Efficient Privacy-preserving Non-exhaustive Nearest Neighbor Search of large-scale databases

Abstract

The Nearest Neighbor Search is a process allowing a query owner to learn the nearest neighbor of his query in a database. Nearest Neighbor Search serves an essential role in a variety of applications of similarity search. A useful technique of Nearest Neighbor Search is Non-exhaustive Nearest Neighbor Search (NENS), allowing one to learn the nearest neighbor of the query in a sublinear time. However, existing research jeopardizes the privacy of the participants when performing NENS: either the query owner must communicate the query to the database owner, or the database owner must provide his customized database to the query owner. Such data disclosures are prohibited by privacy laws when it comes to scenarios with sensitive information. In this paper, we propose two protocols for Privacy-preserving Non-exhaustive Nearest Neighbor Search. Our first protocol accomplishes its purpose by carefully combining cryptography and Inverted File System, which is a basic method to achieve NENS. Building on the first protocol, our second protocol incorporates Product Quantization, which is another vital method for NENS implementation. In addition to privacy issues, we have improved the efficiency of Privacy-preserving Non-exhaustive Nearest Neighbor Search from previous work. Extensive experimental results on datasets with scales of 1 million and 10 million demonstrated that the proposed protocols perform better than the SANNS (Chen et al., 2020). Meanwhile, our protocols can scale to the dataset of 100 million entries, pushing the limit by one order of magnitude.