Reward-Preserving Attacks For Robust Reinforcement Learning

Abstract

Adversarial training in reinforcement learning (RL) is challenging because perturbations cascade through trajectories and compound over time, making fixed-strength attacks either overly destructive or too conservative. We propose reward-preserving attacks, which adapt adversarial strength so that an $\alpha$ fraction of the nominal-to-worst-case return gap remains achievable at each state. In deep RL, perturbation magnitudes $\eta$ are selected dynamically, using a learned critic $Q((s,a),\eta)$ that estimates the expected return of $\alpha$-reward-preserving rollouts. For intermediate values of $\alpha$, this adaptive training yields policies that are robust across a wide range of perturbation magnitudes while preserving nominal performance, outperforming fixed-radius and uniformly sampled-radius adversarial training.