GridFL: A 3D-Grid-based Federated Learning framework

Abstract

Federated Learning (FL) is an emerging distributed machine learning framework that enables a large number of devices to train machine learning models collaboratively without sharing local data. Despite the extensive potential of FL, in practical scenarios, different characteristics of clients lead to the presence of different heterogeneity in resources, data distribution, and data quantity, which poses a challenge for the training of FL. To address this problem, in this paper, we first conduct an exhaustive experimental study on all three kinds of heterogeneity in FL and provide insights into the specific impact of heterogeneity on training performance. Subsequently, we propose GridFL, a 3D-grid-based FL framework, where the three kinds of heterogeneity are defined as three dimensions (i.e., dimensions of training speed, data distribution, and data quantity) independently, and all clients in FL training are assigned to corresponding cells of the 3D grid by a gridding algorithm based on K-means clustering. In addition, we propose a grid scheduling algorithm with a dynamic selection strategy, which can select an optimal subset of clients to participate in FL training per round by adopting different strategies for different dimensions and cells. The simulation experiments show that GridFL exhibits superior performance in heterogeneous environments and outperforms several related state-of-the-art FL algorithms. Thus, the effectiveness of the proposed algorithms and strategies in GridFL are verified.