Storage-aware Joint User Scheduling and Spectrum Allocation for Federated Learning

Abstract

Massive data drives the development of machine learning (ML) for a long time. However, at present, data is starting to hinder ML's development. The first reason is that the privacy of data is increasingly valued by the public. Therefore, Federated Learning (FL) has emerged, which realizes model training through distributed computing and centralized aggregation. Second, due to the popularity of FL, edge devices need to store all data, which may quickly occupy the entire storage space of edge devices, resulting in fatal errors. To address these challenges, we proposed a storage-aware joint user scheduling and spectrum allocation algorithm, named FedSUS, to reduce the storage stress of each device and guarantee traditional FL metrics, i.e., learning accuracy and training latency. First, a probabilistic framework is adopted for user scheduling. Second, we introduce a data influence evaluation method to FL and analyze its convergence. Based on this, two problems are formulated to tradeoff the storage resource, the influence of data, and the learning latency and to minimize the transmission latency, respectively. Then, the closed-form results to the above problems are both developed. Finally, FedSUS is validated by using a popular convolutional neural network (CNN) and datasets (CIFAR-10). And numerical results demonstrate that our algorithm can effectively reduce the local data size while keeping (even improving) the learning accuracy as compared with baseline.