Research map / Large language models / Instruction and fine-tuning
Safety Anchor: Defending Harmful Fine-tuning via Geometric Bottlenecks
Guoxin Lu, Letian Sha, Qing Wang, Peijie Sun, ZHOU HAO, Hua Dai, Fu Xiao
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 SBR (anchor final hidden state) to harmful fine-tuning | Llama3.1-8B SFT · SST-2/AGNEWS/GSM8K/AlpacaEval | average Harmful Score: 70.70 -> 5.68 | Improved |
| add SBR | Llama3.1-8B · SST-2/AGNEWS/GSM8K/AlpacaEval | average Functional Accuracy: 78.07 -> 78.17 | No clear effect |
| use DeepAlign instead of SBR | Llama3.1-8B · SST-2 and AGNEWS | Functional Accuracy: SST-2 94.61 -> 93.12 | Hurt |
| raise poison ratio from 0.05 to 0.3 | SBR · Llama3.1-8B | Harmful Score: 4.10 -> 7.30 | No clear effect |
| increase lambda from 0 to 10 (ablation) | SBR · Llama3.1-8B | Harmful Score: 67.80 -> 5.90 | Improved |
| increase lambda from 100 to 200 (ablation) | SBR · Llama3.1-8B | Functional Accuracy: 94.15 -> 92.89 (lambda 50 -> 200) | Hurt |
| use a single safety anchor (ablation) | SBR · Llama3.1-8B | Harmful Score: 67.8 -> 6 | Improved |
| increase anchor set size to 16 or more (ablation) | SBR · Llama3.1-8B | training cost: 1.07x -> 1.18x to 1.39x | Hurt |
| regularize intermediate layer instead of final layer (ablation) | SBR · Llama3.1-8B | Harmful Score: HS 76.8 (layer 2), 55.9 (layer 30) vs 5.8 (layer 32) | Hurt |
| add SBR under backdoor trigger | Llama3.1-8B SFT · backdoor jailbreak | Harmful Score with trigger: 65.68 -> 4.70 | Improved |
| constrain parameter distance (Lisa/EWC style) | harmful fine-tuning · stress test | Harmful Score: HS 8 -> 73 at constant distance 0.90 -> 0.88 | Hurt |
| constrain representation drift (Vaccine style) | harmful fine-tuning · stress test | Harmful Score: HS 12 -> 59 at drift about 35 | Hurt |
Closest papers by meaning
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.92)
- Safety Layers in Aligned Large Language Models: The Key to LLM Security (ICLR 2025 · similarity 0.91)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.91)
- Antidote: Post-fine-tuning Safety Alignment for Large Language Models against Harmful Fine-tuning Attack (ICML 2025 · similarity 0.91)
- Safety Alignment Should be Made More Than Just a Few Tokens Deep (ICLR 2025 · similarity 0.91)
- Booster: Tackling Harmful Fine-tuning for Large Language Models via Attenuating Harmful Perturbation (ICLR 2025 · similarity 0.90)
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.
