A Matrix Factorization Recommendation Method With Privacy-preserving In Distributed Environment

The emergence and popularity of the Internet has brought convenience to users while generating a lot of information.In order to help people make decisions quickly,the recommendation system needs to collect the historical data of users for modeling and analysis to implement the recommendation service.Matrix factorization recommendation algorithm is one of the most popular algorithms in collaborative filtering,with high recommendation accuracy.The traditional matrix factorization method is mainly used in the centralized environment.Users need to provide their historical data to the server.For the untrusted recommendation server,there is a risk of privacy leakage.With the continuous emergence of personal information leakage,people pay more and more attention to the issue of privacy preserving,and more data owners are reluctant to provide their own data.Therefore,the matrix factorization algorithm applied in the distributed recommendation system came into existence.In order to solve the privacy problem in the distributed recommendation system,There are two main approaches.One is to protect the security of data based on data tampering.Make it meet the differential privacy protection model by adding noise,or using random disturbance method to disturb the gradient data and sent to the server for aggregation.However,these methods based on data tampering sacrifice some of the data utility to protect the privacy of the data,which will lead to a decline in the recommendation effect.The other method uses encryption to ensure the lossless calculation of data,and uses homomorphic properties to perform encryption calculations to ensure the security of data during model training without reducing the recommended accuracy.However,use homomorphic encryption has a lot of encryption and decryption processes,which reducing the time efficiency of the entire recommended algorithm.Based on existing work,the main focus is on privacy issues in matrix factorization under distributed recommendation,designing a new privacy preserving scheme can ensure that the efficiency is improved as much as possible under the premise of data privacy and security,while ensure a better recommendation effect.The main research work of this article is as follows:Firstly,the problem of user privacy being maliciously inferred due to gradient uploading and collaborative calculation in the matrix factorization recommendation algorithm in distributed scenarios is analyzed.Then,aiming at the problem that the privacy-preserving technology based on randomized disturbances During gradient descent,noise will cause excessive errors due to iterative accumulation in distributed matrix factorization.Designed a noise which can be cancel out to process the gradient data,and perform efficient privacy-preserving calculations while reducing the data loss caused by matrix factorization calculations to improve the accuracy of recommendations.This algorithm can ensure the accuracy of recommendations while protecting user privacy.At the same time,it focuses on the issues of computational complexity and communication cost in a distributed environment,Combining the actual situation,divided items into two types:sensitive and non-sensitive,designed a personalized perturbation matrix perturbation algorithm.Only need to disturb sensitive items during the training process,which improves the efficiency of the algorithm.In addition,due to the continuous iterative calculation between client and server,it will involve frequent interaction between the user and the server.Due to the large dimension of the gradient data,users interacts too frequently,and there is a problem of excessive calculation pressure on the user client.To solve the problem,this paper compresses the transmission gradient to reduce the communication overhead in the iterative process.Finally,the detailed system design is given,and compare the matrix factorization recommendation algorithm based on differential privacy and homomorphic encryption technology through experiments,evaluates and measures the performance of the system in terms of prediction error,recommendation accuracy,operating efficiency,and communication overhead,verifying the effectiveness of the algorithm.