Personalized Privacy Protection-Preserving Collaborative Filtering Algorithm for Recommendation Systems

Abstract

With the rapid development of ubiquitous data collection and data analysis, data privacy in a recommended system is facing more and more challenges. Differential privacy technology can provide strict privacy protection while reducing the risk of privacy leakage, but it also introduces unwanted noise, which makes the performance of the recommender system worsen. Among different users, the degree of their sensitivity to privacy is usually different. Thus, through considering the impact of users’ personalized requirements, the collaborative filtering algorithm can be designed to reduce the amount of unwanted noise. Taking the above assertions into account, we propose a collaborative filtering algorithm based on personalized privacy protection. First, it locally classifies ratings by privacy sensitivity on the user side, then utilizes the random flip mechanism to protect the privacy-sensitive ratings. Then, after the server catches the perturbed rating data, we reconstruct the joint item-item distribution through the Bayesian estimation method. Experimental results show that our proposed algorithm can significantly improve the recommendation performance of recommendation systems while protecting users’ privacy.