Clustered Federated Learning Based on Data Distribution

Abstract

Federated learning is a distributed machine learning framework where many clients (e.g. mobile devices or whole organizations) collaboratively train a model under the orchestration of a central server (e.g. service provider), while keeping the training data decentralized. Non-independent and identically distributed data across clients is one of the challenges in federated learning applications which leads to a decline in model accuracy and modeling efficiency. We present a clustered federated learning algorithm based on data distribution and conduct an empirical evaluation. To protect the privacy of data in each client, we apply the encrypted distance computing algorithm in data set similarity measurement. The data experiments demonstrate the approach is effective for improving the accuracy and efficiency of federated learning. The AUC values of the clustered model is about 15% higher than the conventional model while the time cost of clustered modeling is less than 1/2 of that of conventional modeling.