Towards practical personalized recommendation with multi-level differential privacy controls

Recommender systems have been widely applied in many scenarios such as shopping websites, online learning systems and video sharing clients, etc, which need to collect user's historical data for improving the accuracy of recommended results. However, user's private data inadvertently left behind on the internet or clients also will be gathered by service providers, and the risk of user's privacy being leaked may greatly increase. While the problem of private recommended algorithm has been studied in recent years, the latest research findings are still inadequate in security or utility. In this paper, we propose a novel multi-level differential privacy (Multi-DP)scheme under untrusted personalized recommendation system. Specifically, we firstly design a multi-level differential privacy algorithm employing Laplace mechanism to protect both overall privacy towards service provider and each data attributes privacy. Then, randomized response is introduced to sanitize user's historical data for generating useful perturbed data. In this way, the error between raw historical data and perturbed data will be further reduced. Finally, theoretical analysis and extensive experiments demonstrate the high security and data utility of our proposed model compared with existing methods.