Privacy-preserving collaboration in an integrated social environment

Abstract

Privacy and security of data have been a critical concern at the state, organization and individual levels since times immemorial. New and innovative methods for data storage, retrieval and analysis have given rise to greater challenges on these fronts. Online social networks (OSNs) are at the forefront of individual privacy concerns due to their ubiquity, popularity and possession of a large collection of users' personal data. These OSNs use recommender systems along with their integration partners (IPs) for offering an enriching user experience and growth. However, the recommender systems provided by these OSNs inadvertently leak private user information. In this work, we develop solutions targeted at addressing existing, real-world privacy issues for recommender systems that are deployed across multiple OSNs. Specifically, we identify the various ways through which privacy leaks can occur in a friend recommendation system (FRS), and propose a comprehensive solution that integrates both Differential Privacy and Secure Multi-Party Computation (MPC) to provide a holistic privacy guarantee. We model a privacy-preserving similarity computation framework and library named Lucene-P2. It includes the efficient privacy-preserving Latent Semantic Indexing (LSI) extension. OSNs can use the Lucene-P2 framework to evaluate similarity scores for their private inputs without sharing them. Security proofs are provided under semi-honest and malicious adversary models. We analyze the computation and communication complexities of the protocols proposed and empirically test them on real-world datasets. These solutions provide functional efficiency and data utility for practical applications to an extent.