Importance-aware data selection and resource allocation for hierarchical federated edge learning

Abstract

Compared to traditional machine learning approaches, federated learning (FL) is effective in dealing with mobile device data privacy issues. Apart from utilizing the cloud computing server as the model aggregation center, edge computing servers can also be advocated as intermediaries to perform model aggregation near the devices, which can reduce transmission latency and energy consumption. In this paper, we consider a multi-layer federated edge learning framework where both cloud and edge servers are used for FL and design a Data Importance-aware Hierarchical Federated Edge Learning (DHFL) scheme. We develop a joint data selection and resource allocation algorithm based on data importance to maximize learning efficiency in DHFL. To solve this problem, we decompose it into three sub-problems including edge-device association, resource allocation and data selection. By presenting the optimal strategy for each edge-device group, the optimal association between devices and edge servers is achieved through an iterative global cost reduction adjustment process, and data selection is performed by using convex optimization scheme. Extensive simulations are carried out to verify the proposed scheme and show that our proposal can achieve smaller training loss using less than 1/6 of the data and reduce latency by 80% compared to FedAvg.