Incentive Mechanism Design for Hierarchical Federated Learning With Selfishness Queue Stability

Abstract

The potential privacy breaches in centralized artificial intelligence model training have raised significant public concern. Hierarchical federated learning, as a technology addressing privacy and network efficiency issues, coordinates local devices using edge servers for model training and parameter updates, thereby reducing communication with central cloud servers and diminishing the risk of privacy leaks. However, in this context, the rise of node selfishness presents a significant challenge, undermining training efficiency and the quality of local models, thereby impacting the overall system’s performance. This paper addresses the issue by introducing a virtual node selfish queue to characterize dynamic selfishness, considering both training costs and rewards, and formulating the problem of maximizing model quality within the bounds of controlled node selfishness. Utilizing Lyapunov optimization, this issue is divided into two subproblems: controlling the quantity of node data and optimizing node associations. To solve these, we propose the Data Quantity Control and Client Association (DCCA) algorithm, based on the Hungarian method. This algorithm is shown to ensure boundedness, stability, and optimality in the system. Experimental results demonstrate that the DCCA algorithm enhances model quality by 8.43% and 13.83% compared to the Fmore and FedAvg algorithms, respectively.