Federated learning based on stratified sampling and regularization

Abstract Federated learning (FL) is a new distributed framework that different from traditional machine learning: (1) differences in communication, computing, and storage performance among devices (device heterogeneity), (2) data distribution volume (data (3) high communication consumption. Under heterogeneous conditions, the of clients varies greatly, which leads to problem convergence speed training model decreases cannot converge global optimal solution. In this work, an FL algorithm based on stratified sampling regularization (FedSSAR) proposed. FedSSAR, density-based clustering method used divide overall client into clusters, then, some available are proportionally extracted clusters participate realizes unbiased for reduces aggregation weight variance client. At same time, when calculating local loss function, we limit update direction by regular term, so optimized globally direction. We prove FedSSAR theoretically experimentally, demonstrate superiority comparing it with other algorithms public datasets.