Backdoor Detection Method for Vision Transformer Model Based on Attention Map Analysis

With the widespread application of deep learning models across various domains, concerns regarding data privacy and security have become increasingly prominent. In particular, the risk of maliciously tampering with data to implant backdoors during the model training process poses a significant security threat. Although Transformers are widely used, research on detecting backdoors in these models remains insufficient. This paper proposes an attention map-based backdoor detection method for vision Transformers. By extracting the model’s attention maps and transforming them into histograms, we effectively analyze the differences between clean and poisoned data. Experimental results demonstrate that the proposed method successfully identifies anomalous data across multiple datasets. The findings confirm that leveraging attention map analysis substantially enhances the accuracy of backdoor detection, offering an effective solution to mitigate data contamination in deep learning models.