Wenbo Guo, UC Berkeley; Xian Wu, Northwestern University; Lun Wang, UC Berkeley; Xinyu Xing, Northwestern University; Dawn Song, UC Berkeley
Recent advances in deep reinforcement learning (DRL) takes artificial intelligence to the next level, from making individual decisions to accomplishing sophisticated tasks via sequential decision makings, such as defeating world-class human players in various games and making real-time trading decisions in stock markets. Following these achievements, we have recently witnessed a new attack specifically designed against DRL. Recent research shows by learning and controlling an adversarial agent/policy, an attacker could quickly discover a victim agent's weaknesses and thus force it to fail its task.
Due to differences in the threat model, most existing defenses proposed for deep neural networks (DNN) cannot be migrated to train robust policies against adversarial policy attacks. In this work, we draw insights from classical game theory and propose the first provable defense against such attacks in two-player competitive games. Technically, we first model the robust policy training problem as finding the nash equilibrium (NE) point in the entire policy space. Then, we design a novel policy training method to search for the NE point in complicated DRL tasks. Finally, we theoretically prove that our proposed method could guarantee the lowerbound performance of the trained agents against arbitrary adversarial policy attacks. Through extensive evaluations, we demonstrate that our method significantly outperforms existing policy training methods in adversarial robustness and performance in non-adversarial settings.
Open Access Media
USENIX is committed to Open Access to the research presented at our events. Papers and proceedings are freely available to everyone once the event begins. Any video, audio, and/or slides that are posted after the event are also free and open to everyone. Support USENIX and our commitment to Open Access.
author = {Wenbo Guo and Xian Wu and Lun Wang and Xinyu Xing and Dawn Song},
title = {{PATROL}: Provable Defense against Adversarial Policy in Two-player Games},
booktitle = {32nd USENIX Security Symposium (USENIX Security 23)},
year = {2023},
isbn = {978-1-939133-37-3},
address = {Anaheim, CA},
pages = {3943--3960},
url = {https://www.usenix.org/conference/usenixsecurity23/presentation/guo-wenbo},
publisher = {USENIX Association},
month = aug
}