Learner Referral for Cost-Effective Federated Learning Over Hierarchical IoT Networks

Abstract

Addressing data privacy concerns, Federated Learning (FL) has been recognized for its ability to train parameters locally on resource-constrained clients in a distributed manner. However, the problem of optimization of FL client selection and resource allocation in hierarchical Internet of Things (HieIoT) networks, where clients move in and out of each others’ D2D communication coverage and no FL server knows all the data owners, remains open. To bridge this gap, we propose a learner referral aided federated client selection (LRef-FedCS) approach, complemented by communications and computing resource scheduling, along with local model accuracy optimization (LMAO). LRef-FedCS enhances cost efficiency and FL model quality by enabling data owners to share FL task details within their trusted local networks, increasing the opportunity of the FL server choosing the optimal clients. Using Lyapunov optimization, the problem is transformed into a joint optimization problem (JOP). To address the JOP’s complexities, we combine a centralized method for LRef-FedCS and the self-adaptive global best harmony search algorithm for LMAO. For enhance scalability, a distributed LRef-FedCS based on a matching game is proposed. Numerical experiments on the Fashion-MNIST dataset show LRef-FedCS outperforms existing state-of-the-art approaches, delivering enhanced model accuracy with notable cost savings.