ML Security PapersTowards Identification and Intervention of Safety-Critical Parameters in Large Language Models
Weiwei Qi 1, Zefeng Wu 1, Tianhang Zheng 1,2, Zikang Zhang 1, Xiaojun Jia 3, Zhan Qin 1,2, Kui Ren 1,2
Published on arXiv
2604.08297
Prompt Injection
OWASP LLM Top 10 — LLM01
Key Finding
SET reduces attack success rates by over 50% updating only 1% of weights in 100 iterations; SPA limits safety degradation to within 1% after 1,000-iteration instruction tuning
ESI (Expected Safety Impact)
Novel technique introduced
Ensuring Large Language Model (LLM) safety is crucial, yet the lack of a clear understanding about safety mechanisms hinders the development of precise and reliable methodologies for safety intervention across diverse tasks. To better understand and control LLM safety, we propose the Expected Safety Impact (ESI) framework for quantifying how different parameters affect LLM safety. Based on ESI, we reveal distinct safety-critical patterns across different LLM architectures: In dense LLMs, many safety-critical parameters are located in value matrices (V) and MLPs in middle layers, whereas in Mixture-of-Experts (MoE) models, they shift to the late-layer MLPs. Leveraging ESI, we further introduce two targeted intervention paradigms for safety enhancement and preservation, i.e., Safety Enhancement Tuning (SET) and Safety Preserving Adaptation (SPA). SET can align unsafe LLMs by updating only a few safety-critical parameters, effectively enhancing safety while preserving original performance. SPA safeguards well-aligned LLMs during capability-oriented intervention (e.g., instruction tuning) by preventing disruption of safety-critical weights, allowing the LLM to acquire new abilities and maintain safety capabilities. Extensive evaluations on different LLMs demonstrate that SET can reduce the attack success rates of unaligned LLMs by over 50% with only a 100-iteration update on 1% of model weights. SPA can limit the safety degradation of aligned LLMs within 1% after a 1,000-iteration instruction fine-tuning on different tasks. Our code is available at: https://github.com/ZJU-LLM-Safety/SafeWeights-ACL.
Key Contributions
- Expected Safety Impact (ESI) framework for quantifying parameter-level safety criticality in LLMs using gradient analysis and parameter standard deviation
- Safety Enhancement Tuning (SET) that aligns unsafe LLMs by updating only 1% of safety-critical weights, reducing attack success rates by 50%+ in 100 iterations
- Safety Preserving Adaptation (SPA) that maintains safety during instruction fine-tuning by protecting critical weights, limiting degradation to <1% after 1,000 iterations