A Stackelberg Incentive Mechanism for Wireless Federated Learning With Differential Privacy

Abstract

In recent years, data privacy and security have attracted increasing attention in the age of artificial intelligence. Although federated learning (FL) can avoid data leakage by only sharing the machine learning models, it still suffers from differential attacks which erode the privacy of data owners. In wireless networks, the inherent channel noise can be utilized for differential privacy (DP) protection. However, the problem of incentivizing mobile devices, i.e., data owners, with the demand for DP protection to complete FL tasks has received limited attention so far. In this letter, we establish a system model for DP preserving wireless federated learning and propose an incentive mechanism based on the Stackelberg game. Our theoretical proof and numerical results demonstrate that the proposed game model can achieve the Nash equilibrium and the superior performance in maximizing the server’s utility.