An efficient aggregation algorithm based on synchronous-asynchronous mechanism for federated learning

With the widespread application of Federated Learning (FL) in distributed data environments, enhancing system communication efficiency while maintaining model performance has become a critical challenge in this field. In this paper, we propose SaAS-FL, an innovative FL algorithm that aims to balance model accuracy and communication efficiency. First, it employs a synchronous training mode to obtain a relatively stable baseline global model. Then, it adopts an asynchronous update approach to achieve efficient aggregation, while introducing a delay factor based on client staleness to dynamically adjust aggregation weights, thereby mitigating the adverse effects of stale clients on model performance. Finally, an accuracy-based decision mechanism is employed to determine whether to update the global model, which avoids the distribution of ineffective models and effectively prevents model degradation. Experimental results demonstrate that SaAS-FL achieves high communication efficiency while maintaining high model accuracy, exhibiting strong robustness and adaptability across diverse heterogeneous data environments. This approach offers novel insights for enhancing FL efficiency.