Privacy Preserved Federated Learning with Attention-Based Aggregation for Biometric Recognition

TL;DR

The paper tackles privacy-preserving biometric recognition under non-IID data by proposing A3-FL, a federated learning framework that employs server-side attention to adaptively weight client updates. Each client trains a Siamese-CNN to generate embeddings, while the central server computes attention scores based on update relevance, producing a global model via $W_{t+1} = \sum_i \alpha_i W_t^i$ with $\alpha_i$ derived from $e_i$. Differential privacy and secure update protocols are incorporated to maintain data confidentiality without sacrificing accuracy, yielding a peak verification accuracy of $0.8413$ on the FVC2004 fingerprint dataset, and $0.8330$ under DP. The work demonstrates improved convergence, robustness to data heterogeneity, and practical scalability for privacy-sensitive biometric recognition in distributed settings.

Abstract

Because biometric data is sensitive, centralized training poses a privacy risk, even though biometric recognition is essential for contemporary applications. Federated learning (FL), which permits decentralized training, provides a privacy-preserving substitute. Conventional FL, however, has trouble with interpretability and heterogeneous data (non-IID). In order to handle non-IID biometric data, this framework adds an attention mechanism at the central server that weights local model updates according to their significance. Differential privacy and secure update protocols safeguard data while preserving accuracy. The A3-FL framework is evaluated in this study using FVC2004 fingerprint data, with each client's features extracted using a Siamese Convolutional Neural Network (Siamese-CNN). By dynamically modifying client contributions, the attention mechanism increases the accuracy of the global model.The accuracy, convergence speed, and robustness of the A3-FL framework are superior to those of standard FL (FedAvg) and static baselines, according to experimental evaluations using fingerprint data (FVC2004). The accuracy of the attention-based approach was 0.8413, while FedAvg, Local-only, and Centralized approaches were 0.8164, 0.7664, and 0.7997, respectively. Accuracy stayed high at 0.8330 even with differential privacy. A scalable and privacy-sensitive biometric system for secure and effective recognition in dispersed environments is presented in this work.

Figures (5)

• Figure 1: Architectures of the proposed approach.
• Figure 2: ROC Curves for All Methods.
• Figure 3: Accuracy & Loss per Round.
• Figure 4: EER, FAR, FRR per Round.
• Figure 5: Accuracy Comparison Line Graph of Methods per rounds.