the input data via black-box models with the goal of reducing the optimality of the agent. We focus on pre-processing adversarial images before they enter the network to reconstruct the ground-truth images.