Research On Key Technologies Of Data Element Value Multi-party Sharing Service Platform

Since data silos for various reasons are hindering the use of the large amounts of data needed to train machine learning models,researchers are looking for a way to complete training without having to gather all the data into a central storage point.The federated learning framework was first proposed by Google.With the development of the federated learning framework,more and more usage scenarios have begun to apply federated learning technology.However,some participants only want to profit from the provided training data and do not need to obtain the global model.The scenario,that is to say,in this case,there is currently no solution for the scenario where one party initiates joint modeling by purchasing data from other parties.At the same time,when the initiator has a large number of participants to choose from,due to the influence of subjective and objective factors such as the uncertainty and ambiguity of the conditions of the participants themselves,and the initiator’s personal preference,the initiator cannot give the choice of the participant.A reasonable judgment mechanism.Moreover,for the participants who cooperate with the initiator to complete the joint training,a reasonable mechanism is also needed to achieve a reasonable distribution of benefits.In view of the above problems,the following researches were carried out:(1)Research on the training method of security promotion tree model for the protection of the rights and interests of the initiator.On the basis of federated learning,a boosting tree RPBT is proposed to protect the rights and interests of the initiator.RPBT can ensure that the participants will not obtain the final joint training model at the end of the training,and the local data of all participants will not be leaked,ensuring the security of the participants’ data,and using the data of multiple participants to jointly train the promotion tree model,which improves the accuracy of the model.(2)Research on the optimal partner intelligent matching method for multi-party value win-win.Based on the current research of many scholars on the factors affecting the contribution of the federation,combined with the principle of index construction,the index system for partner selection is determined.Design and realize the optimal partner intelligent matching algorithm based on the gray relational evaluation model,comprehensively evaluate and sort the partners,and obtain the optimal partner list.At the same time,the Shapley value income distribution model of the basic participant contribution index is designed and implemented.In the established evaluation index value income distribution model,the participants can be fairly distributed according to the contribution degree.(3)Research on the design and implementation method of data element value multi-party sharing service platform.Based on the above-mentioned security promotion tree model oriented to the protection of the initiator’s rights and interests and the optimal partner intelligent matching method oriented to multi-party value win-win,this paper realizes the data element value multi-party sharing service platform.In the implementation process,the unified storage technology of multi-source data format and the computing architecture supporting massive data sample training are used to solve the problem of supporting multiple types of data sources and efficient training of massive data.While improving the efficiency of model building tasks and the quality of training models,the platform fairly distributes the income generated by the joint training model to different data contributors to achieve a win-win situation for multiple parties.