Attention and Gradient-Based Transformer Interpretability Method Proposed

Researchers have introduced a novel method for interpreting Transformer models by guiding gradient direction—specifically, attention direction—to achieve more comprehensive feature region interpretation and detail interpretation. The approach leverages differences in how Vision Transformer (ViT) and humans perceive images, enabling class rewriting that is nearly imperceptible to the human eye, which may pose security risks in certain scenarios. The work is published on arXiv under the title 'Transformer Interpretability from Perspective of Attention and Gradient'.