Tactics of Robust Deep Reinforcement Learning with Randomized Smoothing

Chung-En Sun, Sicun Gao, Tsui-Wei Weng

Primary Area: reinforcement learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Reinforcement Learning, Robustness, Adversarial Attack

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: Despite randomized smoothing being proven to give a robustness guarantee, the standard performance of a smoothed deep reinforcement learning (DRL) agent exhibits a significant trade-off between its utility and robustness. Naively introducing randomized smoothing during the training or testing can fail completely in the DRL setting. To address this issue, we proposed new algorithms to train smoothed robust DRL agents while attaining superior clean reward, empirical robustness, and robustness guarantee in discrete and continuous action space. Our proposed DS-DQN and AS-PPO outperform prior state-of-the-art robustly-trained agents in robust reward by $1.6\times$ on average and exhibit strong guarantees that previous agents failed to achieve. Moreover, a stronger adversarial attack for smoothed DQN agents is proposed, which is $4.6\times$ more effective in decreasing the rewards compared to existing adversarial attacks.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3197