Overview of Federated learning, its challenges and algorithms

Details

Supervisors

Pircalabelu, Eugen

Faculty

Faculté des sciences

Degree label

Master [120] en science des données, orientation statistique, à finalité spécialisée

Abstract

At the time when we are witnessing a great rise in new technologies using data, the protection of the user data becomes imperative. This in various fields such as commerce, education. For this purpose, several new approaches have emerged, particularly one, introduced for the first time by Google in 2016, called Federated learning. Federated learning is a recent technology whose primary objective is the acquisition of high quality models while preserving data privacy. In this work, we provide an overview of the field of Federated learning, introducing its parts and types. Then we discuss about the main challenges and complexities associated with the distributed optimization such as heterogeneity of data and propose solutions to deal with them like data normalization. Afterwards, we focus on algorithms of Federated learning existing in the literature such as FedAvg, FedProx and on their main properties. Then, we inspect their implementation as well as the influence of their hyperparameters and models on the performance through simulations and applications on real data under different scenarios and using different metrics like the accuracy.