Research map / Adversarial robustness and security / Black-box and transfer attacks
SEBA: Sample-Efficient Black-Box Attacks on Visual Reinforcement Learning
Tairan Huang, Yulin Jin, Junxu Liu, Qingqing Ye, Haibo Hu
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 discriminator in Stage 2 (ablation) | SEBA · Pixel MuJoCo | FID: 97.18 -> 62.43 | Improved |
| add discriminator in Stage 2 (ablation) | SEBA · Cheetah Run (pixel MuJoCo) | reward under attack: 3.15 -> 1.61 | No clear effect |
| train shadow critic on perturbed states in Stage 1 (ablation) | SEBA · Walker Walk (pixel MuJoCo) | reward under attack: 118.11 -> 35.74 | Improved |
| train shadow critic on perturbed states in Stage 1 (ablation) | SEBA · Cheetah Run (pixel MuJoCo) | reward under attack: 19.25 -> 1.61 | Improved |
| add world model for synthetic rollouts (ablation) | SEBA · Pixel MuJoCo | training env queries: 800K -> 160K | Improved |
| add world model for synthetic rollouts (ablation) | SEBA · Cheetah Run (pixel MuJoCo) | reward under attack: 0.18 vs 1.61 | No clear effect |
| use SEBA instead of PGD (white-box) | Pixel MuJoCo attack · Cheetah Run (pixel MuJoCo) | reward under attack: 150.72 -> 1.61 | Improved |
| use SEBA instead of OPTIMAL (black-box RL attacker) | Pixel MuJoCo attack · Reacher Hard (pixel MuJoCo) | reward under attack: 592.64 -> 0.3 | Improved |
| use SEBA instead of PA-AD in training cost | Pixel MuJoCo attack · Pixel MuJoCo | training env queries: 4M -> 160K | Improved |
| use SEBA instead of PA-AD | Atari attack · Pong (Atari) | reward under attack: -21 -> 3 | Hurt |
| use SEBA instead of OPTIMAL | Atari attack · Alien (Atari) | reward under attack: 7400 -> 982 | Improved |
| use SEBA objective instead of PGD for targeted attack | Targeted attack · Walker Walk targeted | success rate (%): 37.2 -> 91.3 | Improved |
Closest papers by meaning
- Reinforcement Learning Platform for Adversarial Black-box Attacks with Custom Distortion Filters (AAAI 2025 · similarity 0.88)
- Sparse Black-Box Video Attack with Reinforcement Learning (IJCV 2022 · similarity 0.87)
- Natural Black-Box Adversarial Examples against Deep Reinforcement Learning (AAAI 2022 · similarity 0.87)
- Efficient Reinforcement Learning Through Adaptively Pretrained Visual Encoder (AAAI 2025 · similarity 0.85)
- Diffusion Guided Adversarial State Perturbations in Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- Efficient Black-box Adversarial Attacks via Bayesian Optimization Guided by a Function Prior (ICML 2024 · similarity 0.85)
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.
