Federal Learning Incentive Optimization Method Based On Evolutionary Game

As an emerging machine learning technology,federated learning(Federated Learning,FL)provides a new information unlocking method for training model parameters on multiple edge devices or servers with local data.In federated learning,model users hope to obtain high-performance training models at a small cost,while participants hope to reap more benefits by completing federated learning tasks.However,when the participants train the model,they will inevitably consume their own various resources.Therefore,in the process of federated learning,not only data privacy issues should be considered,but also a certain amount of rewards should be provided for federated learning participants to compensate for their loss during training.In recent years,the development speed of federated learning has been getting faster and faster,and the scope of application has gradually expanded.Many companies and individuals have invested in the wave of federated learning.What follows is that there are often low-quality data in federated learning but malicious requests.Incentivize the participants of the income,which will destroy the fairness of the federal incentive distribution and even affect the normal operation of the federated learning.Existing research lacks consideration of such situations,and most researchers ignore the important factor that participants cannot achieve complete rationality when designing federated learning incentive methods.Therefore,this paper establishes a federated incentive model based on evolutionary games to solve the problem of incentive benefits between federated participants and model users.In this paper,three main aspects are studied.(1)To address the payoff game problem in the federation,an evolutionary game model of the federation participant-federation organizer is established in the federation learning system.By considering the imperfect rationality of both sides of the game,the payoff choice problem of the participant and the organizer is modeled by using the evolutionary game,and the optimal payoff strategies of both sides in different initial states are derived by combining the replicated dynamic equations.(2)To address the payoff game problem in federal learning,a federal participantfederal organizer evolutionary game model is developed in the federal learning system,taking into account the imperfect rationality of both sides of the game,introducing an evolutionary game to model the incentive payoff choice problem of federal participants and federal organizers,and combining replicated dynamic equations to derive the optimal payoff strategies of both sides in different initial states.(3)To evaluate the credibility of federated participants based on their behavior of claiming incentive benefits,and to propose an optimal incentive allocation method by combining the model quality evaluation results,so as to weaken the incentive benefits of false claims,improve the motivation of other participants,and enhance the attractiveness of the federation.