Cor-Split: Defending Privacy in Data Re-publication from Historical Correlations and Compromised Tuples

Abstract

Several approaches have been proposed for privacy preserving data publication. In this paper we consider the important case in which a certain view over a dynamic dataset has to be released a number of times during its history. The insufficiency of techniques used for one-shot publication in the case of subsequent releases has been previously recognized, and some new approaches have been proposed. Our research shows that relevant privacy threats, not recognized by previous proposals, can occur in practice. In particular, we show the cascading effects that a single (or a few) compromised tuples can have in data re-publication when coupled with the ability of an adversary to recognize historical correlations among released tuples. A theoretical study of the threats leads us to a defense algorithm, implemented as a significant extension of the m-invariance technique. Extensive experiments using publicly available datasets show that the proposed technique preserves the utility of published data and effectively protects from the identified privacy threats.