pFedFace: Personalized Federated Learning for Face Recognition

Abstract

The emerging paradigm of federated learning facilitates the collaborative training of face recognition models without compromising the privacy of local data, thus effectively addressing the growing public concerns surrounding data privacy in face recognition. However, existing federated learning approaches typically focus on generic face representations, often overlooking the personalized requirements of individual local clients. To address this issue, we introduce pFedFace, a personalized framework for federated learning face recognition. Firstly, we propose a scheme of adaptive batch normalization (BN) which duplicates each BN layer into two variants: a generic BN and a client-specific one. The generic variant is designed to achieve consensus among clients for global generalization, while the client-specific one mitigates domain gaps for local personalization. Additionally, we introduce a personalized discriminative loss that compels the local model to extract features more discriminately than the general model. Our experiments on large-scale benchmarks have demonstrated the effectiveness of the proposed pFedFace.