Unsupervised Federated Learning for Face Recognition in Decentralized Environments

Recent advancements in face recognition rely on training models on a single computer, often using sensitive personal data, which raises privacy concerns. To mitigate this, researchers are exploring federated learning for unsupervised face recognition, leveraging decentralized edge devices. Each device trains the model locally and transmits only the results to a secure aggregator. To further enhance privacy, we employ GANs to generate diverse synthetic data, eliminating the need for raw data transmission. The aggregator then integrates these locally trained models into a unified global model, which is redistributed to edge devices for continuous refinement. Experiments on the CelebA dataset show that federated learning not only safeguards privacy but also maintains strong performance.