Beyond Worst-case Attacks: Robust RL with Adaptive Defense via Non-dominated Policies
Xiangyu Liu, Chenghao Deng, Yanchao Sun, Yongyuan Liang, Furong Huang
TL;DR
This work addresses robustness of reinforcement learning policies under state-adversarial attacks beyond worst-case scenarios. It introduces PROTECTED, a framework that pre-trains a finite set of non-dominated policies $\TildeΠ$ and employs online no-regret adaptation over this set at test time to minimize regret against adaptive attackers, instead of optimizing for a single worst-case policy. The authors prove intrinsic hardness for sublinear regret with unrestricted policy classes and provide iterative, finite-policy discovery methods that guarantee near-optimality up to a gap $δ$, along with practical optimization strategies. Empirical evaluation on Mujoco tasks shows that PROTECTED improves natural performance while maintaining robustness across static and dynamic attacks, with efficient adaptation even for small policy sets. Overall, the approach offers a practical balance between robustness and test-time efficiency by combining training-time policy discovery with online adaptation in a finite policy space.
Abstract
In light of the burgeoning success of reinforcement learning (RL) in diverse real-world applications, considerable focus has been directed towards ensuring RL policies are robust to adversarial attacks during test time. Current approaches largely revolve around solving a minimax problem to prepare for potential worst-case scenarios. While effective against strong attacks, these methods often compromise performance in the absence of attacks or the presence of only weak attacks. To address this, we study policy robustness under the well-accepted state-adversarial attack model, extending our focus beyond only worst-case attacks. We first formalize this task at test time as a regret minimization problem and establish its intrinsic hardness in achieving sublinear regret when the baseline policy is from a general continuous policy class, $Π$. This finding prompts us to \textit{refine} the baseline policy class $Π$ prior to test time, aiming for efficient adaptation within a finite policy class $\TildeΠ$, which can resort to an adversarial bandit subroutine. In light of the importance of a small, finite $\TildeΠ$, we propose a novel training-time algorithm to iteratively discover \textit{non-dominated policies}, forming a near-optimal and minimal $\TildeΠ$, thereby ensuring both robustness and test-time efficiency. Empirical validation on the Mujoco corroborates the superiority of our approach in terms of natural and robust performance, as well as adaptability to various attack scenarios.