Designing Incentive Mechanisms for Sustainable and High-Quality Data Sharing in Federated Learning

Federated learning is a promising privacy-preserving approach for data collaboration, where the continued participation of data owners is crucial for sustainability. In this paper, we propose incentive mechanisms for FL to ensure stable and high-quality data collaboration. We model the long-term data sharing problem in FL as a repeated game and design the incentive mechanisms based on fine-grained payoff construction and data characteristics. The designed incentive mechanisms provide a reasonable profit allocation to participants. We derive the boundary conditions of long-term cooperation with two Bayesian strategies, addressing the inflexibility and limitations of classic strategies. Our findings show that stable cooperation depends on various factors, including task difficulty, participants' data quality, cost sensitivity, and discount rate. Among the incentive mechanisms we provide, the marginal-improvement-based scheme proves to be the most sensitive to data quality, tending to promote cooperation among clients with high data quality. Extensive numerical simulations and case studies on data sharing are conducted to validate the theoretical analysis. Our research provides insights into stable and high-quality data sharing in various applications.