Algorithm-safe privacy-preserving data publishing

Abstract

This paper develops toolsets for eliminating algorithm-based disclosure from existing privacy-preserving data publishing algorithms. We first show that the space of algorithm-based disclosure is larger than previously believed and thus more prevalent and dangerous. Then, we formally define Algorithm-Safe Publishing (ASP) to model the threats from algorithm-based disclosure. To eliminate algorithm-based disclosure from existing data publishing algorithms, we propose two generic tools for revising their design: worst-case eligibility test and stratified pick-up. We demonstrate the effectiveness of our tools by using them to transform two popular existing l-diversity algorithms, Mondrian and Hilb, to SP-Mondrian and SP-Hilb which are algorithm-safe. We conduct extensive experiments to demonstrate the effectiveness of SP-Mondrian and SP-Hilb in terms of data utility and efficiency.