Research map / Adversarial robustness and security / Backdoor and poisoning
Uncovering and Aligning Anomalous Attention Heads to Defend Against NLP Backdoor Attacks
Haotian Jin, Yang Li, Haihui Fan, Lin Shen, Xiangfang Li, Bo Li
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 attention-head alignment to head-wise fine-tuning (ablation) | Fine-tuning-only defense · SST-2 BadNets, BERT-base | ASR: 53.37 -> 11.23 | Improved |
| add attention-head alignment to head-wise fine-tuning (ablation) | Fine-tuning-only defense · SST-2 BadNets, Llama2-7B | ASR: 56.17 -> 25.61 | Improved |
| add attention-head alignment to head-wise fine-tuning (ablation) | Fine-tuning-only defense · Targeted refusal Sleeper Agent, Mistral-7B | ASR: 31.09 -> 8.54 | Improved |
| add head-wise fine-tuning after alignment (ablation) | Alignment-only defense · SST-2 BadNets, BERT-base | clean accuracy: 83.39 -> 91.09 | Improved |
| add head-wise fine-tuning after alignment (ablation) | Alignment-only defense · SST-2 BadNets, Llama2-7B | clean accuracy: 80.15 -> 85.19 | Improved |
| add head-wise fine-tuning after alignment (ablation) | Alignment-only defense · SST-2 BadNets, Llama2-7B | ASR: 23.41 -> 25.61 | No clear effect |
| add head-wise fine-tuning after alignment (ablation) | Alignment-only defense · SST-2 BadNets, BERT-base | ASR: 18.36 -> 11.23 | Improved |
| use proposed defense instead of PURE | BERT-base · SST-2 HiddenKiller | ASR: 35.26 -> 15.66 | Improved |
| use proposed defense instead of GraCeFul | Llama2-7B · Sentiment steering VPI | ASR: 10.98 -> 24.57 | Hurt |
| use proposed defense instead of MuScleLoRA | Llama2-7B · Sentiment steering VPI | clean accuracy: 78.19 -> 87.13 | Improved |
| vary safety threshold tau (ablation) | Attention head classification · backdoor defense | ASR/CA | No clear effect |
| use learning rate 2e-4 for safe heads and 5e-6 for suspicious heads (ablation) | Head-wise fine-tuning · backdoor defense | ASR and CA | Improved |
Closest papers by meaning
- Defense against Backdoor Attack on Pre-trained Language Models via Head Pruning and Attention Normalization (ICML 2024 · similarity 0.91)
- Backdoor Token Unlearning: Exposing and Defending Backdoors in Pretrained Language Models (AAAI 2025 · similarity 0.90)
- Backdoor Cleaning without External Guidance in MLLM Fine-tuning (NeurIPS 2025 · similarity 0.90)
- Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models (ICLR 2025 · similarity 0.89)
- Backdoor Attacks in Token Selection of Attention Mechanism (ICML 2025 · similarity 0.89)
- BackdoorLLM: A Comprehensive Benchmark for Backdoor Attacks and Defenses on Large Language Models (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.
