An efficient multi-party scheme for privacy preserving collaborative filtering for healthcare recommender system

作者：Kaur Harmanjeet; Kumar Neeraj^*; Batra Shalini

来源：Future Generation Computer Systems-The International Journal of eScience, 2018, 86: 297-307.

DOI：10.1016/j.future.2018.03.017

• Recommender system
• Collaborative filtering
• Privacy preserving
• Healthcare
• Homomorphic encryption

Patient oriented decision-making in medical domains can enhance the efficiency of the modern healthcare recommender system provided the data scattered across different geographical regions is collected, mined and analyzed efficiently. Different sites, having Arbitrary Distributed Data (ADD) of healthcare services at various nodes can collaborate with each other to generate customer's preference leading to mutual advantage and overcoming of the issues related to insufficient ratings of various medical services. However, due to privacy, financial and legal issues, different parties defer from sharing their confidential data. If the parties are assured of data confidentiality, they might agree for fruitful collaboration. Few existing studies proposed Privacy Preserving Collaborative Filtering (PPCF) on ADD, but these techniques considered only two parties. Moreover, the computation cost of off-line model generation process is high since these techniques use homomorphic encryption techniques. To fill these gaps, this paper propose PPCF scheme on ADD based on multi-party random masking and polynomial aggregation techniques. In the proposal, two phases are considered namely as: off-line model generation and online prediction generation. Three protocols have been considered for privacy preservation so that analysis of each protocol is performed separately. The Paillier homomorphic encryption system is also used to calculate the length of vector X securely, so that only additive property of homomorphic encryption is used. Analysis of the proposed scheme has been done for security, accuracy, coverage and performance on healthcare and Movieslens datasets. It has been experimentally demonstrated that the proposed scheme maintains data owner's confidentiality, and privacy measure so that it does not affect the accuracy of prediction generation on integrated data. Comparative analysis of the proposed scheme has also been done with other related schemes based on off-line and online computation overheads. The results obtained demonstrated that the proposed scheme has significant improvement by a factor of 36% (approx) with respect to the aforementioned parameters.

• 出版日期2018-9