Federated Learning : Introduction to privacy-preserving collaborative machine learning

Details

Supervisors

Van Oirbeek, Robin

Faculty

Faculté des sciences

Degree label

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

Abstract

Federated Learning (FL) has emerged as a new approach to training machine learning models across distributed data sources without the need to centralize this data. This method safeguards user privacy, addresses data transfer challenges, and provides a solution to scenarios where traditional centralized approaches falter. FL essentially embodies the strategy of "bringing the code to the data, instead of the data to the code". This thesis delves deep into the world of FL, starting with a comprehensive introduction to its structure and benefits. While the potential of FL is vast, its implementation is riddled with challenges, from understanding distributed optimization to selecting the right algorithm for the job. Our exploration focuses on benchmarking widely recognized algorithms such as FedAvg, FedProx, and FedNova, evaluating their capabilities, strengths, and potential drawbacks across various heteroneous settings. Our experimental results, based on three distinct datasets, reveal the effectiveness of federated learning as a concept. In essence, this thesis provides a foundational understanding of Federated Learning, its challenges, its benefits and its potential impact on the future of decentralized machine learning.