Research map / Adversarial robustness and security / Backdoor and poisoning
Mitigating Safety Fallback in Editing-based Backdoor Injection on LLMs
Houcheng Jiang, Zetong Zhao, Junfeng Fang, Haokai Ma, Ruipeng Wang, Xiang Wang, Xiangnan He, Yang Deng
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 dynamic loss weighting (ablation) | DualEdit · DAN | ASR: 71.42 -> 81.51 | Improved |
| add dynamic loss weighting (ablation) | DualEdit · DAN | SFR: 36.32 -> 22.67 | Improved |
| add dynamic loss weighting (ablation) | DualEdit · DNA | ASR: 66.13 -> 77.93 | Improved |
| add dynamic loss weighting (ablation) | DualEdit · Misuse | SFR: 46.48 -> 40.22 | Improved |
| add value anchoring (ablation) | DualEdit · DAN | ASR: 75.28 -> 81.51 | Improved |
| add value anchoring (ablation) | DualEdit · DNA | SFR: 31.69 -> 27.64 | Improved |
| add both dynamic loss weighting and value anchoring (ablation) | DualEdit · DNA | ASR: 56.40 -> 77.93 | Improved |
| add both dynamic loss weighting and value anchoring (ablation) | DualEdit · DNA | SFR: 52.13 -> 27.64 | Improved |
| add DualEdit backdoor edit | DualEdit · MMLU, LLaMA-3.1-8B | accuracy: 72.95 -> 71.81 | No clear effect |
| add DualEdit backdoor edit | DualEdit · Average of 7 capability benchmarks, Qwen2.5-7B | avg score: 84.50 -> 81.54 | Hurt |
| add DualEdit backdoor edit | DualEdit · Average of 7 capability benchmarks, LLaMA-2-7B | avg score: 60.56 -> 61.01 | No clear effect |
| use DualEdit instead of JailbreakEdit | DualEdit · DAN, LLaMA-3.1-8B | ASR with trigger: 75.43% -> 88.07% | Improved |
Closest papers by meaning
- Simulate and Eliminate: Revoke Backdoors for Generative Large Language Models (AAAI 2025 · similarity 0.90)
- Improving LLM Safety Alignment with Dual-Objective Optimization (ICML 2025 · similarity 0.90)
- Probe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models (ICLR 2025 · similarity 0.90)
- BadEdit: Backdooring Large Language Models by Model Editing (ICLR 2024 · similarity 0.89)
- BackdoorAlign: Mitigating Fine-tuning based Jailbreak Attack with Backdoor Enhanced Safety Alignment (NeurIPS 2024 · similarity 0.89)
- SORRY-Bench: Systematically Evaluating Large Language Model Safety Refusal (ICLR 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.
