FedPSFV: Personalized Federated Learning via Prototype Sharing for Finger Vein Recognition

Finger vein recognition algorithms based on deep learning techniques are widely used in many fields. However, the training of finger vein recognition models is hindered by privacy issues and the scarcity of public datasets. Although applying federated learning techniques to finger vein recognition can effectively address privacy concerns, data heterogeneity across clients limits the performance of the models, especially on small datasets. To address these problems, in this paper, we propose a new federated finger vein recognition algorithm (FedPSFV). The algorithm is based on the federated learning framework, which increases the interclass distance of each dataset by sharing the prototypes among clients to solve the data heterogeneity problem. The algorithm also integrates and improves the margin-based loss function, which advances the feature differentiation ability of the model. Comparative experiments based on six public datasets (SDUMLA, MMCBNU, USM, UTFVP, VERA, and NUPT) show that FedPSFV has better accuracy and generalizability; the TAR@FAR = 0.01 is improved by 5.00-11.25%, and the EER is reduced by 81.48-90.22% compared to the existing methods.