Research map / Adversarial robustness and security / Backdoor and poisoning
Toward Subspace-Perturbed Trajectory-Aware Backdoor Attacks in Deep Reinforcement Learning
Yaguan Qian, Taining Zhang, Qiqi Bao, Yanru Guo, Lufang Zhang, Zhaoquan Gu, Shouling Ji, Bin Wang, Zhen Lei
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 random sampling with value-guided (RTG + TD-error) strategic sampling of poisoned steps (ablation) | SpecDRL · Pong, Breakout, Atlantis (early training) | ASR | Improved |
| Use only RTG or only TD-error selection instead of the combined dual filter (ablation) | SpecDRL · Pong, Breakout, Atlantis (early training) | ASR | Improved |
| Embed triggers in PCA least-variance subspace instead of pixel patch | SpecDRL · Neural Cleanse detection | anomaly score: 11.57 -> 1.21 (max score vs TrojDRL) | Improved |
| Use value-guided sampling and Bellman-consistent reward with lower poison rate (0.01% vs 0.03%) | SpecDRL · Atari Breakout (100k) | ASR: 62.69% (SleeperNets) -> 99.75% | Improved |
| Low poisoning budget on Gopher and Atlantis | SpecDRL · Gopher, Atlantis (0.01% budget) | ASR: 19.19% -> 20.12% (Gopher vs SleeperNets); 28.65% -> 30.12% (Atlantis vs TrojDRL) | No clear effect |
| Apply Bellman-consistent reward poisoning with differential backfilling on benign performance | SpecDRL · Atari long training | BPI | Improved |
| Apply SpecDRL on highway merge | SpecDRL · Highway merge | ASR: 31.24% -> 39.24% | Mixed |
Closest papers by meaning
- Adversarial Inception Backdoor Attacks against Reinforcement Learning (ICML 2025 · similarity 0.88)
- BadRL: Sparse Targeted Backdoor Attack against Reinforcement Learning (AAAI 2024 · similarity 0.87)
- SHINE: Shielding Backdoors in Deep Reinforcement Learning (ICML 2024 · similarity 0.87)
- PolicyGuard: Towards Test-time and Step-level Adversary Defense for Reinforcement Learning Agent (ICML 2026 · similarity 0.86)
- Diffusion Guided Adversarial State Perturbations in Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- Multi-level Certified Defense Against Poisoning Attacks in Offline Reinforcement Learning (ICLR 2025 · 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.
