Mitigating Statistical Non-Independence in Distributed Facial Recognition Systems through Personalized Federated Learning and Feature Alignment Techniques

Abstract

Statistical non-independence poses a critical challenge in federated learning systems for facial recognition, where distributed client data exhibits inherent correlations that violate fundamental independence assumptions. This paper proposes a novel framework combining personalized federated learning (PFL) with adaptive feature alignment techniques to mitigate these statistical dependencies. Our method introduces a two-stage approach: (1) local personalization that captures client-specific data distributions through adaptive model parameters, and (2) cross-client feature alignment that normalizes learned representations to a common latent space, reducing inter-client correlations. We introduce two novel evaluation metrics—Feature Independence Score (FIS) and Cross-Client Alignment Quality (CCAQ)—specifically designed to quantify statistical non-independence and feature homogeneity in distributed systems. Comprehensive experiments on VGGFace2 and MS-Celeb-1M datasets demonstrate that our approach achieves 95.7% accuracy on the proposed framework, surpassing standard federated learning by 2.3 percentage points while reducing feature dependency by 0.892 on the FIS metric. Results validate that addressing statistical non-independence significantly improves model convergence and generalization in privacy-preserving facial recognition systems.