Research map / Adversarial robustness and security / Backdoor and poisoning
Beware Untrusted Simulators -- Reward-Free Backdoor Attacks in Reinforcement Learning
Ethan Rathbun, Wo Wei Lin, Alina Oprea, Christopher Amato
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 Daze dazing instead of reward poisoning (TrojDRL-C) | Daze · Hopper (continuous, PPO) | attack success rate: 25.6% -> 94.1% | Improved |
| use Daze instead of SleeperNets-C | PPO agent · Half Cheetah | attack success rate: 3.3% -> 92.4% | Improved |
| use Daze instead of SleeperNets-C on Intersection | PPO agent · Intersection | attack success rate: 99.8% -> 92.3% | Hurt |
| use SleeperNets-C instead of Daze (benign return cost) | PPO agent · Intersection | benign return: 238.4 -> 158.6 | Hurt |
| add Daze to Atari agents vs no poisoning | PPO agent · Pacman | benign return: 628.7 -> 591.8 | No clear effect |
| add Daze vs TrojDRL on Q*bert | PPO agent · Q*bert | attack success rate: 88.2% -> 99.3% | Improved |
| use Daze vs Q-Incept on Q*bert | PPO agent · Q*bert | attack success rate: 100.0% -> 99.3% | Hurt |
| increase daze factor k from 1 to 16 (ablation) | Daze · Hopper (beta=1.0%) | attack success rate: 40.29% -> 95.56% | Improved |
| increase poisoning rate beta from 1.0% to 2.0% at k=16 (ablation) | Daze · Hopper | attack success rate: 95.56% -> 85.82% | Hurt |
| increase poisoning rate beta from 1.0% to 2.0% at k=16 (ablation) | Daze · Hopper | benign return: 2222.4 -> 1407.7 | Hurt |
| switch victim algorithm from PPO to TD3 | Daze · Hopper | attack success rate: 94.1% -> 99.9% | Improved |
| switch victim algorithm from PPO to DQN | Daze · Breakout (beta=1%) | attack success rate: 97.6% -> 90.1% | Mixed |
Closest papers by meaning
- PolicyGuard: Towards Test-time and Step-level Adversary Defense for Reinforcement Learning Agent (ICML 2026 · similarity 0.91)
- Adversarial Inception Backdoor Attacks against Reinforcement Learning (ICML 2025 · similarity 0.91)
- BadRL: Sparse Targeted Backdoor Attack against Reinforcement Learning (AAAI 2024 · similarity 0.90)
- SHINE: Shielding Backdoors in Deep Reinforcement Learning (ICML 2024 · similarity 0.89)
- Stealthy Imitation: Reward-guided Environment-free Policy Stealing (ICML 2024 · similarity 0.88)
- Optimal Attack and Defense for 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.
