TiFedCrowd: Federated Crowdsourcing With Time-Controlled Incentive

Crowdsourcing provides an effective way to collect labeled data and train models for machine learning. However, crowd workers performing data collection may expose them to the risk of privacy breaches, which lowers their willingness to participate in the tasks. To address this issue, this paper proposes a novel incentive mechanism for a federated learning paradigm, namely TiFedCrowd, which mitigates the risk of privacy breaches for crowd workers and obtains high-quality crowdsourcing data and learning models at the minimum cost. TiFedCrowd allows clients to keep privacy-related data locally, just needing to upload trained models (parameters) to the server. The server then aggregates local models into a global model. TiFedCrowd models the federated crowdsourcing process as a two-stage Stackelberg game and motivates more workers to complete tasks with high quality and efficiency. It also provides a time control mechanism to manage the quality of the submitted outcomes, the range of data freshness, and the waiting time. TiFedCrowd requires the client to complete the federated crowdsourcing task within a specified time interval while globally maximizing the utility of both clients and the server by solving the Nash equilibrium. We also extended TiFedCrowd to the multiple heterogeneous federated crowdsourcing scenario. Extensive simulation shows that TiFedCrowd not only ensures the fairness of the incentive but also has significant effects on accelerating convergence speed, selecting data/model quality, and saving the budget. Moreover, we also confirm the effectiveness of TiFedCrowd in real-world crowdsourcing tasks.