Evaluation of Data Heterogeneity in FL Environment

Paradigm of federated learning (FL) suggests a practical solution to privacy preserving distributed machine learning, and one of certain benefits of FL is its capability to naturally expand the set of data on which machine learning algorithms are trained by including data with limited access which could not be made publicly available or shared to third party. However, the application of federated learning in practice still faces open challenges. One of such challenges relates to the processing of heterogeneous data owned by different parties in FL setting. This paper examines the different types of heterogeneity that arise when using federated learning and analyses methods that quantify the heterogeneity of the data.