Research map / Adversarial robustness and security / Backdoor and poisoning
Don't Shift the Trigger: Robust Gradient Ascent for Backdoor Unlearning
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 GA with RGA | GA · SST-2 AddSent, BERT | dPACC: 39.32 -> 4.55 | Improved |
| replace NPO with RGA | NPO · SST-2 AddSent, BERT | dPACC: 38.97 -> 4.55 | Improved |
| replace GA with RGA | GA · SST-2 AddSent, Llama2 (7B) | dPACC: 43.86 -> 2.71 | Improved |
| replace NPO with RGA | NPO · SST-2 BadNets, Llama2 (7B) | dPACC: 1.10 -> 5.20 | Hurt |
| replace NPO with RGA | NPO · SST-2 HiddenKiller, Llama2 (7B) | dPACC: 12.10 -> 20.39 | Hurt |
| add the regularization term (term iii) to DGA (ablation) | DGA · SST-2 AddSent, Llama2 (7B, 10 epochs) | dPACC: 26.18 -> 4.16 | Improved |
| add the regularization term (term iii) to DGA (ablation) | DGA · SST-2 HiddenKiller, Llama2 (7B, 10 epochs) | dPACC: 2.23 -> 13.16 | Hurt |
| increase alpha from 1 to 4 (ablation) | RGA · SST-2 HiddenKiller, Llama2 (7B) | dPACC: 27.79 -> 13.02 | Improved |
| increase alpha from 1 to 4 (ablation) | RGA · SST-2 AddSent, Llama2 (7B) | dPACC: 2.09 -> 4.17 | Hurt |
| increase alpha from 1 to 4 (ablation) | RGA · AG-News HiddenKiller, Llama2 (7B) | dPACC: 1.60 -> 12.50 | Hurt |
| use RGA instead of ReTrain for removal | ReTrain · SST-2 BadNets, Llama2 (7B) | time per epoch (s): 403 -> 39 | Improved |
Closest papers by meaning
- Unveiling and Mitigating Backdoor Vulnerabilities based on Unlearning Weight Changes and Backdoor Activeness (NeurIPS 2024 · similarity 0.89)
- Backdoor Attacks via Machine Unlearning (AAAI 2024 · similarity 0.87)
- Backdoor Token Unlearning: Exposing and Defending Backdoors in Pretrained Language Models (AAAI 2025 · similarity 0.87)
- LOTUS: Evasive and Resilient Backdoor Attacks through Sub-Partitioning (CVPR 2024 · similarity 0.87)
- Towards LLM Unlearning Resilient to Relearning Attacks: A Sharpness-Aware Minimization Perspective and Beyond (ICML 2025 · similarity 0.87)
- Backdoor Mitigation via Invertible Pruning Masks (NeurIPS 2025 · similarity 0.87)
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.
