Storage-Aware Joint User Scheduling and Bandwidth Allocation for Federated Edge Learning

Abstract

In Federated Edge Learning (FEEL) networks, edge devices exchange the model parameters with each other to protect data privacy, instead of directly transmitting data samples. However, the learning performance may decrease due to the limited computation, communication, and storage resources. On the one hand, devices may not have sufficient storage for the redundant data samples. On the other hand, the model transmission and computation cause a large training latency. To address these issues, we develop a storage-aware user scheduling and bandwidth allocation Federated Learning (FL) algorithm with data cleansing by taking into consideration the storage resource, data influence, and channel state information. First, a data influence evaluation method is introduced by analyzing the model divergence in a communication round aroused by the data sample. Secondly, a probability-based user scheduling scheme is proposed by minimizing the weighted sum of the storage consumption, data influence, and uploading latency. Accordingly, the joint user scheduling and bandwidth allocation scheme is developed to minimize the maximum latency for local gradient uploading. Extensive experiments demonstrate that the proposed algorithm can significantly reduce the storage pressure and the training latency while improving the learning accuracy.