Online Robustness Training for Deep Reinforcement Learning
This addresses the vulnerability of deep RL agents to adversarial attacks, which is an incremental improvement in robustness training methods.
The paper tackles the problem of adversarial attacks disrupting deep reinforcement learning agents by proposing Robust Student-DQN (RS-DQN), which enables online robustness training while maintaining competitive performance, resulting in an agent resilient to strong attacks during both training and evaluation.
In deep reinforcement learning (RL), adversarial attacks can trick an agent into unwanted states and disrupt training. We propose a system called Robust Student-DQN (RS-DQN), which permits online robustness training alongside Q networks, while preserving competitive performance. We show that RS-DQN can be combined with (i) state-of-the-art adversarial training and (ii) provably robust training to obtain an agent that is resilient to strong attacks during training and evaluation.