Research map / Adversarial robustness and security / Backdoor and poisoning
Purifying Generative LLMs from Backdoors without Prior Knowledge or Clean Reference
Jianwei Li, Jung-Eun Kim
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 6 synthetic backdoor variants instead of 1 for signature extraction (ablation) | Backdoor signature purification (LoRA) · Targeted Refusal BadNets, LLaMA-2-7B-Chat | ASR: 40.91 -> 10.66 | Improved |
| Add cross-variant alignment term to norm-only scoring (ablation) | Backdoor signature scoring · Targeted Refusal BadNets, LLaMA-2-7B-Chat | ASR / utility: ASR 10.26 -> 10.66; utility 58.86 -> 59.42 | Mixed |
| Use norm term instead of alignment-only scoring (ablation) | Backdoor signature scoring · Targeted Refusal BadNets, LLaMA-2-7B-Chat | ASR: 77.04 -> 10.66 | Improved |
| Apply signature-guided neuron suppression plus light finetuning vs plain finetuning | LLaMA-2-7B-Chat, sentiment steering (full-model) · Sentiment Steering ASR avg | ASR: 29.96 -> 2.10 | Improved |
| Use signature-based suppression instead of CROW at standard learning rates | LLaMA-2-13B-Chat targeted refusal (full-model) · Targeted Refusal ASR avg | ASR: 68.43 -> 26.46 | Improved |
| Include attention heads in signature for Mistral (not only MLP channels) (ablation) | Mistral-7B-Instruct LoRA purification · BadNets sentiment steering | ASR | Improved |
| Zero out poisoned attention updates while keeping MLP updates (ablation) | Poisoned LLaMA-2-7B-Chat · Sanity check | backdoor persistence | No clear effect |
| Zero out poisoned MLP updates while keeping attention updates (ablation) | Poisoned LLaMA-2-7B-Chat · Sanity check | backdoor persistence | Improved |
| Pruning (magnitude) or quantization as defense | LLaMA-2-13B-Chat · MT-Bench BadNets sentiment | MT-Bench: 3.86 -> 3.02 (pruning) | Hurt |
| Plain finetuning on 200 clean samples as defense | LLaMA-2-7B-Chat sentiment (full) · BadNets ASR | ASR: 59.30 -> 60.0 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- CROW — CROW: Eliminating Backdoors from Large Language Models via Internal Consistency Regularization (ICML 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Simulate and Eliminate: Revoke Backdoors for Generative Large Language Models (AAAI 2025 · similarity 0.94)
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (NeurIPS 2025 · similarity 0.92)
- Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models (ICLR 2025 · similarity 0.91)
- Backdoor Token Unlearning: Exposing and Defending Backdoors in Pretrained Language Models (AAAI 2025 · similarity 0.90)
- BadChain: Backdoor Chain-of-Thought Prompting for Large Language Models (ICLR 2024 · similarity 0.89)
- Backdoor Cleaning without External Guidance in MLLM Fine-tuning (NeurIPS 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.
