Dynamic Personalized Federated Learning for Cross-Spectral Palmprint Recognition

Area of Science:

• Biometrics
• Computer Vision
• Machine Learning

Background:

• Palmprint recognition offers high accuracy, robustness, and security but faces privacy issues with centralized deep learning.
• Non-independent and identically distributed (non-IID) multi-spectral palmprint data degrades recognition performance.

Purpose of the Study:

• To propose a dynamic personalized federated learning model (DPFed-Palm) for privacy-preserving cross-spectral palmprint recognition.
• To address the challenges of data privacy and non-IID data in multi-spectral palmprint recognition.

Main Methods:

• Developed a novel combination of loss functions for effective local model training and enhanced feature representation.
• Implemented a hybrid aggregation strategy combining Federated Averaging (FedAvg) and Personalized Federated Learning (PFL).
• Introduced a dynamic weight selection strategy for optimal personalized global model selection via cross-spectral testing.

Main Results:

• DPFed-Palm demonstrated superior privacy-preserving capabilities compared to existing methods.
• The proposed model achieved enhanced recognition performance on public datasets (PolyU, IITD, CASIA).
• Experimental results validated the effectiveness of the dynamic weight selection and hybrid aggregation strategies.

Conclusions:

• DPFed-Palm effectively enhances privacy and recognition performance in cross-spectral palmprint recognition.
• The model successfully mitigates issues related to non-IID data and centralized training in federated learning settings.
• This approach offers a promising solution for secure and accurate biometric identification systems.