Tactics of Adversarial Attack on Deep Reinforcement Learning Agents

Yen-Chen Lin, Zhang-Wei Hong, Yuan-Hong Liao, Meng-Li Shih, Ming-Yu Liu, Min Sun

Abstract: We introduce two novel tactics for adversarial attack on deep reinforcement learning (RL) agents: strategically-timed and enchanting attack. For strategically- timed attack, our method selectively forces the deep RL agent to take the least likely action. For enchanting attack, our method lures the agent to a target state by staging a sequence of adversarial attacks. We show that both DQN and A3C agents are vulnerable to our proposed tactics of adversarial attack.

TL;DR: We propose two tactics of adversarial attacks for deep reinforcement learning and show their strength.

Keywords: Deep learning, Reinforcement Learning

Conflicts: nvidia.com, nthu.edu.tw, merl.com