Fairness-Aware Federated Learning With Unreliable Links in Resource-Constrained Internet of Things

Abstract

Listen

Translate article

In order to make full use of the network data and guarantee user privacy simultaneously, federated learning (FL) is proposed to enable distributed intelligence for local nodes without sharing data with each other. However, in practice, due to resource limitations, traditional FL suffers from node scheduling and parameter transmission failure, which not only affects the final performance but also further reduces the fairness of the participating nodes. This article addresses the challenge and proposes an FL method to enhance the performance of FL on the basis of guaranteeing the fairness of the local nodes in a resource-constrained Internet of Things (IoT) network. Specifically, an analytical model is first constructed to characterize the performance of FL with joint considerations of node fairness, unreliable parameter transmissions as well as resource limitations. Thereafter, a statistically reweighted aggregation (SRA) scheme is proposed for parameter aggregation and the corresponding model is proved to be unbiased to that based on ideal parameter transmissions. With the knowledge of time dependency of the global model, we further extend SRA and propose a reliable SRA (RSRA) scheme. Additionally, we prove RSRA is able to achieve higher stability performance than SRA in model training. Furthermore, the convergence bound of the proposed RSRA is derived analytically, based on which an adaptive local training scheme is proposed under a given resource budget. Finally, extensive experiments are carried out with a public data set to validate the effectiveness of the proposed scheme with comparisons of other baseline schemes.