Quality-Aware Incentive Mechanism for Efficient Federated Learning in Mobile Crowdsensing
Hui Zhang, Ning Ti, Dongdong Wang, Xinyu Du, Qin Wang, Wenchao Xia
Abstract
Federated Learning (FL), through which mobile users (MUs) optimize a shared model without revealing the private raw data, has opened up possibilities for mobile crowdsensing (MCS). However, challenges for the MCS system with FL still exist in terms of an incentive mechanism for selecting suitable MUs to participate in and motivating MUs to contribute to model training. Our objective is to design an incentive mechanism that maximizes the overall quality of model training and improves communication efficiency while operating within a limited budget. This paper addresses a practical scenario where we lack information about the quality of MUs' local training and their computation and communication capacities. Due to lacking information on MUs, we propose a novel approach that formulates the MU selection problem as a multi-armed bandit (MAB) model. Then, we propose an effective incentive scheme combining reverse auction and discounted upper confidence bound (UCB) to incentivize the MUs to participate in the FL training process. Moreover, the proposed incentive scheme improves learning performance and reduces the FL latency by jointly selecting suitable MUs in terms of the local model quality and reputation of latency. We also prove that the proposed incentive scheme provides truthfulness, individual rationality and computationally efficiency. Compared with the existing schemes, extensive simulations demonstrate that the proposed incentive scheme can promote the quality of model aggregation and enhance communication efficiency in FL.