AFL-HCS: asynchronous federated learning based on heterogeneous edge client selection

Federated learning (FL) constitutes a potent machine learning paradigm extensively applied in edge computing for training models on vast datasets. However, the challenges of data imbalance, edge dynamics, and resource constraints in edge computing pose formidable obstacles to sustaining FL efficiency. In addressing these challenges and enhancing the effectiveness of training across heterogeneous devices in unpredictable communication networks, we introduce an asynchronous federated learning framework called AFL-HCS. Within the AFL-HCS framework, client updates transmitted to the parameter server are aggregated in each epoch based on their arrival sequence at the parameter server. Furthermore, the system incorporates a cloud cache structure to store client-submitted training progress for subsequent rounds of global model updates. This mechanism optimally leverages the local progress of clients, expediting the enhancement of the global model's performance. Experimental results demonstrate that AFL-HCS has significant advantages over the original federated learning protocol. Specifically, AFL-HCS shortens the duration of federated rounds, accelerates the convergence of the global model, and improves the accuracy of the global model, even in unstable edge environments.