Federated Learning on Mask recognition with its local simulation and optimization

Abstract

Even in the post-pandemic era, mask wearing remains widespread. Not only are masks crucial for stopping and managing the spread of illness, but many humanities studies have also shown that people keep wearing masks also because of a variety of psychosocial-behavioural factors. It makes the development of efficient mask recognition technology crucial. Most current research focuses on centralized training. However, especially for mask recognition which involves large-scale data and privacy issues, federated training process has a much larger advantage. Therefore, this paper investigates a basic distributed training of mask recognition model with multi-client participation and central server aggregation. The model is trained on a real-world dataset under multiple model configuration combination (local epoch and client number), then model results like training accuracy, training loss, and training time under different settings are tested under several statistic tests. Finally, this paper explores some balancing strategies regarding local epochs and client numbers, reveals some interactions between two configuration argument (local epoch and client number) and proposes some locally optimal configuration combinations.