Contract-Theory-Based Incentive Mechanism for Federated Learning in Health CrowdSensing

Abstract

Federated learning (FL) provides an effective solution for multiparty data processing under privacy preserving, and becomes a good choice for crowd intelligence extraction in Health CrowdSensing. The quality of the local model submitted by the data holder determines the quality of the global model in FL, and the quality of the local model depends on the data quantity, data quality, and computing power of the data holder. However, in the process of model training, the data holder will inevitably spend the cost of communication and local model training, and higher quality data acquisition and higher quality local model training require higher cost. Therefore, how to motivate data holders with a large amount of high-quality data and computing power to participate in FL has become an urgent problem to be solved. This article transforms the problem of motivating data holders into an optimization problem of utility from the perspective of maximizing the utility of the data holder, establishes the incentive mechanism based on the Contract Theory, and proves that the optimal strategy set of the data holders reaches Nash Equilibrium. A large number of experiments based on public data sets of UCI and MNIST verify that the incentive mechanism can make the baseline algorithm converge faster, while resisting malicious behaviors, such as free-riding and collusive attacks. Furthermore, the data holder with a large amount of high-quality data and computing power can obtain higher revenue.