Reseach On High-performance Privacy Preservation Based On CF Recommendation Algorithm

With the advent and continuous development of the era of big data,the information overload problem on the Internet is becoming increasingly serious.In order to solve the dilemma of users on the Internet becoming increasingly difficult to quickly find the required information from massive data,Recommendation Algorithms have emerged.The main idea of the Recommendation Algorithm is to mine the correlations among users,items,users and items based on the historical behavior data of users,to predict the user's interest and generate recommendations.Because the Recommendation Algorithm mainly uses the user 's historical behavior data for calculation and recommendation,the participating users face a serious threat of privacy leakage.Differential Privacy protection technology has attracted wide attention due to its advantages of security provability,controllable security level and high privacy preservation efficiency.The advantages of Differential Privacy protection technology make it very suitable to solve the privacy protection problem of recommendation algorithm.Therefore,the research on the reasonable application of Differential Privacy protection technology in recommendation algorithm is of great significance.This paper firstly investigates the classical recommendation algorithms and the Differential Privacy technologies,and then focuses on how to apply Differential Privacy efficiently and reasonably in the Collaborative Filtering recommendation algorithm to solve its privacy problem.The main research work is as follows.(1)For the Collaborative Filtering recommendation algorithm,the direct application of Differential Privacy protection technology can easily lead to a significant decline in the recommendation performance.To solve this problem,an efficient and feasible Collaborative Filtering algorithm based on MinHash algorithm and Differential Privacy protection(MH-DP CF)is designed in this paper.This algorithm not only guarantees users'privacy security,but also gives consideration to the recommendation accuracy.The specific approach is to pre-process the initial user-item dataset to select a dense data matrix with more data records,and use MinHash algorithm to reduce the data dimension of the dense data matrix,so as to reduce the noise range generated by Differential Privacy technology and the computational cost.Then,in order to further optimize recommendation accuracy,exporting the hole neighbor items by the exponential mechanism of Differential Privacy protection during the selection of the neighbor items,rather than using the exponential mechanism to select each neighbor item for many times,so that greatly avoiding the introduction of lots of unnecessary noise.Finally,the target users are recommended based on the above neighbor sets.At the end,the MH-DP CF algorithm is proved to meet the privacy protection definition of Differential Privacy,and the simulation experiment is also carried out.The experimental results prove the feasibility and effectiveness of MH-DP CF algorithm.(2)In view of the fact that the actual recommendation server is often not trusted,this paper designs a privacy protection scheme based on agent forwarding mechanism(PPBAF).PPBAF combines the Localized Differential Privacy protection method with the agent forwarding mechanism,not only to ensure the safety of the users' privacy,but also avoid causing serious loss of recommendation accuracy.The main idea of the PPBAF solution is to firstly select an agent from the user group containing the users who request recommendation service at the same time.Then,ordinary users use the MinHash algorithm to process their respective data and send them to the agent user.The agent user fi rstly uses localized Differential Privacy technology to disturb his own data,then hashes his disturbed data by MinHash,and finally mixes his data into the data set of all users and uniformly sends it to the recommendation server to request the recommendation service.In order to improve the accuracy of neighbors selected,a comparison protocol is designed on the recommendation server to obtain neighbors.In the whole recommendation process of PPBAF scheme,users do not interact with the recommendation server directly,but interact with the recommendation server through agents,which reduces the risk of user privacy disclosure.Finally,the simulation results prove that the PPBAF scheme is feasible,and the privacy analysis also proves that the scheme is safe.