d347: Deep Reinforcement Learning Vulnerability, Policy Induction Attacks

Deep Reinforcement Learning Vulnerability to Policy Induction Attacks: https://arxiv.org/abs/1701.04143v1

Deep Reinforcement Learning Vulnerability

PDF: https://arxiv.org/pdf/1701.04143v1.pdf | Review: http://www.shortscience.org/paper?bibtexKey=journals/corr/1701.04143

Abstract: “In this work, we establish that reinforcement learning techniques based on Deep Q-Networks (DQNs) are also vulnerable to adversarial input perturbations, and verify the transferability of adversarial examples across different DQN models. Furthermore, we present a novel class of attacks based on this deep reinforcement learning vulnerability that enable policy manipulation and induction in the learning process of DQNs. We propose an attack mechanism that exploits the transferability of adversarial examples to implement policy induction attacks on DQNs, and demonstrate its efficacy and impact through experimental study of a game-learning scenario.