Research map / Adversarial robustness and security / Backdoor and poisoning
DUP: Detection-guided Unlearning for Backdoor Purification in Language Models
Man Hu, Yahui Ding, Yatao Yang, Liangyu Chen, Yanhao Jia, Shuai Zhao
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 |
|---|---|---|---|
| select top-k layers by Calinski-Harabasz score instead of all layers (ablation) | MS detector · BadNets on SST-2 (BART) | AUC (%): 74.73 -> 99.27 | Improved |
| use last-half layers instead of first-half layers (ablation) | MS detector · BadNets on SST-2 (BERT) | AUC (%): 88.10 -> 99.84 | Improved |
| add spectral signature score to Mahalanobis distance (ablation) | MS detector · BadNets on SST-2 (BART) | AUC (%): 85.71 -> 99.27 | Improved |
| add spectral signature score (ablation) | MS detector · BadNets on SST-2 (LLaMA 3B) | AUC (%): 95.61 -> 94.39 | Hurt |
| add preservation loss (cross-entropy on clean samples) (ablation) | DUP · SST-2 (BERT) | CACC | Improved |
| add unlearning loss (KL divergence on detected poisoned samples) (ablation) | DUP · SST-2 (BERT) | ASR | Improved |
| increase lambda_asr from 1 to 5 (ablation) | DUP · SST-2 (BERT) | ASR | Improved |
| increase lambda_acc from 1 to 5 (ablation) | DUP · SST-2 (BERT) | CACC | Improved |
| attacker adds feature-level regularization (adaptive attack) (ablation) | DUP · Stylebkd on SST-2 (BERT) | ASR (%): 5.81 -> 12.28 | Hurt |
| attacker adds feature-level regularization (adaptive attack) (ablation) | DUP · AddSent on SST-2 (BERT) | ASR (%): 0.22 -> 0 | No clear effect |
Closest papers by meaning
- Backdoor Token Unlearning: Exposing and Defending Backdoors in Pretrained Language Models (AAAI 2025 · similarity 0.92)
- Uncovering, Explaining, and Mitigating the Superficial Safety of Backdoor Defense (NeurIPS 2024 · similarity 0.89)
- Lie Detector: Unified Backdoor Detection via Cross-Examination Framework (NeurIPS 2025 · similarity 0.88)
- Simulate and Eliminate: Revoke Backdoors for Generative Large Language Models (AAAI 2025 · similarity 0.88)
- Defense against Backdoor Attack on Pre-trained Language Models via Head Pruning and Attention Normalization (ICML 2024 · similarity 0.88)
- Backdoor Cleaning without External Guidance in MLLM Fine-tuning (NeurIPS 2025 · similarity 0.88)
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.
