Finger vein recognition methods, as emerging biometric technologies, have attracted increasing attention in identity verification due to their high accuracy and live detection capabilities. However, as privacy protection awareness increases, traditional centralized finger vein recognition algorithms face privacy and security issues. Federated learning, a distributed training method that protects data privacy without sharing data across endpoints, is gradually being promoted and applied. Nevertheless, its performance is severely limited by heterogeneity among datasets. To address these issues, this paper proposes a dual-decoupling personalized federated learning framework for finger vein recognition (DDP-FedFV). The DDP-FedFV method combines generalization and personalization. In the first stage, the DDP-FedFV method implements a dual-decoupling mechanism involving model and feature decoupling to optimize feature representations and enhance the generalizability of the global model. In the second stage, the DDP-FedFV method implements a personalized weight aggregation method, federated personalization weight ratio reduction (FedPWRR), to optimize the parameter aggregation process based on data distribution information, thereby enhancing the personalization of the client models. To evaluate the performance of the DDP-FedFV method, theoretical analyses and experiments were conducted based on six public finger vein datasets. The experimental results indicate that the proposed algorithm outperforms centralized training models without increasing communication costs or privacy leakage risks.

Finger vein recognition is a popular biometric technology. It has the characteristics of live detection and non-contact and is widely used in fields with high security requirements. Due to the influence of the environment and fingers' misplacement, the quality of images collected in this situation is poor. Therefore, the recognition accuracy is usually not ideal. To address this problem, this paper proposes a finger vein recognition algorithm based on self-calibrated convolution. First, with ResNet50 as the backbone network, a Coordinate Attention module is added after the input layer. Then, the original convolution module is replaced by the self-calibration convolution module (SCC). Finally, the joint supervision loss function is introduced to train the model. The simulation results show that, the proposed algorithm can effectively reduce the equal error rate (EER) and improve the recognition accuracy.