Coalitional Federated Learning: Improving Communication and Training on Non-IID Data With Selfish Clients

Abstract

In this work, we propose a new paradigm of Federated Learning (FL) for Internet of Things (IoT) devices called <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Coalitional Federated Learning</i> . The proposed paradigm aims to address the challenges of (1) non-independent and identically distributed (non-IID) data across clients; (2) communication overhead due to the large number of messages exchanged between the server and clients; and (3) selfish clients that seek to obtain the latest global models without efficiently contributing to the training of the FL model. Our novel paradigm consists of three main components, i.e., (1) client-to-client trust establishment mechanism that relies on subjective and objective sources to enable clients to establish credible trust relationships toward one another; (2) trust-enabled coalitional game to enable clients to autonomously form harmonious coalitions of FL trainers; and (3) coalitional federated learning in which multiple local aggregations take place at the level of each coalition to mitigate the problems of non-IID data and communication bottleneck. Extensive experiments suggest that our solution outperforms both the standard vanilla FL approach and one state-of-the-art trust-based FL approach in terms of increasing the accuracy of the global FL model and decreasing the presence of selfish devices participating in the training.