不训练模型，通过调整权重让视觉语言模型更安全
Safe Vision-Language Models via Unsafe Weights Manipulation
- 用安全与不安全样本对比激活值，定位关键权重并反转其值
- 在不安全输入上提升安全性，且对安全输入性能无损甚至更好
- 适合希望零训练改进模型安全性的研究者和工程师
视觉语言模型常继承大规模训练数据中的偏见和不安全关联。现有方法虽能缓解不安全行为，但评估仅关注模型在不安全输入上的表现，忽视其在安全输入上的潜在缺陷。本文提出SafeGround新评估体系，从多粒度衡量安全性，发现训练类方法反而使模型在安全输入上更不安全。基于此，我们提出无需训练的Unsafe Weights Manipulation（UWM）：利用一组包含安全与不安全样本的校准集，比较内容激活差异，识别处理不安全内容的关键参数，并通过取反操作进行调整。实验表明，UWM在安全性与知识保留间取得最佳平衡，在不安全查询上持续提升性能，且在安全输入上优于现有最先进训练方法。
原文摘要 · Abstract (English)
Vision-language models (VLMs) often inherit the biases and unsafe associations present within their large-scale training dataset. While recent approaches mitigate unsafe behaviors, their evaluation focuses on how safe the model is on unsafe inputs, ignoring potential shortcomings on safe ones. In this paper, we first revise safety evaluation by introducing SafeGround, a new set of metrics that evaluate safety at different levels of granularity. With this metric, we uncover a surprising issue of training-based methods: they make the model less safe on safe inputs. From this finding, we take a different direction and explore whether it is possible to make a model safer without training, introducing Unsafe Weights Manipulation (UWM). UWM uses a calibration set of safe and unsafe instances to compare activations between safe and unsafe content, identifying the most important parameters for processing the latter. Their values are then manipulated via negation. Experiments show that UWM achieves the best tradeoff between safety and knowledge preservation, consistently improving VLMs on unsafe queries while outperforming even training-based state-of-the-art methods on safe ones.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。