Online Robustness Training for Deep Reinforcement Learning
arXiv:1911.00887
Abstract
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.
References in corpus (7)
- Rainbow: Combining Improvements in Deep Reinforcement Learning
- Defensive Distillation is Not Robust to Adversarial Examples
- Adversarial Attacks on Neural Network Policies
- Whatever Does Not Kill Deep Reinforcement Learning, Makes It Stronger
- Adversary A3C for Robust Reinforcement Learning
- Deep Robust Kalman Filter
- Collaborative Deep Reinforcement Learning