Research map / Adversarial robustness and security / Backdoor and poisoning
Broadening the Backdoor Basin: Understanding LLM Backdoors Collapse and Making Backdoors Persistent
Xingyi Zhao, Tian Xie, Xiaojun Qi, Depeng Xu, Shuhan Yuan
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 AdamW with BAD-BOOM poisoning (Fisher-ellipsoid SAM) | backdoor training · Sentiment Steering VPI, SST-2 SFT, Qwen3-0.6B | post-SFT ASR: 0.0 -> 97.0 | Improved |
| replace AdamW with BAD-BOOM poisoning | backdoor training · Sentiment Steering AddSent, SST-2 SFT, Qwen3-0.6B | post-SFT ASR: 7.3 -> 97.1 | Improved |
| replace AdamW with BAD-BOOM poisoning | backdoor training · Sentiment Steering AddSent, GSM8K SFT, Qwen3-0.6B | post-SFT ASR: 63.7 -> 93.1 | Improved |
| replace vanilla SAM with BAD-BOOM (Fisher-induced ellipsoid) (ablation) | backdoor training · Sentiment Steering AddSent, SST-2 SFT, Qwen3-0.6B | post-SFT ASR: 0.0 -> 97.1 | Improved |
| replace vanilla SAM with BAD-BOOM (ablation) | backdoor training · Sentiment Steering VPI, SST-2 SFT, Llama3.2-1B | post-SFT ASR: 0.0 -> 90.5 | Improved |
| replace AdamW with vanilla SAM | backdoor training · Sentiment Steering Sleeper, SST-2 SFT, Qwen3-0.6B | post-SFT ASR: 48.2 -> 95.4 | Improved |
| replace AdamW with vanilla SAM | backdoor training · Sentiment Steering AddSent, SST-2 SFT, Qwen3-0.6B | post-SFT ASR: 7.3 -> 0.0 | No clear effect |
| use BAD-BOOM instead of AdamW on a stronger model | backdoor training · Sentiment Steering AddSent, SST-2 SFT, Qwen3-1.7B | post-SFT ASR: 96.4 -> 97.1 | No clear effect |
| lower poisoning ratio to 1% with BAD-BOOM (ablation) | backdoor training · Sentiment Steering Sleeper, SST-2 SFT, Qwen3-0.6B | post-SFT ASR: 97.1 -> 97.1 | No clear effect |
| use BAD-BOOM instead of AdamW at 5% poisoning | backdoor training · Sentiment Steering Sleeper, SST-2 SFT, Qwen3-0.6B | post-SFT ASR: 0.1 -> 97.1 | Improved |
| raise radius rho from 0.001 to 0.005 (ablation) | BAD-BOOM · Sentiment Steering VPI, SST-2 SFT, Qwen3-0.6B | post-SFT ASR: 36.3 -> 97.0 | Improved |
| raise radius rho from 0.005 to 0.01 (ablation) | BAD-BOOM · Sentiment Steering VPI, SST-2 SFT, Qwen3-0.6B | post-SFT ASR: 97.0 -> 97.0 | No clear effect |
Closest papers by meaning
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.91)
- Simulate and Eliminate: Revoke Backdoors for Generative Large Language Models (AAAI 2025 · similarity 0.91)
- BadEdit: Backdooring Large Language Models by Model Editing (ICLR 2024 · similarity 0.90)
- Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models (ICLR 2025 · similarity 0.90)
- BadChain: Backdoor Chain-of-Thought Prompting for Large Language Models (ICLR 2024 · similarity 0.89)
- Revisiting Backdoor Attacks against Large Vision-Language Models from Domain Shift (CVPR 2025 · 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.
