A Comparative Study of Federated Learning Methods for Face Recognition on Non-IID Dataset

This study focuses on developing facial recognition technology through deep learning. Investigating accurate models requires large and diverse datasets; however, facial data are inherently private and entail significant privacy and consent considerations. To address this issue, this research proposes a Federated Learning (FL)-based facial recognition using Non-IID datasets. We investigate how differential privacy and various facial recognition models impact the performance of FL. The Transformer-Based Face Recognition (TFR) model outperforms others by delivering higher accuracy and lower runtime. In addition, differential privacy introduces a trade-off: as the noise level increases to ensure privacy, the model's accuracy decreases.