FRIMFL: A Fair and Reliable Incentive Mechanism in Federated Learning

초록

Federated learning (FL) enables data owners to collaboratively train a machine learning model without revealing their private data and sharing the global models. Reliable and continuous client participation is essential in FL for building a high-quality global model via the aggregation of local updates from clients over many rounds. Incentive mechanisms are needed to encourage client participation, but malicious clients might provide ineffectual updates to receive rewards. Therefore, a fair and reliable incentive mechanism is needed in FL to promote the continuous participation of clients while selecting clients with high-quality data that will benefit the whole system. In this paper, we propose an FL incentive scheme based on the reverse auction and trust reputation to select reliable clients and fairly reward clients that have a limited budget. Reverse auctions provide candidate clients to bid for the task while reputations reflect their trustworthiness and reliability. Our simulation results show that the proposed scheme can accurately select users with positive contributions to the system based on reputation and data quality. Therefore, compared to the existing schemes, the proposed scheme achieves higher economic benefit encouraging higher participation, satisfies reward fairness and accuracy to promote stable FL development.

키워드