Research map / Adversarial robustness and security / Backdoor and poisoning
Policy Teaching in Reinforcement Learning via Environment Poisoning Attacks
Amin Rakhsha, Goran Radanovic, Rati Devidze, Xiaojin Zhu, Adish Singla
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 |
|---|---|---|---|
| allow joint reward and transition poisoning (JAttack) instead of reward-only (RAttack) or transition-only (DAttack) (ablation) | offline poisoning attack · Chain and navigation environments (offline) | attack cost | Improved |
| allow manipulating target-action entries too (JAttack) instead of non-target only (NT-JAttack) (ablation) | offline poisoning attack · Chain and navigation environments (offline) | attack cost | Improved |
| increase desired epsilon margin of target policy (ablation) | offline attacks · Chain and navigation environments (offline) | attack cost / feasibility | Hurt |
| increase number of states |S| from 4 to 100 (ablation) | JAttack · Chain environment run time | run time (s): 8.35s -> 273.20s | Hurt |
| restrict to non-target only manipulations (NT-JAttack) vs full JAttack (ablation) | offline attack solving · Chain environment, |S|=100 | run time (s): 273.20s -> 1.27s | Improved |
| use NT-JAttack instead of JAttack for online attacks (ablation) | online poisoning attack · Chain and navigation environments (online) | cumulative attack cost | Improved |
| online NT-JAttack against Q-learning (gamma=0.99, constant exploration) rather than UCRL (gamma=1) (ablation) | NT-JAttack · Online attacks | AvgMiss and AvgCost | Mixed |
Closest papers by meaning
- Optimal Attack and Defense for Reinforcement Learning (AAAI 2024 · similarity 0.91)
- PolicyGuard: Towards Test-time and Step-level Adversary Defense for Reinforcement Learning Agent (ICML 2026 · similarity 0.88)
- Multi-level Certified Defense Against Poisoning Attacks in Offline Reinforcement Learning (ICLR 2025 · similarity 0.87)
- Stealthy Imitation: Reward-guided Environment-free Policy Stealing (ICML 2024 · similarity 0.87)
- Data Poisoning to Fake a Nash Equilibria for Markov Games (AAAI 2024 · similarity 0.87)
- BadRL: Sparse Targeted Backdoor Attack against Reinforcement Learning (AAAI 2024 · similarity 0.87)
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.
