Differentially Private Auction for Federated Learning with Non-IID Data

Abstract

With the increase in clients’ concerns about their privacy, federated learning, as a new model of machine learning process, was proposed to help people complete learning tasks on the basis of privacy protection. But the large-scale application of federated learning depends on the extensive participation of individual clients. This motivates the incentive mechanism design to increase clients’ willingness to participate. However, the incentive mechanism should take into account non-IID issues and privacy protection of clients’ sensitive information of data distribution. These two aspects are not well studied jointly in the existing incentive mechanism design. In this paper, we propose a differentially private auction for federated learning with non-IID data. It can not only protect clients’ private information of data distribution with differential privacy but also incentivize clients with suitable data distribution to deal with non-IID issues. Finally, we prove that the designed mechanism meets the design objective through detailed theoretical analysis.