Affiliations: [a] School of Computer and Information, Anhui Normal University, Wuhu, Anhui, China
| [b] Anhui Provincial Key Laboratory of Network and Information Security, Wuhu, Anhui, China
Correspondence:
[*]
Corresponding author. Yonglong Luo, School of Computer and Information, Anhui Normal University, No. 189 Jiuhua South Road, Wuhu 241003, Anhui, China. Tel.: +86 0553 5910645; E-mail: [email protected].
Abstract: With the popularity of networks and the increasing number of online users, recommender systems have suffered from the privacy leakage of sensitive information. While people enjoy recommender services, their information is exposed to the networks. To protect the privacy of users when using the recommender services, we propose a multi-level combined privacy-preserving model that maintains high accuracy of recommendation with privacy protection and alleviates the data sparsity problem. Our scheme contains two steps of recommendation. First, a multi-level combined random perturbation (MCRP) model is proposed on the client side. Our model dynamically divides multiple disturbance levels and adds noise of different ranges to the rating matrix according to Gaussian and uniform mixed disturbances. Second, on the server side, we propose a pseudo rating prediction filling (PRPF) algorithm based on the matrix factorization model. Combining the PRPF algorithm with the MCRP method significantly improves the recommender accuracy and effectively increases privacy security. Sensitive analysis and comparison experiments show that the proposed privacy method has certain advantages in security and recommender accuracy by using three publicly available datasets.
Keywords: Recommender system, matrix factorization, privacy protection, random perturbation, sparse data
