Stackelberg Game Based Local Device Participation Incentive Method for Federated Learning

This paper proposes an incentive method for local device participation in federated learning (FL). A local device participation incentive model for FL is built, and the utility functions of local devices and servers are designed. Then, the servers are modeled as leaders that provide rewards to local devices to incentivize them to train models. Local devices, as followers, contribute local data resources to participate in training. It is theoretically proved that there is a unique Nash equilibrium solution for the two-stage Stackelberg game, and the optimal reward strategy of the servers and the optimal data volume strategy of local devices are obtained. Simulation results show that the proposed local device participation incentive method proposed in can maximize the utility functions of the servers and local devices respectively, and improve the enthusiasm of local devices in FL to participate in FL.