Centralized training of face recognition models requires large-scale collection of sensitive biometric data, raising serious privacy and regulatory concerns. Federated Learning (FL) enables collaborative model training across distributed clients while keeping all facial images localized on user devices. In this work, we investigate federated face recognition using ArcFace-based deep embeddings and present a comparative evaluation of three backbone architectures: AlexNet, ResNet50, and MobileNetV2. Training is performed using the FedAvg aggregation strategy over five identity-disjoint, highly non-IID client partitions of the Labeled Faces in the Wild (LFW) dataset. Each client performs local optimization for multiple epochs per communication round, and only model parameters are shared with the server. Performance is evaluated using ArcFace classifier accuracy, cosine similarity based 1 -NN identification, and verification AUC derived from ROC analysis. Experimental data show that MobileNetV2 obtains the most accurate verification area/under curve (AUC) and rapid convergence when trained utilising Federated Learning (FL), thus demonstrating robustness to data heterogeneity and communication limitations. ResNet50 produces similar embedding quality, however it does so at the cost of slower convergence times compared to MobileNetV2, owing to a higher model complexity. AlexNet, however, produces embeddings with little ability for discrimination when trained in a decentralised manner. Therefore, the findings of this study indicate that lightweight architectures and angular-margin loss functions are more effective for conducting privacy-preserving biometrics in real world federated systems environments.