Using diffusion theory as a scalpel: detecting and purifying toxic dimensions caused by backdoors or bias in pretrained language models

The researchers propose the Fine-purifying method, which uses diffusion theory to study fine-tuning dynamics and detect and purify toxic dimensions injected by backdoors or bias in pretrained language models. The method locates dimensions with anomalous dynamics based on the relationship between each dimension's parameter drift and the Hessian, resets them to clean pretrained weights, and then fine-tunes on a small-scale clean dataset. The authors claim this is the first use of diffusion-theory-guided dynamics research for security defense, and experiments verify its effectiveness.