Incentive Mechanism Design for Asynchronous Federated Learning in Vehicular Networks

Abstract

The inherent stochasticity of mobile users' (MUs') request moments and data processing durations, coupled with the aggregated bandwidth constraints during peak moments, often limits the performance of synchronous federated learning (FL) systems. To overcome these challenges, this paper proposes an incentive mechanism for asynchronous federated learning (AFL) systems within a Stackelberg game framework. In this model, MUs function as both consumers of communication resources and providers of computational services. Additionally, we derive a closed-form solution for determining the optimal number of local iterations and employ Newton's method for numerical iteration to optimize the rewards for the cloud network controller (CNC). The numerical results demonstrate the effectiveness of the proposed scheme in enhancing system performance.