Joint Client Association and Resource Utilizing for Energy-Efficient Semi-Decentralized Federated Learning with Non-IID Data

Federated learning (FL), a distributed learning paradigm, has gained widespread application in wireless Internet of Things (IoT) scenarios. To accelerate the convergence of the global model, it is expected to allow more users to participate in the federation for training. However, in heterogeneous scenarios, unrestricted expansion of the FL scale, on the contrary, may reduce the training speed of the FL or even lead to the failure of the model convergence. In this paper, we explore the heterogeneity in wireless networks and propose a semi-decentralized FL architecture that includes a cluster head (CH) selection scheme, as well as resource allocation scheme and device matching method. The selection of the CH is constrained by the minimum overlapping clusters and limited radio resource blocks (RRBs). Additionally, in the resource and device allocation scheme, we reduce overall energy consumption under time constraints while ensuring FL model performance in non-IID environments. To achieve this, we propose a joint optimization problem considering IoT device scheduling and computation frequency allocation. Furthermore, this paper analyzes the low resource efficiency caused by stragglers and proposes a suboptimal solution with lower computational complexity for large-scale applications. Comparisons with several benchmark schemes confirm that our proposed algorithm outperforms them in terms of energy efficiency, model performance, and system stability.