A Hierarchical Federated Aggregation Method for Heterogeneous Clients Based on Clustering

Federated learning is a collaborative machine learning framework that trains global models without obtaining private data from the client. However, traditional federated learning does not fully consider the heterogeneity of the client, resulting in poor training results and making it easy to cause model bias. In order to solve the impact of heterogeneous data on model performance, this paper proposes a hierarchical federated aggregation method for heterogeneous clients based on clustering. The edge layer uses a clustering algorithm to assign weights to heterogeneous data sets and synchronously train aggregate client data to solve the impact of client heterogeneity on model aggregation performance. The cloud layer asynchronously aggregates the responses of each edge region according to the update frequency setting power of the edge region aggregation, avoiding the influence of model bias on model performance and aggregation efficiency. Experimental results show that compared with the traditional hierarchical federated learning algorithm, the proposed algorithm has good robustness in aggregating heterogeneous data sets, achieves better training accuracy on MNIST and Fashion-MNIST data sets, reduces the training time of heterogeneous data, and improves the model aggregation efficiency.