Federated Learning over sets of heterogeneous devices

Abstract

The objective of this PhD thesis is to face the issue of data heterogeneity when training a Federated Learning model across several devices, or participants. There exist lot of previous approaches that aim to solve this kind of situation. In contrast with them, this dissertation starts with a bird's eye view of the whole data heterogeneity casuistry. Performing this analysis, data heterogeneity can be rigorously classified, making specific problems more approachable. Two main branches of problems are identified: the ones caused purely by data skewness across the devices, and the ones were data evolves over time. At the same time, each of these two branches can be further split according to the kind of data probability that is affected. This split is based on mathematical properties of the data probability. After classifying the data heterogeneity in such a way, two different algorithms are proposed, each of them dedicated to improve previous state-of-the-art results in one of the main branches mentioned before. On the one hand, we present an algorithm designed for handling time-evolving sets of data. On the other hand, we display an algorithm that focuses on identifying conflicting data distributions and manages to reach an FL model that obtains groundbreaking results in this kind of situations.