Research map / Adversarial robustness and security / Backdoor and poisoning
Angel or Demon: Investigating the Plasticity Interventions' Impact on Backdoor Threats in Deep Reinforcement Learning
Oubo Ma, Ruixiao Lin, Yang Dai, Jiahao Chen, Chunyi Zhou, Linkang Du, Shouling Ji
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 |
|---|---|---|---|
| apply SAM during backdoor post-training | DRL backdoor attacks (TM-Post), robotic tasks · robotic tasks | ASR: 0.178 -> 0.326 | Improved |
| apply SAM during backdoor post-training | DRL agent benign performance, robotic tasks · robotic tasks TM-Post | BTP: 0.745 -> 0.814 | Improved |
| apply Weight Clipping in TM-Post | DRL backdoor attacks · TM-Post | ASR: -17.46% average | Hurt |
| apply Spectral Normalization in TM-Post | DRL backdoor attacks · TM-Post | ASR: -11.78% average | Hurt |
| apply interventions during TM-Scratch backdoor training | DRL backdoor attacks · TM-Scratch | ASR: Layer Normalization -8.84% | No clear effect |
| apply Layer Normalization in TM-Scratch | benign task performance on sparse-reward tasks · Acrobot and MountainCar | BTP: reduces BTP by over 30% | Hurt |
| apply Weight Clipping in TM-Post | benign task performance · TM-Post | BTP: -20.19% average | Hurt |
| apply SAM in TM-Post on physics control tasks | DRL backdoor attacks · physics control tasks | ASR: +26.07% relative | Improved |
| combine SAM with other interventions (Plastic, SLac, SSW) | DRL backdoor attacks (TM-Post) · robotic tasks | ASR: 0.178 -> 0.418 (SSW) | Improved |
| combine interventions (Swiss Cheese, Lac) for mitigation | DRL backdoor attacks · TM-Post | ASR | No clear effect |
| larger pathological distance among combined interventions | SAM combinations · robotic tasks TM-Post | ASR: PD 9.43 -> 18.64; ASR 0.368 -> 0.418 | Improved |
| inject backdoor during conventional training | DRL agent training · loss landscape sharpness | sharpness range: range increases by 635.22% | Hurt |
Closest papers by meaning
- SHINE: Shielding Backdoors in Deep Reinforcement Learning (ICML 2024 · similarity 0.89)
- Adversarial Inception Backdoor Attacks against Reinforcement Learning (ICML 2025 · similarity 0.88)
- 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 0.86)
- A Study of Plasticity Loss in On-Policy Deep Reinforcement Learning (NeurIPS 2024 · similarity 0.85)
- RAT: Adversarial Attacks on Deep Reinforcement Agents for Targeted Behaviors (AAAI 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.
