Research map / Large language models / Instruction and fine-tuning
Antibody: Strengthening Defense Against Harmful Fine-Tuning for Large Language Models via Attenuating Harmful Gradient Influence
Quoc Minh Nguyen, Trung Le, Jing Wu, Anh Tuan Bui, Mehrtash Harandi
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 flatness (sharpness) regularization in alignment (ablation) | SFT · GSM8K Llama-2-7B | harmful score: 23.94 -> 13.02 | Improved |
| add flatness (sharpness) regularization in alignment (ablation) | SFT · GSM8K Llama-2-7B | fine-tuning accuracy: 10.90 -> 16.00 | Improved |
| add weighted fine-tuning update (ablation) | SFT + flatness alignment · GSM8K Llama-2-7B | harmful score: 13.02 -> 4.44 | Improved |
| add weighted fine-tuning update (ablation) | SFT + flatness alignment · GSM8K Llama-2-7B | fine-tuning accuracy: 16.00 -> 13.83 | Hurt |
| add refusal loss in alignment (ablation) | SFT + flatness + weighted FT · GSM8K Llama-2-7B | harmful score: 4.44 -> 1.24 | Improved |
| add refusal loss in alignment (ablation) | SFT + flatness + weighted FT · GSM8K Llama-2-7B | fine-tuning accuracy: 13.83 -> 15.07 | Improved |
| replace Lisa with Antibody | Lisa · Average of 4 datasets (Llama-2-7B) | harmful score: 15.29 -> 7.04 | Improved |
| replace Booster with Antibody | Booster · Average of 4 datasets (Llama-2-7B) | fine-tuning accuracy: 65.20 -> 63.51 | Hurt |
| replace SFT with Antibody | SFT · Average over 3 models, GSM8K | harmful score: 23.51 -> 0.92 | Improved |
| replace SFT with Antibody | SFT · GSM8K harmful ratios average | harmful score: 16.53 -> 1.12 | Improved |
| replace Booster with Antibody | Booster · GSM8K Llama-2-7B, p=0.2 | harmful score: 9.06 -> 1.24 | Improved |
| replace Booster with Antibody | Booster · AlpacaEval Llama-2-7B | harmful score: 36.91 -> 24.18 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 on most shared results.
- Booster — Booster: Tackling Harmful Fine-tuning for Large Language Models via Attenuating Harmful Perturbation (ICLR 2025) · better on 3 of 6 shared results
- Vaccine — Vaccine: Perturbation-aware Alignment for Large Language Models against Harmful Fine-tuning Attack (NeurIPS 2024) · better on 6 of 6 shared results
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
- Panacea: Mitigating Harmful Fine-tuning for Large Language Models via Post-fine-tuning Perturbation (NeurIPS 2025 · similarity 0.92)
- Antidote: Post-fine-tuning Safety Alignment for Large Language Models against Harmful Fine-tuning Attack (ICML 2025 · similarity 0.90)
- Booster: Tackling Harmful Fine-tuning for Large Language Models via Attenuating Harmful Perturbation (ICLR 2025 · similarity 0.90)
- Benign Samples Matter! Fine-tuning On Outlier Benign Samples Severely Breaks Safety (ICML 2025 · similarity 0.89)
- Safe LoRA: The Silver Lining of Reducing Safety Risks when Finetuning Large Language Models (NeurIPS 2024 · similarity 0.89)
- Vaccine: Perturbation-aware Alignment for Large Language Models against Harmful Fine-tuning Attack (NeurIPS 2024 · 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.
