Research On Privacy Protection And Incentive Mechanism Of Federated Learning Based On Cloud-Edge Computing Collaboration

With the emergence of new businesses represented by intelligent driving and intelligent medicine,in order to meet the demands of such businesses on delay,privacy and computing power,the industry integrates the federal learning technology and cloud-edge computing collaborative technology to make comprehensive use of all levels of computing power in the network to protect user privacy and improve user service quality.However,in the scenario of federated learning based on cloud-edge computing collaboration,attackers can still obtain the private data information of federated members through such attack means as inference attack and reconstruction attack.In order to deal with the privacy threats mentioned above and protect the data privacy of federated learning members,differential privacy technology,as a simple and effective way of privacy protection,is often used to protect the privacy of federated learning.However,the disturbance brought by differential privacy technology will have a negative impact on the training results of federated learning.Therefore,how to balance the effect of data privacy protection and the accuracy of training results has become a big challenge in related fields.In addition,the current federated learning mechanism in the federated learning based on cloud-edge computing collaboration still faces challenges such as difficult coordination of heterogeneous devices and easy single point of failure.How to design an efficient,credible and decentralized federated learning mechanism is also an urgent problem to be solved.To solve the above problems,this thesis mainly carried out the following work:1.Aiming at the difficulty of balancing the privacy protection effect with the training result precision in the current federated learning,this thesis proposes a federated learning incentive mechanism for privacy protection under the scenario of cloud-edge computing collaboration.First,the utility functions of participating nodes and parameter servers in federated learning under collaborative cloud computing scenarios are constructed.Secondly,the existence and uniqueness of Nash equilibrium points in this game are proved by analyzing the behavioral strategies of both parties.Finally,numerical simulation and analysis are carried out.The results of numerical simulation prove that the proposed scheme not only guarantees the data privacy of the members participating in the federation learning,but also improves the accuracy of the training results.2.Aiming at the problems of poor training effect and privacy data leakage caused by device heterogeneity in federated learning under the current collaborative cloud-edge computing scenario,this thesis proposes an efficient asynchronous federated learning mechanism for privacy protection.Firstly,a federated learning mechanism based on directed acyclic graph is designed,and the training iteration process of edge nodes under this mechanism is described.Secondly,the intelligent contract is used to design the incentive mechanism of reward distribution according to node contribution to promote node participation in the learning process.Finally,experiments are designed to verify the performance of the proposed federal learning mechanism,which proves the privacy protection of the proposed federal learning mechanism.Moreover,compared with the benchmark algorithm BlockFL algorithm,the proposed algorithm has better convergence and accuracy.In summary,this thesis takes the privacy protection and incentive mechanism of federated learning based on cloud-edge computing collaboration as the research direction,providing a new idea for the design of the privacy protection mechanism and incentive mechanism of federated learning based on cloud-edge computing collaboration,and also provides theoretical support for the design and deployment of federated learning based on cloud-edge computing collaboration.