Research map / Adversarial robustness and security / Attacks on text and graphs
SPARD: Defending Harmful Fine-Tuning Attack via Safety Projection with Relevance–Diversity Data Selection
Shuhao Chen, Weisen Jiang, Yeqi Gong, Shengda Luo, Chengxiang Zhuo, Zang Li, James Kwok, Yu Zhang
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 trust-region radius eta_safe (ablation) | SPARD · GSM8K (Qwen-2.5-7B) | GSM8K accuracy: 81.92 -> 85.77 | Improved |
| add trust-region radius eta_safe (ablation) | SPARD · GSM8K (Qwen-2.5-7B) | average ASR (%): 5.03 -> 9.45 | Hurt |
| use Relevance-Diversity DPP instead of random selection (ablation) | SPAG · GSM8K (Qwen-2.5-7B) | average ASR (%): 15.85 -> 9.45 | Improved |
| use Relevance-Diversity DPP instead of max-quality selection (ablation) | SPAG · GSM8K (Qwen-2.5-7B) | average ASR (%): 16.51 -> 9.45 | Improved |
| add SPAG safety projection with random safe data versus Lisa (ablation) | Lisa · GSM8K (Qwen-2.5-7B) | average ASR (%): 19.12 -> 15.85 | Improved |
| add SPAG safety projection to DPP-selected data (ablation) | standard fine-tuning · GSM8K (Qwen-2.5-7B) | average ASR (%): 74.90 -> 9.45 | Improved |
| select safe data by Relevance-Diversity DPP instead of D_High-Sim with SPAG (ablation) | SPAG · GSM8K (Qwen-2.5-7B) | average ASR (%): 16.51 -> 9.45 | Improved |
| select safe data by DPP instead of D_Low-Sim with SPAG (ablation) | SPAG · GSM8K (Qwen-2.5-7B) | average ASR (%): 58.82 -> 9.45 | Improved |
| SPARD versus SFT on harmful fine-tuning | Qwen-2.5-7B-Instruct · GSM8K | average ASR (%): 87.93 -> 9.45 | Improved |
| SPARD versus SFT on harmful fine-tuning | LLaMA-3.2-3B-Instruct · GSM8K | average ASR (%): 91.36 -> 12.09 | Improved |
| SPARD versus SafeGrad | Qwen-2.5-7B-Instruct · OpenBookQA | average ASR (%): 20.63 -> 14.54 | Improved |
| full fine-tuning instead of LoRA with SPARD | SmolLM2-1.7B-Instruct · GSM8K (BeaverTails) | ASR (%): 74.0 -> 17.4 | Improved |
Closest papers by meaning
- Adaptive Defense against Harmful Fine-Tuning for Large Language Models via Bayesian Data Scheduler (NeurIPS 2025 · similarity 0.88)
- Do as I do (Safely): Mitigating Task-Specific Fine-tuning Risks in Large Language Models (ICLR 2025 · similarity 0.88)
- Panacea: Mitigating Harmful Fine-tuning for Large Language Models via Post-fine-tuning Perturbation (NeurIPS 2025 · similarity 0.88)
- Representation Noising: A Defence Mechanism Against Harmful Finetuning (NeurIPS 2024 · similarity 0.87)
- Fundamental Limitations in Pointwise Defences of LLM Finetuning APIs (NeurIPS 2025 · similarity 0.87)
- Safety Layers in Aligned Large Language Models: The Key to LLM Security (ICLR 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.
