Research On User Incentive Mechanism For Wireless Federated Learnin

Federated learning,as a novel distributed machine learning method,allows mobile edge devices to jointly train machine learning models without exposing their original data,avoiding uploading a large amount of original data to the cloud center,which can greatly reduce data transmission delay and save valuable communication resources.Although federated learning is a promising machine learning method,the participation of mobile devices in federated learning will consume a lot of private resources,such as computing power,communication resources and precious data resources.Considering the selfishness of users,they are not willing to sacrifice their mobile device resources to assist federated learning without rewards.The selfishness of users limits the application of federated learning in mobile networks,mobile edge computing and other scenarios.Therefore,it is necessary to design reasonable incentive mechanisms.In response to this requirement,this paper studies the following works:Firstly,one-to-many(O2M)federated learning incentive mechanism algorithm based on Stackelberg game.To address the lack of a reasonable balance mechanism between high-quality global models and users’ economic benefits in existing O2 M federated learning architectures,this paper proposes a O2 M federated learning incentive algorithm based on the Stackelberg game algorithm.The proposed algorithm models the interaction between the base station(task publisher)and multiple mobile devices(data owners)as a Stackelberg game.The base station gives a reward strategy,and the mobile device adjusts the local accuracy according to the reward.In particular,the mobile device uses the first-order optimality condition to determine its own optimal strategy,that is,the optimal local accuracy;The base station adjusts its reward strategy to maximize the benefit function according to the optimal strategy of the mobile device,and then we can obtain the Nash equilibrium.The simulation results show that proposed algorithm is significantly superior to random-ε scheme and high-ε scheme in terms of training loss and test accuracy.In addition,compared with the low-ε scheme,while providing almost the same quality of federated learning services,the proposed algorithm has improved mobile devices’ revenue/cost by 73.2%,increased BS’s revenue by12.3%,and reduced the average training delay of the overall system by 21.1%.Secondly,many-to-many(M2M)federated learning incentive mechanism algorithm based on Stackelberg game.Based on the work above,we further consider the scenario of multiple base stations and multiple mobile devices.In a M2 M federation learning architecture,it is necessary to maximize the total benefit of the system by optimizing the association relationship between base stations and users,where the benefit function of base stations is defined as a weighted sum of the benefits provided by the global model and the total reward given to mobile devices,and the benefit function of mobile devices is defined as the weighted sum of the rewards given by the base station and the total energy consumption of participating in the federated learning.To solve this problem,we propose a M2M federated learning incentive mechanism algorithm based on Stackelberg game.First,each mobile device determines its own strategy according to the benefit function.Then,each base station adjusts its own strategy according to the optimal solution of the mobile device to maximize its own benefits.The 0-1 optimization problem of multiple base stations and multiple mobile devices is converted into a linear problem and solved by relaxing the binary variables representing the correlation relationship into continuous variables,and the optimal correlation relationship is finally obtained.The simulation results show that proposed algorithm is obviously superior to random scheme in terms of training loss and test accuracy.In addition,compared with the random scheme,the algorithm increases the revenue of the base station by about 25.8% and the average revenue of the mobile device by about 293.7%.