Abstract:Federated learning is a promising solution to ensure privacy preservation while solving data silos issue, it however suf-fers from severe statistical heterogeneity of data since it does not need to transfer the private end-user data to a centralized cloud server for training. Existing techniques to alleviate statistical het-erogeneity still lack analysis of the nature and degree of statistical heterogeneity in federated tasks. In this paper, we propose Four-Dimensional Quantitative Measure, which quantitatively analyzes several typical statistical heterogeneity problem types common in federated learning image tasks with a set of quantitative strategies. Simulations show that the quantitative strategy can clearly and quantitatively reflect the type and degree of statistical heterogeneity, and can contribute the design of optimal strategies for federated learning algorithms.

FDQM: Four-Dimensional Quantitative Measure for Statistical Heterogeneity in Federated Learning