Quantifying Location Privacy Risks Under Heterogeneous Correlations

Abstract

Currently, increasingly ubiquitous location-based services are facilitating the activities of people in daily life. However, releasing real locations could lead to serious concerns about privacy. To remedy these issues, a number of location privacy protection mechanisms (LPPMs) have been proposed, e.g., spatial cloaking, dummy location generation, query caching, and perturbation. However, these LPPMs are vulnerable to inference attacks because of the incompleteness of the captured privacy risks caused by heterogeneous correlations in location data, e.g., semantical, temporal, and social correlations. Consequently, they cannot provide sufficient privacy guarantees due to the absence of embedded heterogeneous correlations in the design process of LPPM. To address these issues, we present QUAD, a framework for quantifying location privacy risks under heterogeneous correlations. QUAD has three features: 1) it enables the modeling and seamless fusion of multiple kinds of correlations that are available to adversaries; 2) it provides a probabilistic representation of the privacy risks faced under heterogeneous correlations; and 3) it achieves the quantification of privacy risks for multiple kinds of LPPMs that are widely used in the literature. To mitigate privacy threats, we propose a defense mechanism embedded with the quantified privacy risks. Extensive experiments on two real-world datasets confirm that QUAD can capture more privacy risks than competitors, and the risks can be dramatically reduced by our defense mechanism.