IMFGR: Incentive Mechanism With Fine-Grained Reputation for Federated Learning in Mobile Crowdsensing

With the proliferation of portable devices, mobile crowdsensing has rapidly developed. However, traditional mobile crowdsensing faces problems such as single points of failure, data security, privacy concerns, insufficient participation and malicious users. To tackle these challenges, this paper proposes a novel mobile crowdsensing architecture based on federated learning and introduce blockchain to ensure the security of the crowdsensing process. Federated learning, a contemporary distributed machine learning system, facilitates the collaborative training of high-performance models across several devices while safeguarding data privacy. In federated learning tasks, users with elevated reputation values possess a better likelihood of contributing high-quality training data. To this end, we designe a fine-grained reputation evaluation scheme to characterize the behavior of these users. Additionally, to tackle the issue of insufficient participation, we designe a reverse auction model based on historical reputation to attract high-quality users to participate in tasks, thereby achieving fair reward distribution among users with different reputation levels and encouraging users to honestly participate in task execution. A series of experiments demonstrated that the proposed scheme effectively identifies malicious users, attracts high-quality users, and improves the performance of federated learning.