Research On User Selection Method For Vertical Federated Learning Under Feature Redundancy

With the vigorous development of artificial intelligence(AI),data privacy in AI has always been the focus of researchers.Due to the existence of isolated islands of data in different organizations,the laws of various countries on data privacy protection and other practical factors,AI technology has faced a series of key challenges in data,communication and other aspects.In recent years,the proposal of federated learning has realized data sharing and privacy protection in distributed learning,and completed the training of the model without leaving the original data.Therefore,federated learning has broad application prospects and research significance in different fields.In practical applications,the global model performance of federated learning depends on whether high-quality and reliable users can be recruited to participate in learning.Because users will inevitably bear certain costs when participating in federated learning,including data leakage risk,computing resources,communication resources and time,users may not be willing to participate in training or share their high-quality models without sufficient returns.At present,more and more research has focused on the work of user incentive strategies in the field of federal learning,aiming at recruiting active and reliable users to participate in federated learning.Most of the existing work on user selection in federated learning focuses on two directions: how to fairly provide reward schemes and how to reasonably evaluate the data value of individual users.In real scenes,the phenomenon of feature redundancy of samples between multiple users is endless,and the existence of feature redundancy seriously affects the performance of the vertical federated learning model.However,most of the existing studies in the field of vertical federated learning user selection are oriented to the selection of users in the case of non-redundant feature data,without considering how to reasonably select a group of high-quality users in the scenario of feature redundancy among users,which is not conducive to the sustainable development and implementation of vertical federated learning.Therefore,in the case of feature redundancy,selecting a group of high-quality users to perform the training of vertical federated learning model has become the key research issue of this paper.In order to solve the above problems,this paper focuses on how to measure the impact of feature redundancy between multiple users in vertical federated learning on the global model and how to select high-quality winning users to complete the learning,and proposes a two-stage user selection method.The specific content is:Aiming at the problem of data privacy protection in vertical federated learning,this paper adopts secure multi-party computing to transform the local private data set owned by users into a secret that can be shared by multiple users,and then trains the global machine learning model.Faced with the problem that it is difficult to grasp the real utility of each user on the model and the impact of the redundant features between users on the model in the early stage before the actual training of the model,this paper simulates the interaction between the server and users according to the reverse auction framework,and splits the user selection problem into two stages: the first stage of winning user selection is based on the idea of "explore and exploit" in the multi-armed bandit problem and the historical utility estimation results of a single user,and in the second stage a combined user utility estimation strategy is designed based on the historical utility of a single user and the degree of overlap of multi-user features to complete user selection based on the combined user utility.In view of the above framework and strategy,this paper provides theoretical proof.Finally,this paper evaluates our method on three real data sets and corresponding machine learning models.The experimental results show that the proposed method can achieve excellent performance compared with other benchmark methods.