Research On Probabilistic Matrix Factorization Recommendation Algorithm With Differential Privacy

In recent years,with the rapid development of the Internet economy,the growth of network information has become a rapid trend,making it difficult for users to quickly filter out information of interest from massive information.Some online search engines(such as Baidu,Google,etc.)develop and adopt some special search algorithms,so that they can implement targeted search based on keywords input by users.However,sometimes the results obtained in this way cannot meet the actual needs of users to quickly obtain the required information.The recommendation system aims to push information of interest according to the user's preference,which can greatly reduce the workload of filtering a large amount of information for the user,and bring convenience to the user's life and work.The core part of the recommendation system is the recommendation algorithm,and the high-performance recommendation algorithm naturally becomes the key to building a high-quality recommendation system.However,a large amount of historical data(such as a score record,a browsing history of a web page,etc.)of the user is required in the training process of the recommendation algorithm.Through effective analysis of the user's historical data,the system can provide users with corresponding recommendation services according to the user's preferences,thereby enhancing the user's experience of the Internet and promoting the integration of the Internet economy.However,the training set of the recommendation algorithm contains the user's private information.As the privacy concept is improved,users may worry about the disclosure of their private information when enjoying the network service.For this,it is necessary to provide secure privacy protection for user data in the recommendation system.In recent years,the differential privacy technology,which has become a research hotspot,protects the user's private information by adding controllable noise,and does not change the overall pattern characteristics of the data,thereby satisfying the above-mentioned privacy protection and recommendation service requirements.This thesis mainly focuses on the differential privacy protection mechanism of the recommended scheme.The main work includes the following two aspects:(1)A recommendation scheme based on personalized differential privacy protection(PDP-PMF)is designed.Firstly,based on the probability matrix factorization recommendation algorithm and differential privacy technology,a recommendation scheme with general differential privacy protection(DP-PMF)is proposed.Furthermore,in response to the user community's need for privacy protection differentiation,this thesis develop an improved sampling mechanism,using matrix expression score data(in line with the bounded differential privacy mechanism),and propose a PDP-PMF recommendation based on personalized differential privacy protection.The scheme implements item-level privacy protection to meet the needs of users' personalized privacy protection.From a theoretical point of view,this thesis conducts a security analysis of the DP-PMF and PDP-PMF schemes,which rigorously proves that it satisfies the differential privacy mechanism.In addition,this thesis implement the above two schemes and conduct a series of experiments on the allocation of multiple user levels and privacy budgets,as well as sampling threshold optimization,compared with the DP-MF scheme and the DP-CF scheme,to verify the superiority of the PDP-PMF scheme in terms of recommendation accuracy.(2)A recommended scheme based on two-stage differential privacy protection is designed.This scheme is particularly suitable for situations where user privacy information may be compromised during the referral process and the recommender is not trusted.Considering that the existing schemes generally ignore the problem of user privacy data protection in the recommendation process,this thesis use a combined differential privacy technique and a probability matrix factorization algorithm to construct a two-stage privacy protection recommendation scheme.The scheme provides secure and effective privacy protection for users' data in the recommendation process in a scenario where the recommender is not trusted.Combining the combination theorem of differential privacy,this thesis proves the security of the scheme from a theoretical perspective.The results of multiple sets of experiments show that the accuracy of the proposed scheme is better than the relevant comparison scheme,and its safety is higher than that of the comparison scheme.