Evaluating Coalition Stability in Federated Learning Under Voluntary Client Participation

Federated learning has become the de facto standard solution for decentralized training of machine learning models across multiple clients. However, when client participation is voluntary, the final outcome can become unpredictable, especially if data quality varies among clients. A thesis topic in this field is to investigate a game theoretic framework that models federated learning with voluntary participation as a coalition game. This determines stable client coalitions as Nash equilibria (NEs) so as to predict possible final outcomes in terms of federated model performance. The research will explore, both empirically and theoretically, the conditions under which a grand coalition of all clients (all clients wants to join in FL) constitutes a NE. In particular, it will examine how the proportion of low-quality clients impacts the overall stability of the coalition.