Research On Federated Learning Based Incentive Mechanisms For Crowdsourcing

With the rapid commercialization of Internet of Things(IoT)smart services,the collection of massive data in IoT service scenarios faces huge challenges.As an emerging paradigm of data perception,crowdsourcing relies on crowds to accomplish large-scale data perception and has been widely applied in smart city,smart ocean,and smart transportation scenarios.Due to the risks of user privacy leakage in crowdsourcing,Federated Learning(FL),as a distributed learning technology,is considered a promissing solution.Currently,research on crowdsourcing technology that integrates FL has received extensive attention.Neverthless,there are still technical challenges to be addressed in the design of incentive mechanisms,such as how to motivate users to actively participate in crowdsourcing tasks,provide high-quality federated learning training models,and improve federated learning training efficiency.To effectively incentivize high-quality user participation,this thesis studies federated learning based incentive mechanisms for Crowdsourcing.The relevant work is as follows:(1)The incentive mechanism of crowdsourcing based on federated learning is studied.First,the crowdsourcing architecture and workflow using federated learning are studied.Then,the basic concept of incentive mechanisms in federated learning is studied,and the user evaluation method is described in detail.Finally,the application of reinforcement learning in incentive mechanisms is studied.(2)From the perspective of improving model training quality and training efficiency,a new incentive mechanism for federated learning based on the Multi-Armed Bandit(MAB)is proposed.First,the user selection problem is modeled as a Combinatorial Multi-Armed Bandit(CMAB)problem,and user selection is combined with model training quality and reputation value.An incentive mechanism that combines reverse auctions and the Discounted-UCB(Upper Confidence Bound)algorithm is proposed to incentivize user participation.Simulation results show that the proposed incentive mechanism can effectively promote high-quality model aggregation and improve communication efficiency.(3)A continuous incentive mechanism for crowdsourcing with Two-Stage Stackelberg Game is proposed to effectively incentivize user participation in crowdsourcing tasks based on Federated Learning.First,the user selection problem is modeled as a CMAB problem,and virtual queue technology is introduced to improve fairness in user selection.Secondly,the interaction between the federated learning platform and the user is modeled as a two-stage Stackelberg game,determining the optimal pricing strategy for the platform and the optimal training strategy for the user.Finally,an algorithm based on user reputation is proposed to ensure sufficient user participation in federated learning training.Experimental simulations demonstrate the effectiveness of the proposed incentive mechanism.