DDP-FedFV: A Dual-Decoupling Personalized Federated Learning Framework for Finger Vein Recognition.

Area of Science:

• Biometrics
• Machine Learning
• Data Privacy

Background:

• Finger vein recognition offers high accuracy for identity verification but faces privacy concerns with centralized methods.
• Federated learning (FL) addresses data privacy by training models without data sharing, but its performance suffers from dataset heterogeneity.
• Existing FL approaches struggle to balance global model generalizability with individual client model personalization.

Purpose of the Study:

• To propose a novel federated learning framework, DDP-FedFV, specifically designed for finger vein recognition.
• To enhance both the generalizability of the global model and the personalization of client models in a distributed setting.
• To address the performance limitations of federated learning caused by data heterogeneity in biometric applications.

Main Methods:

• Introduced a dual-decoupling mechanism (model and feature decoupling) to optimize feature representations and global model generalizability.
• Implemented a personalized weight aggregation method (FedPWRR) to tailor client models based on data distribution.
• Evaluated the DDP-FedFV framework using theoretical analyses and experiments on six public finger vein datasets.

Main Results:

• The DDP-FedFV framework effectively combines generalization and personalization for finger vein recognition.
• The dual-decoupling mechanism improved feature representation and global model generalizability.
• FedPWRR enhanced client model personalization by optimizing parameter aggregation based on data distribution.
• Experimental results demonstrated superior performance compared to traditional centralized training models.

Conclusions:

• The proposed DDP-FedFV framework offers a privacy-preserving and effective solution for finger vein recognition.
• The dual-decoupling and personalized aggregation strategies successfully address FL challenges in heterogeneous biometric data.
• DDP-FedFV achieves high accuracy without compromising data privacy or increasing communication overhead.