Explainable Vision Transformers for Vein Biometric Recognition

In order to foster trust and ensure accountability in the domain of deep learning, it is imperative to understand the rationale underlying the decisions rendered by an automated system. Explainable Artificial Intelligence (XAI), which elucidates a model's operational behavior, serves as a valuable tool in this context. Our investigation centers on vein-based biometric recognition, with a particular emphasis on scrutinizing the pivotal regions of a wristvein image for purposes of identity validation. We utilize Vision Transformers (ViTs), which autonomously identify and focus on the most relevant segments of the data through self-attention mechanisms. By employing two wrist-vein pattern datasets, namely FYO-wrist and PUT-wrist, we refine these models. Through the analysis of the attention maps generated by the trained networks, we can ascertain which regions of the vein patterns contribute to the identification process. Our study illustrates that, under open-set conditions, ViTs exhibit superior verification efficacy compared to established methods when they prioritize regions characterized by prominent vein patterns. This enhancement also augments the comprehensibility and reliability of the process, thereby promoting transparency through explainability.