Research map / Adversarial robustness and security / Backdoor and poisoning
TrojanTO: Action-Level Backdoor Attacks Against Trajectory Optimization Models
Yang Dai, Oubo Ma, Xingxing Liang, Longfei Zhang, Xiaochun Cao, Shouling Ji, Jiaheng Zhang, Jincai Huang, Li Shen
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 TrojanTO post-training attack instead of Baffle | TO models · D4RL avg over 6 envs, 3 target actions, 3 TO models | CP: 0.342 -> 0.701 | Improved |
| use TrojanTO instead of IMC | TO models · D4RL avg over 6 envs, 3 target actions, 3 TO models | CP: 0.551 -> 0.701 | Improved |
| add trajectory filtering (ablation) | TrojanTO · D4RL avg | BTP: 0.850 -> 0.914 | Improved |
| add batch poisoning (single poisoned transition per batch) (ablation) | TrojanTO · D4RL avg | ASR: 0.528 -> 0.719 | Improved |
| add alternating trigger/model training (ablation) | TrojanTO · D4RL avg | ASR: 0.507 -> 0.719 | Improved |
| add trajectory filtering (ablation) | TrojanTO · D4RL avg | ASR: 0.678 -> 0.719 | Mixed |
| use interior target action '0' instead of boundary '1' (ablation) | backdoor · Walk | ASR: 0.993 -> 0.110 | Hurt |
| use trigger dims (1,10,14) instead of (1,2,3) (ablation) | backdoor · Half | ASR: 0.915 -> 0.000 | Hurt |
| use learnable trigger values instead of fixed/dataset triggers (ablation) | backdoor · Half | ASR: 0.000 -> 0.557 | Improved |
| manipulate rewards of target action during backdoor training (ablation) | TO model backdoor · Walk | ASR and BTP | No clear effect |
| increase trigger noise to 10% (ablation) | TrojanTO · Walk | ASR: 0.980 -> 0.777 | Hurt |
| extend persistence of target action to 15 steps (ablation) | TrojanTO · Walk | CP: 0.993 -> 0.973 | No clear effect |
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)
- Adversarial Inception Backdoor Attacks against Reinforcement Learning (ICML 2025 · similarity 0.86)
- Adversarial Backdoor Attack by Naturalistic Data Poisoning on Trajectory Prediction in Autonomous Driving (CVPR 2024 · similarity 0.85)
- Temporal Logic-Based Multi-Vehicle Backdoor Attacks against Offline RL Agents in End-to-end Autonomous Driving (NeurIPS 2025 · similarity 0.85)
- BadVLA: Towards Backdoor Attacks on Vision-Language-Action Models via Objective-Decoupled Optimization (NeurIPS 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.
