Flexible Clustered Federated Learning for Client-Level Data Distribution Shift

Federated Learning (FL) enables the multiple participating devices to collaboratively contribute a global neural network model while keeping training data locally. Unlike centralized setting, non-IID, imbalanced (statistical heterogeneity) and distribution shifted of FL is distributed in federated network, which will increase divergences between local models model, further degrading performance. In this paper, we propose flexible clustered learning (CFL) framework named FlexCFL, 1) group clients based on similarities clients’ optimization directions for lower divergence; 2) implement an efficient newcomer device cold start mechanism scalability practicality; 3) flexibly migrate meet challenge client-level shift. FlexCFL can achieve improvements by dividing joint into groups sub-optimization strike balance accuracy communication efficiency shift environment. The convergence complexity are analyzed demonstrate FlexCFL. We also evaluate several open datasets made comparisons with related CFL frameworks. results show that significantly improve absolute test $+10.6\%$ FEMNIST compared FedAvg , notation="LaTeX">$+3.5\%$ FashionMNIST FedProx notation="LaTeX">$+8.4\%$ MNIST FeSEM notation="LaTeX">$+4.7\%$ Sentiment140 compare IFCA . experiment