Research on Aggregation Strategy of Federated Learning Parameters under Non-Independent and Identically Distributed Conditions

Federated learning is an emerging distributed machine learning framework that can build a shared model under the condition that the data of each participant is not available locally. Non-IID data is one of the fundamental challenges in the application of federated learning. The existing federated learning aggregation algorithm FedAvg (Federated Averaging) will cause model accuracy and efficiency reduction in case of non-IID data. This paper proposes the grouped federated averaging algorithm GFedAvg (Grouping federated averaging), which improves the FedAvg method. The experimental result shows that the GFedAvg algorithm improves the model accuracy by 10% on average and reduces the running time by 50% compared with the FedAvg algorithm under the same conditions.