QSAFL: A Quality-Based Semi-Asynchronous Federated Learning Strategy

Federated learning (FL) addresses high data upload costs and privacy issues of centralized machine learning by en-abling distributed model training, which is a hot topic in wireless communication. Among FL strategies, synchronous strategy is subject to the slowest client, and asynchronous strategy faces challenges in global model convergence. Semi-asynchronous FL strategy balances client idle time and global accuracy. However, non-independent and identically distributed (Non-lID) samples among clients further complicates model aggregation. This paper introduces Quality-Based Semi-Asynchronous Federated Learning (QSAFL), a strategy mitigating two common issues in FL within the field of wireless communication: straggler effect and global accuracy problem caused by Non-lID samples. In QSAFL, the client requests a server connection immediately after finishing training. If the server is idle, it connects; otherwise, the client waits. Once connected, clients upload their models, after that, the server aggregates global model and distributes it back. The server then disconnects and becomes idle, ready for new connections. In the meantime, sample size and uniformity of distribution, evaluated by Gini index, are used to weight the client local model during aggregation. This ensures high-quality models predominate, enhancing global accuracy. According to the experiment results, QSAFL can reduce the overall time by 80 % and decrease clients' idle time by 83 % to reach the target global accuracy in Non-lID samples compared with traditional FedAvg. Within the same time frame, QSAFL's global accuracy is 10% higher than that of FedAvg.