Abstract

The incentive mechanism in federated learning is a critical area for research. Establishing a fair system to incentivise data owners to share useful data is required to encourage all data owners to actively contribute their data for model training. An effective incentive system allows all participants to efficiently train models continuously, leading to improved accuracy of the final trained federated model. This paper introduces a novel algorithm for optimising the incentive mechanism. First, clients with high-quality data are able to participate in training based on their reputation value. Next, to improve the effectiveness of local training and address the issue of performance disparity among clients, the client entrusts the high-performance fog node with the training of local data by auctioning local training assignments to it. Finally, malicious clients are eliminated from the local gradient by the global gradient aggregation algorithm. Simulation results indicate that the proposed algorithm performs more effectively than existing algorithms.