Aggregation of Incentivized Learning Models in Mobile Federated Learning Environments

Abstract

We propose a reputation and budget-constrained selection methodology along with an auction-driven incentive scheme in a Federated Learning (FL) setting. The reputation score is built on the performance metrics of the local models, and the incentive aims to ensure all participants are rewarded according to the quality of their contributions. With the dynamic adjustment of the user compensation to distribute the benefits more fairly, the proposed incentive increases user utilities with the increasing platform budget. The incentive achieves positive user utility without compromising the aggregated model accuracy. Moreover, the platform utility remains consistent with respect to the baselines.