Privacy-preserving Cross-datasets Collaborative Filtering

Recommender systems have become extremely common in recent years,and are widely utilized in various domains.Recommender systems offer users personalized recommendations for many kinds of items that meet their preferences.One of the main techniques on which recommender systems are based is collaborative filtering(CF).The accuracy of the CF depends on sufficient users' preference data.A typical CF-based recommender system is designed in a central scenario,basing its recommendations solely on users' preference data collected by the system itself.Although the central CF recommender system has high operating efficiency,it cannot obtain a large amount of diverse data to ensure the accuracy of recommendations,because of the limitation of service.In fact,if several recommender systems(or vendors)share their data to enrich the users' preference information,the quality of recommendations can significantly be improved.However,competitive advantage,privacy concerns and regulations,and issues surrounding data sovereignty and jurisdiction prevent many vendors from openly sharing their data.Therefore,a privacy-preserving multiparty protocol for CF recommender systems is needed to break the isolation of data and to realize the controllable data sharing.In order to achieve this goal,it is necessary to introduce a secure multi-party computing(MPC)technology that uses cryptography to protect data privacy,specifically through the protocols such as Yao garbled circuit and secret sharing.It allows multiple data owners without mutual trust to collaboratively perform computation,but no one can get any information other than the result of the calculation.Therefore,this thesis focuses on the distributed scenario in which multiple data sources make recommendations based on their shared data.We propose a privacy-preserving multiparty scheme for collaborative filtering,using mixed MPC protocols combined with Yao garbled circuit and additive secret sharing.The scheme can solve privacy leakage problem during federated training phase and federated recommendation phase.The main contributions are as follows:(1)We propose a federated training scheme based on the mixed multiparty computation.The scheme introduces a two-server model to perform efficient secure two party computation protocols.A sorting network based on Yao garbled circuit is designed to solve the matrix sparsity,and gradient computation is performed with additive secret sharing to obtain federated recommendation model.(2)We propose a privacy-preserving federated recommendation scheme.Yao garbled circuit and additive secret sharing are used to complete rating prediction and sorting.Then servers interact with the data sources to generate recommender lists.In addition,the thesis also utilizes the particularity of horizontal distribution and vertical distribution of data to optimize the federated recommendation scheme.(3)The proposed scheme is implemented in C++.The implementation is optimized by the comparison reusing and parallelization.The accuracy and efficiency of the proposed scheme is verified under the real dataset.The experimental results show that the scheme of this thesis can ensure the accuracy of the recommendation results while ensuring the data privacy of multiple data sources.