Research map / Adversarial robustness and security / Backdoor and poisoning
Fox in the Henhouse: Supply-Chain Backdoor Attacks Against Reinforcement Learning
Shijie Liu, Andrew Craig Cullen, Paul Montague, Sarah Monazam Erfani, Benjamin I. P. Rubinstein
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 |
|---|---|---|---|
| use LSTM instead of CNN victim (ablation) | SCAB · Boxing PPO | trigger success rate: 35.0% -> 91.3% | Improved |
| use LSTM instead of CNN victim (ablation) | SCAB · Surround PPO | trigger success rate: 67.3% -> 83.2% | Improved |
| use LSTM instead of CNN victim (ablation) | SCAB · Pong PPO | trigger success rate: 34.4% -> 48.0% | Improved |
| raise testing-time trigger proportion from 0% to 30% | SCAB · Boxing LSTM PPO | victim average return: 82.6 -> 13.1 | Hurt |
| raise trigger injection probability from 1% to 5% (ablation) | SCAB · Pong CNN PPO | trigger success rate: 33.4% -> 55.4% | Improved |
| raise trigger injection probability to 3% and 5% (ablation) | SCAB · Pong CNN PPO | training-time return: 19.4 -> 15.8 (3%), 12.8 (5%) | Hurt |
| lengthen trigger to 8 no-ops (T2) with backdoor B1 (ablation) | SCAB · Surround LSTM PPO | trigger success rate: 83.2 -> 86.5 | Mixed |
| lengthen backdoor action to 8 no-ops (B2) with trigger T1 (ablation) | SCAB · Surround LSTM PPO | trigger success rate: 83.2 -> 67.4 | Hurt |
| compare SCAB with white-box BackdoorL | SCAB · Boxing LSTM PPO | trigger success rate: 93.6 -> 91.3 | No clear effect |
| fine-tune victim against normal players as defense (ablation) | SCAB defense · Surround LSTM PPO | trigger success rate: 83.2% -> 69.5% | No clear effect |
| train trigger injection timing with an RL network (ablation) | SCAB · Pong | testing trigger success rate | Hurt |
| use continuous action space (Simple Push) | SCAB · Simple Push PPO CNN, 5% TIP | trigger success rate: 83.4% | Improved |
Closest papers by meaning
- BadRL: Sparse Targeted Backdoor Attack against Reinforcement Learning (AAAI 2024 · similarity 0.88)
- PolicyGuard: Towards Test-time and Step-level Adversary Defense for Reinforcement Learning Agent (ICML 2026 · similarity 0.87)
- SHINE: Shielding Backdoors in Deep Reinforcement Learning (ICML 2024 · similarity 0.87)
- Adversarial Inception Backdoor Attacks against Reinforcement Learning (ICML 2025 · similarity 0.86)
- Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Use (ICML 2026 · similarity 0.85)
- Optimal Attack and Defense for Reinforcement Learning (AAAI 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.
