Auction-Based Incentive Mechanism with Personalized Privacy Protection in Federated Learning

Abstract

Differential privacy in federated learning (DP-FL) aims to prevent the leakage of sensitive information through client-uploaded model parameters. However, existing DP-FL frameworks typically enforce uniform privacy protection across all clients, neglecting individual privacy preferences and ultimately degrading the performance of the global model. To address this limitation, we propose an auction-based incentive mechanism for personalized privacy protection in federated learning (IMPP-FL). By integrating clients’ reported private information with data quality assessments, our mechanism encourages clients to truthfully disclose their privacy preferences and maintain high model quality. This, in turn, enhances the overall performance of federated learning. Theoretically, we prove that our mechanism satisfies dominant strategy incentive compatibility, budget constraints, and individual rationality, enabling clients to reveal their minimum acceptable privacy budgets, data volumes, and training costs. Extensive experiments on MNIST, Fashion-MNIST, and CIFAR-10 (IID & Non-IID) show that, under a total privacy budget of 10,000, IMPP-FL achieves up to 14% higher test accuracy than UPSM-FL, delivers up to 4% improvement over KGM-FL, and remains within 1% of the noise-free CDSM-FL upper bound on MNIST-IID.