Cost-Constrained Incentive Mechanism for Federated Learning

Abstract

: Federated learning (FL) is an effective solution to address the problem of data silos, where multiple clients collaborate to train a model. Model training itself consumes significant communication and computational resources, implying that significant economic compensation needs to be paid to incentivize clients to participate. Existing incentive mechanisms primarily focus on the effectiveness of incentives while ignoring the incentive cost constraints that need to be considered in practice, leading to scenarios where achieving the same incentive effect necessitates higher costs. The incentive costs, composed of substantial training compensation and unnecessary rewards, make it difficult for task publishers to bear. We propose a cost-constrained incentive mechanism to enable high-quality task completion within the affordability of task publishers. This mechanism designs penalty terms to compel clients to submit honest bids, thereby reducing resource wastage with repeated auctions. Additionally, it utilizes an asynchronous decision-making exploration algorithm to adjust the reward weight for clients, aiming to ensure incentive effectiveness while minimizing the reward cost. Furthermore, we introduce a resource allocation optimization strategy that allocates reward based on remaining costs, transferring more incentive costs to those who contribute positively to enhancing model accuracy. Finally, we theoretically prove the feasibility of our solution and experimentally verify that our approach achieves 5% higher accuracy than previous solutions under cost constraints while maintaining the same accuracy with at least 45% cost savings. Our code is available at link.