Research map / Adversarial robustness and security / Backdoor and poisoning
From Poisoned to Aware: Fostering Backdoor Self-Awareness in LLMs
Guangyu Shen, Siyuan Cheng, Xiangzhe Xu, Yuan Zhou, Hanxi Guo, ZHUO ZHANG, Xiangyu Zhang
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 reversal augmentation (R-SFT) to SFT | Llama-3.1-8B and Qwen2.5-Coder-7B poisoned models · Jailbreak and sleeper agent | Awareness@k: 0 -> 0.02 and 0.042 | No clear effect |
| add GRPO RL stage after R-SFT | poisoned models · five backdoors | Awareness@5: 0.008 -> 0.5409 | Improved |
| add buffer replay to GRPO (ablation) | Llama-3.1-8B jailbreak (ILoveAppleJuice) · jailbreak backdoor | training reward: plateaus at around 0.3 without buffer | Improved |
| add R-SFT before RL (ablation) | Llama-3.1-8B · jailbreak backdoor | backdoor awareness emergence | Improved |
| unlearn with RL-articulated triggers instead of GCG triggers | Llama-3.1-8B · Jailbreak | ASR with trigger (%): 73.87 -> 4.74 | Improved |
| unlearn with RL-articulated triggers instead of R-SFT triggers | Llama-3.1-8B · Clean-label jailbreak | ASR with trigger (%): 68.34 -> 5.10 | Improved |
| unlearn with RL-articulated triggers instead of BEEAR | Llama-3.1-8B · SHIP | XSTest (%): 10.80 -> 59.20 | Mixed |
| unlearn with RL-articulated triggers instead of BEEAR | Llama-3.1-8B · DoS | MMLU-Pro (%): 11.87 -> 39.33 | Improved |
| use RL-articulated trigger guardrail instead of BEAT | Llama-3.1-8B · SHIP | detection accuracy (%): 47.8% -> 99.19% | Improved |
| use RL-articulated trigger guardrail instead of BEAT | Llama-3.1-8B · DoS | detection accuracy (%): 92.5% -> 89.00% | Hurt |
| apply RL training to other architectures (Qwen2.5-7B, Ministral-8B, DeepSeek-R1-Distill-Llama-8B) (ablation) | RL self-awareness · jailbreak SUDO | training reward | No clear effect |
Closest papers by meaning
- Simulate and Eliminate: Revoke Backdoors for Generative Large Language Models (AAAI 2025 · similarity 0.91)
- Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models (ICLR 2025 · similarity 0.89)
- ICLShield: Exploring and Mitigating In-Context Learning Backdoor Attacks (ICML 2025 · similarity 0.89)
- Demystifying Poisoning Backdoor Attacks from a Statistical Perspective (ICLR 2024 · similarity 0.89)
- BadRL: Sparse Targeted Backdoor Attack against Reinforcement Learning (AAAI 2024 · similarity 0.89)
- Universal Jailbreak Backdoors from Poisoned Human Feedback (ICLR 2024 · similarity 0.89)
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.
