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

Zijie Guo, Jian Guo, Yanan Huang, Yibo Zhang, Hengyi Ren

TL;DR

This paper introduces DDP-FedFV, a novel federated learning framework for finger vein recognition that tackles the challenge of data heterogeneity across clients. By employing a dual-decoupling strategy and a personalized aggregation method, DDP-FedFV achieves superior recognition accuracy compared to existing methods, while maintaining data privacy and reducing communication costs.

What Problem Does This Paper Address?

Finger vein recognition is a highly accurate biometric technology, but traditional deep learning methods require large amounts of data, which are difficult to collect and share due to privacy concerns. Federated learning (FL) offers a solution by enabling collaborative model training without data sharing. However, existing FL methods struggle with the significant heterogeneity in finger vein datasets across different clients (e.g., varying image quality, lighting, and equipment). This heterogeneity leads to suboptimal performance and a "client drift" issue, where models trained on one client's data perform poorly on another's. This paper aims to address this by developing a personalized FL framework that can effectively handle such data variations.

Methods

The proposed DDP-FedFV framework operates in two stages:

  • Stage 1: Generalization Phase (Dual-Decoupling)
    • Model Decoupling: The model is split into a feature extractor and a classifier. Only the feature extractor parameters are shared for aggregation to improve generalizability.
  • Feature Decoupling: This is the core innovation. The feature extractor is further divided into two parts:
  • Domain-Invariant Feature Extractor: Captures general vein patterns common across datasets. These parameters are uploaded to the server.
  • Domain-Specific Feature Extractor: Captures personalized information (e.g., background noise, lighting variations) unique to each client's data. These parameters remain local.
  • This dual-decoupling aims to create a robust global model by aggregating only the common, generalizable features.
  • Stage 2: Personalization Phase (FedPWRR)
    • Personalized Weight Aggregation: A novel method called Federated Personalization Weight Ratio Reduction (FedPWRR) is introduced.
  • FedPWRR calculates personalized aggregation weights for each client based on the similarity of their domain-specific features (from Stage 1) and their dataset sizes.
  • This allows the global model to be fine-tuned to each client's specific data distribution, enhancing personalization and overall performance.

Datasets

The study utilizes six public finger vein datasets to simulate real-world heterogeneous scenarios:

  • SDUMLA-HMT
  • MMCBNU-6000
  • HKPU-FV
  • NUPT-FV
  • VERA
  • UTFVP

These datasets are treated as individual clients in the federated learning setup.

Key Insights and Ideas

  • Decoupling is Crucial for Heterogeneity: The paper strongly argues that simply aggregating all model parameters in FL is insufficient for heterogeneous biometric data. Separating generalizable features from client-specific ones is key.
  • Domain-Invariant vs. Domain-Specific Features: The distinction between these two types of features is fundamental. Domain-invariant features are essential for building a strong global model, while domain-specific features are vital for personalization.
  • Personalized Aggregation Matters: Simply having a good global model isn't enough. Tailoring the aggregation process using client similarity and data size (as done by FedPWRR) significantly boosts individual client performance.
  • Two-Stage Approach Synergizes: The generalizability achieved in the first stage provides a solid foundation for the personalization efforts in the second stage, leading to superior overall results.

Key Results

  • Superior Performance: DDP-FedFV consistently outperforms "Local" (client-independent) and "Centralized" training methods across various metrics.
  • Competitive with Centralized: The proposed FL framework achieves performance comparable to or even superior to centralized training, while preserving data privacy.
  • Effectiveness of Dual-Decoupling: Ablation studies show that the first stage alone significantly improves performance and reduces variability compared to baseline FL methods like FedPer.
  • Impact of Personalization: The second stage (FedPWRR) further enhances performance, reducing EER by 25.902% and increasing TAR@FAR=0.01 by 1.163% on average compared to the first stage alone.
  • Robustness: The "worst performance" metrics demonstrate that DDP-FedFV maintains high accuracy even for clients with significant data heterogeneity. For example, the worst EER achieved was 2.563%, a substantial improvement over other methods.

Limitations and Future Work

  • Theoretical Convergence: While theoretical convergence is proven, the practical implications and convergence speed in highly complex real-world scenarios might warrant further investigation.
  • Computational Overhead: Although the paper claims low computational overhead for FedPWRR, the dual-decoupling in the first stage might introduce some additional complexity compared to standard FL.
  • Future Work: The authors suggest exploring more advanced personalized aggregation schemes and investigating the framework's applicability to other biometric modalities. They also mention the potential for further optimization of the dual-decoupling mechanism.

Summary updated

BibTeX
@article{guo2024ddp,
  author = {Zijie Guo and Jian Guo and Yanan Huang and Yibo Zhang and Hengyi Ren},
  title = {{DDP-FedFV: A Dual-Decoupling Personalized Federated Learning Framework for Finger Vein Recognition}},
  journal = {Sensors},
  year = {2024},
  doi = {10.3390/s24154779},
}
Figure