Design Of Incentive Mechanism For Federated Learning Based On Training Quality

The large amount of data generated by Internet devices such as Internet of Things(Io T)devices and smart wearable devices has greatly promoted the development of artificial intelligence algorithms such as machine learning.In traditional centralized learning,the server can collect and use clients’ data to complete the training task.However,clients are generally reluctant to share their private data due to the growing awareness of data privacy issues.In addition,the sharing of large amounts of data will cause huge data storage and transmission pressure.In this context,Federated Learning(FL)has emerged as an alternative solution to centralized learning.Federated learning is a privacy-preserving distributed learning pattern that enables clients to train machine learning models locally without sharing their private data.After clients’ local training,the server aggregates and updates the local model.However,deploying FL over wireless networks still faces some challenges.First,the limitation of network resources means that the communication bottleneck will occur when multiple clients simultaneously transmit the model during the training process.Secondly,the heterogeneity of clients’ resources and capabilities means that the quality of the models trained by different training clients is different.Finally,the selfishness and autonomy of clients mean that they are unwilling to consume their resources to provide training services for the server.Given the above challenges,this thesis proposes two variants of federated learning incentive mechanisms,and realizes the improvement of federated learning training quality through incentivization.The work contributions of this thesis are as follows:(1)This thesis proposes a data quality function for evaluating the client’s contribution to the training outcome.First,to improve training quality,this thesis theoretically models the federated learning process from the aspects of training time and training outcome(accuracy).Next,to measure the client’s impact of local training on the model performance,it is proposed to estimate the client’s training contribution by data quality.Finally,considering the nonlinear impact of data quality on the training effect,this thesis analyzes the relationship between the training accuracy and the client’s data quality through experiments,and establishes a fitting function for the evaluation of data quality.(2)For the federated learning scenario where clients do not change during training,such as smart streetlights,cameras and other fixed-position Io T devices in the community,to achieve lightweight and efficient incentivization,this thesis proposes a reverse multi-dimensional auction incentive mechanism.To solve the multi-dimensional decision-making problem between the server and the clients and meet the requirements of the server to select trainers,the incentive mechanism based on a multi-dimensional auction process is proposed.First,the client’s multi-dimensional information is analyzed and processed,and a method for evaluating the client’s training contribution is proposed.Next,the client’s auction decision problem is defined as a maximization scoring function problem,and an iterative decomposition algorithm is proposed to reduce the complexity of solving the problem.Finally,to optimize the utilization of bandwidth resources,this thesis proposes that bandwidth allocation should be considered in the incentive mechanism design.In this thesis,the server decision-making problem is modeled as a low-complexity three-stage solution problem: client selection,bandwidth allocation,and reward determination.It is proved that the designed scheme satisfies four necessary properties and realizes lightweight and efficient incentives.The simulation results show that the proposed incentive mechanism can achieve the training accuracy close to the unlimited resources state,and the training accuracy is improved by 2.39% compared with the traditional federated average algorithm.The proposed bandwidth allocation scheme reduces the single round training time by 0.85～12.4 seconds compared with the baseline scheme.(3)To solve the problem that the server cannot choose training clients due to the clients’ change at any time(such as mobile phones and smart wearable devices whose position and battery status can change at any time),this thesis proposes a Stackelberg game incentive mechanism.First,to motivate clients with high data quality to participate in training,a reward distribution scheme that considers the data quality while weighing the computing ability of clients.Next,according to the clients’ training cost,training remuneration and the server’s payment cost,time cost and other information,the utility function of the server and the clients are established.Finally,we model the incentive process as a two-stage game problem that maximizes the utility function and obtain the Stackelberg equilibrium by reverse induction.The simulation results show that the proposed Stackelberg game incentive mechanism can help the server make a flexible tradeoff between training time and payment cost,and can encourage clients with better data to participate in training,thereby effectively reducing the training time and improving the training accuracy by 4.25%.