Research map / Adversarial robustness and security / Certified and empirical robustness
Robust Deep Reinforcement Learning against Adversarial Perturbations on State Observations
Huan Zhang, Hongge Chen, Chaowei Xiao, Bo Li, Mingyan Liu, Duane Boning, Cho-Jui Hsieh
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add state-adversarial regularizer (SA-PPO, SGLD) to PPO | PPO · Hopper best attack | reward under best attack: 733.8 -> 1403.3 | Improved |
| add state-adversarial regularizer (SA-PPO, SGLD) to PPO | PPO · Walker2d best attack | reward under best attack: 913.7 -> 2468.4 | Improved |
| add state-adversarial regularizer (SA-PPO, SGLD) to PPO | PPO · Humanoid best attack | reward under best attack: 884.1 -> 6073.8 | Improved |
| add state-adversarial regularizer to PPO | PPO · Humanoid natural | natural reward: 5270.6 -> 6624.0 | Improved |
| naive adversarial training (100% steps) of PPO | PPO · Hopper natural | natural reward: 3167.6 -> 6.1 | Hurt |
| use convex relaxation instead of SGLD to solve regularizer in SA-PPO (ablation) | SA-PPO · Hopper best attack | reward under best attack: 1403.3 -> 1224.2 | Mixed |
| add state-adversarial regularizer (SA-DDPG) to DDPG | DDPG · Ant best attack | reward under best attack: 142 -> 2007 | Improved |
| add state-adversarial regularizer (SA-DDPG) to DDPG | DDPG · Walker2d natural | natural reward: 1870 -> 4540 | Improved |
| naive adversarial training (50%) of DDPG | DDPG · Walker2d best attack | reward under best attack: 790 -> 837 | No clear effect |
| add SA-DQN regularizer (convex relaxation) to DQN | DQN · RoadRunner PGD-10 | reward under attack: 0.0 -> 44732.0 | Improved |
| use convex relaxation instead of PGD to train SA-DQN (ablation) | SA-DQN · RoadRunner PGD-10 | reward under attack: 20402.0 -> 44732.0 | Improved |
| adversarial training of DQN (50% frames) | DQN · Pong natural | natural reward: 21.0 -> 10.1 | Hurt |
Closest papers by meaning
- Towards Optimal Adversarial Robust Q-learning with Bellman Infinity-error (ICML 2024 · similarity 0.91)
- Breaking the Barrier: Enhanced Utility and Robustness in Smoothed DRL Agents (ICML 2024 · similarity 0.90)
- Enhancing Robustness in Deep Reinforcement Learning: A Lyapunov Exponent Approach (NeurIPS 2024 · similarity 0.90)
- Adversarial Diffusion for Robust Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Beyond Worst-case Attacks: Robust RL with Adaptive Defense via Non-dominated Policies (ICLR 2024 · similarity 0.90)
- Improve Robustness of Reinforcement Learning against Observation Perturbations via l∞ Lipschitz Policy Networks (AAAI 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
