AFSA-FL: A flexible semi-asynchronous federated learning framework

Federated learning has attracted significant attention in recent years, providing a secure and feasible approach for cross-institutional and cross-organizational data collaboration. Synchronous federated learning offers advantages such as easier control of convergence, simpler implementation, and better interpretability. However, it suffers from long waiting times, high stability requirements, and limited scalability. Asynchronous federated learning overcomes the issue of slow nodes in synchronous federated learning, offering higher efficiency and better scalability. However, it also introduces challenges such as complex model convergence and the risk of outdated updates. Inspired by these approaches, we propose a flexible semi-asynchronous federated learning scheme AFSA-FL to address the issues of low round efficiency and slow convergence under extreme conditions. We optimize client selection and global aggregation during the training process of the global model. This approach mitigates the impact of model staleness while ensuring model convergence. Experimental simulations demonstrate that the proposed method effectively reduces training time per round and improves the accuracy of the global model.