Research map / Large language models / Language, safety and interpretability
Fewer Weights, More Problems: A Practical Attack on LLM Pruning
Kazuki Egashira, Robin Staab, Thibaud Gloaguen, Mark Vero, Martin Vechev
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 |
|---|---|---|---|
| apply Wanda/SparseGPT/Magnitude pruning to the attacked (inject then repair) model | pruning-activated attack · Jailbreak, Qwen2.5-7B, Magnitude 20% | ASR (%): 9.3 -> 95.7 | Improved |
| use SparseGPT 2:4 instead of Wanda 50% pruning on content injection | attack on Llama3.1-8B · Content injection | ASR (%): 98.8 -> 4.7 | Mixed |
| increase repair set alpha_rep (ablation) | attack · Content injection | ASR before and after pruning | Hurt |
| repair 1% of parameters (alpha_rep) in jailbreak (ablation) | attack · Jailbreak | unpruned ASR | Improved |
| user calibrates SparseGPT with security-aware dataset (ablation) | defense vs attack, SparseGPT 50% · Jailbreak, Llama3.1-8B | ASR (%): 36.0 -> 0.1 | Improved |
| user calibrates Wanda with security-aware dataset (ablation) | defense vs attack, Wanda 50% · Jailbreak, Llama3.1-8B | ASR (%): 92.3 -> 93.3 | No clear effect |
| patch pruned model with exact repaired parameters (oracle) (ablation) | defense, Wanda 50% · Jailbreak, Llama3.1-8B | ASR (%): 92.3 -> 1.3 | Improved |
| patch with bottom 5% by post-attack score (practical) (ablation) | defense, Wanda 50% · Jailbreak, Qwen2.5-7B | ASR (%): 93.0 -> 75.3 | No clear effect |
Closest papers by meaning
- Exploiting LLM Quantization (NeurIPS 2024 · similarity 0.90)
- Understanding and Mitigating Token-Pruning-Induced Vulnerabilities in VLMs (ICML 2026 · similarity 0.90)
- Let LLM Tell What to Prune and How Much to Prune (ICML 2025 · similarity 0.89)
- Soft Prompt Threats: Attacking Safety Alignment and Unlearning in Open-Source LLMs through the Embedding Space (NeurIPS 2024 · similarity 0.89)
- Hidden No More: Attacking and Defending Private Third-Party LLM Inference (ICML 2025 · similarity 0.89)
- Probe Pruning: Accelerating LLMs through Dynamic Pruning via Model-Probing (ICLR 2025 · 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.
