Research On Federated Learning Incentive Mechanism For Mobile Edge Network

With the rapid development of wireless communication technology and the continuous enhancement of the functionalities of smart devices,the era of the Internet of Things has arrived.Various application scenarios are constantly enriching people’s lives,the number of terminal devices in the network and the amount of data generated by devices are also growing explosively.In order to make up for the deficiencies of traditional network architectures such as centralized mobile cloud computing in Io T,mobile edge computing,as a new distributed computing paradigm,has become a hot spot in the research field of Io T.Mobile edge computing sinks the data storage,transmission,calculation and offloading of real-time services to the communication nodes at the edge of network,such as edge servers and intelligent terminal devices,so as to meet user requirements for low latency and high bandwidth.The massive number of devices and data in the mobile edge network provides fertile ground for the development and application of artificial intelligence technology,which has played an important role in promoting education,medical treatment,e-commerce and other fields.However,as data security issues and privacy leakage threats become increasingly prominent,people pay more attention to personal data and no longer easily share local data.As a typical distributed machine learning framework,federated learning has significant advantages in breaking the ”data island” formed by privacy protection.The stability and efficiency of distributed system depend on the quality and cooperation of each individual,which requires that clients should be encouraged to participate as much as possible in federated learning,and the selected clients should be high-quality and reliable.If a client is unprofitable in the process of federated learning,then it is impossible to voluntarily participate in federated learning.Therefore,the research focus of this paper is on the design of the incentive mechanism for federated learning in different scenarios in the mobile edge network,so as to realize the sustainable and efficient federated learning system.First,we propose a contribution measurement model based on update significance to ensure the contribution fairness of incentive mechanism.Not only the energy consumption of client during the training process is considered,but also the significance of client’s local update on the global model update.We measure the client’s contribution by its training rounds and the deviation of its local model from the global model.In federated learning,the performance and convergence speed of the global model are directly related to the quality of participating clients.However,clients may perform unreliable or malicious local updates.A reputation mechanism based on subjective logic and contribution model is proposed to eliminate unreliable or malicious clients in the learning system in time.Then,for the single-task federated learning,two scenarios of static mobile edge network and dynamic mobile edge network are considered respectively.In both scenarios,incentive mechanisms are designed as a two-stage decision-making model of Stakelberg game,in which the aggregator acts as leader and decides the reward given to clients,and clients act as follower and decide the level of participation.The decision-making process of clients is designed as a non-cooperative game,the existence and uniqueness of the Nash equilibrium of the game are also proved,and a specific algorithm description is given.In dynamic scenarios,the set of clients may be time-varying,and we sink the granularity of incentives into the local update of clients.Therefore,in the dynamic incentive mechanism,the client’s decision is the number of rounds of local training in this round of global update.In the static incentive mechanism,client decides the number of rounds to participate in the global update.Experimental simulation results show that the proposed schemes can select a high-quality set of clients,thereby accelerating the convergence of global model and improving the performance of global model,while maximizing the overall social welfare.Finally,for the multi-task federated learning,an incentive mechanism with balanced energy consumption and satisfaction is proposed,and the incentive scheme is transformed into a Stackelberg game.The alternating direction multiplier method is used for distributed solution.On the one hand,the aggregator hopes that all learning tasks can be trained by the best matching set of clients,and want to maximize the satisfaction function based on the task matching degree.On the other hand,due to the limited computing resources,the client needs to make the optimal allocation of computing resources,and want to minimize energy consumption.To balance satisfaction and energy consumption,aggregators need to incentivize clients.The experimental simulation results show that the proposed incentive mechanism has fast convergence and can effectively balance energy consumption and satisfaction,and obtain the optimal federated learning training scheme.