Research map / Adversarial robustness and security / Backdoor and poisoning
BIRD: Generalizable Backdoor Detection and Removal for Deep Reinforcement Learning
Xuan Chen, Wenbo Guo, Guanhong Tao, Xiangyu Zhang, Dawn Song
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 |
|---|---|---|---|
| Replace reward-based detection metric with ||∆||0 (ablation) | BIRD detection · Qbert | F1 score: 0.909 -> 0.571 | Hurt |
| Replace reward-based detection metric with ||∆||0 (ablation) | BIRD detection · COMA | F1 score: 1.000 -> 0.347 | Hurt |
| Replace reward-based detection metric with ||∆||0 (ablation) | BIRD detection · YSNP | F1 score: 0.667 -> 0.263 | Hurt |
| Add neuron re-initialization before retraining (ablation) | BIRD removal · Qbert poisoned | reward: 0.0%±0.0% -> 12694±2194 | Improved |
| Add neuron re-initialization before retraining (ablation) | BIRD removal · COMA poisoned | winning rate: 6.7%±2.1% -> 90.7%±3.0% | Improved |
| Add neuron re-initialization before retraining (ablation) | BIRD removal · YSNP poisoned | winning rate: 29.5%±4.1% -> 46.7%±3.9% | Improved |
| Use BIRD instead of Finetuning | backdoor removal · SpaceInvaders poisoned | reward: 21±65 -> 543±52 | Improved |
| Use BIRD instead of PD | backdoor removal · SpaceInvaders clean | reward: 238±94 -> 529±92 | Improved |
| Use BIRD instead of PD | backdoor removal · Qbert poisoned | reward: 6180±1817 -> 12694±2194 | Improved |
| Apply BIRD in clean environment versus original policy | clean performance · SpaceInvaders | reward: 610±97 -> 529±92 | Mixed |
| Change trigger size to 5 × 5 | BIRD detection · SpaceInvaders | F1 score: 1.0 -> 0.833 | Hurt |
| Change trigger shape to circle | BIRD detection · SpaceInvaders | F1 score: 1.0 -> 0.909 | Hurt |
Closest papers by meaning
- Fox in the Henhouse: Supply-Chain Backdoor Attacks Against Reinforcement Learning (ICML 2026 · similarity 0.90)
- SHINE: Shielding Backdoors in Deep Reinforcement Learning (ICML 2024 · similarity 0.90)
- BadRL: Sparse Targeted Backdoor Attack against Reinforcement Learning (AAAI 2024 · similarity 0.89)
- PolicyGuard: Towards Test-time and Step-level Adversary Defense for Reinforcement Learning Agent (ICML 2026 · similarity 0.89)
- Adversarial Inception Backdoor Attacks against Reinforcement Learning (ICML 2025 · similarity 0.89)
- Stealthy Imitation: Reward-guided Environment-free Policy Stealing (ICML 2024 · similarity 0.86)
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.
