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

Area of Science:

• Biometrics and Pattern Recognition
• Machine Learning and Artificial Intelligence

Background:

• Deep learning for finger vein recognition faces challenges due to data privacy concerns and limited public datasets.
• Federated learning (FL) offers a solution for privacy but struggles with data heterogeneity across clients, impacting model performance, especially with small datasets.

Purpose of the Study:

• To propose a novel federated finger vein recognition algorithm (FedPSFV) that overcomes data heterogeneity and enhances model performance.
• To improve feature differentiation and interclass distance within the federated learning framework for finger vein recognition.

Main Methods:

• FedPSFV utilizes a federated learning framework incorporating prototype sharing among clients to increase interclass distance.
• The algorithm integrates an improved margin-based loss function to enhance the model's feature differentiation capabilities.

Main Results:

• Comparative experiments on six public datasets (SDUMLA, MMCBNU, USM, UTFVP, VERA, NUPT) demonstrate FedPSFV's superior accuracy and generalizability.
• FedPSFV achieved a 5.00-11.25% improvement in TAR@FAR=0.01 and an 81.48-90.22% reduction in EER compared to existing methods.

Conclusions:

• FedPSFV effectively addresses data heterogeneity in federated finger vein recognition through prototype sharing and improved loss functions.
• The proposed algorithm significantly enhances recognition accuracy and generalizability, offering a promising solution for privacy-preserving biometric identification.