Federated Learning: Tackling Heterogeneous Network Challenges in Distributed Deep Learning

Federated Learning is an emerging paradigm in distributed machine learning that allows AI models to be trained directly on user devices without the need to transfer data to a central server, providing scalability and privacy. However, large-scale networks present significant challenges, such as data heterogeneity and the potential presence of malicious or corrupted agents. This thesis explores the key algorithms and methods that address these challenges and enable versatile, robust, and secure learning. It also introduces FedGW, an innovative algorithm based on a rigorous mathematical framework that aims to identify homogeneous clusters of clients by estimating their interactions through a Gaussian reward mechanism applied to local loss functions. By exploiting the spectral properties of the client network graph, this approach uncovers latent structures in the data, allowing for more effective client clustering and improving both robustness and accuracy of training in the distributed and heterogeneous environment of Federated Learning.