FEDERATED LEARNING: ADVANCEMENTS, CHALLENGES, AND APPLICATIONS

Abstract

In Federated Learning (FL), various clients cooperate to train a single model, trading minimal information among themselves instead of their actual data. Thanks to this architecture, your data is more secure, the risk of communication issues is minimised and you are able to comply with GDPR and HIPAA rules. A review of FedAvg, FedProx and FedNova is presented in this work, showing the way these methods function under IID and non-IID conditions. In healthcare, IoT and NLP, each algorithm’s performance is studied concerning its convergence, accuracy and the ease with which it can be used in practise. According to results from previous studies, FedAvg works satisfactorily when the data is identical, but it struggles where there are differences between the data. By including proximal regularisation, FedProx reduces the problem of model instability. Because of its handling of client updates, FedNova improves both fairness and synchrony, mainly under conditions when data is not IID. The tests also cover efficient communication, ability to resist attacks and fairness, with Jain’s index used and FedNova comes out on top for balance. In short, this thesis explores what FL systems do best and which features need work and it recommends future directions for research. Among them are new ways to aggregate adaptively, to use learning in real time, to update models securely and to provide personalised models. The importance of standard scales and markers is emphasised as well. The review summarises the fieldowrns and helps design AI systems that are ethical, scalable and safeguard privacy using federated concepts.