FedHSFV: Federated Learning for Finger Vein Recognition via Hierarchical Decoupling and Subspace Metric

Area of Science:

• Biometrics
• Machine Learning
• Computer Vision

Background:

• Finger vein recognition (FVR) offers high accuracy and liveness detection.
• Centralized training faces privacy challenges due to regulations.
• Federated learning (FL) addresses privacy but suffers performance degradation on Non-IID data.

Purpose of the Study:

• To address performance degradation in FL for FVR caused by dual-heterogeneity (domain shift and label skew).
• To propose a personalized federated learning framework for robust FVR.

Main Methods:

• Developed a hierarchical parameter decoupling architecture for feature extractors.
• Introduced an additive parameter decomposition with global full-rank and local low-rank adapters.
• Implemented a subspace similarity matching strategy using principal angles on the Grassmann manifold for personalized aggregation.

Main Results:

• The proposed framework significantly improves overall recognition performance.
• Effectively mitigates performance degradation caused by data heterogeneity in FVR.
• Demonstrated strong results across six public finger vein datasets.

Conclusions:

• The personalized federated learning framework effectively handles dual-heterogeneity in FVR.
• Hierarchical parameter decoupling and subspace matching enhance robustness and accuracy.
• Offers a promising solution for privacy-preserving, high-performance FVR in cross-institutional settings.