GPU Accelerated Bayesian Inference for Quasi-Identifier Discovery in High-Dimensional Data

Abstract

Determining unique attribute combinations as quasi-identifiers is a common starting point for both re-identification attacks and data anonymisation schemes. The efficient discovery of those quasi-identifiers (QIDs) has been a combinatoric nightmare, actually an enumeration problem [1,2,3] given its W2-complete nature [4,5,6]. Proper privacy guarantees are required to fulfil highest ethical standards and privacy legislation like CCPA or GDPR, yet also enable the most modern data-driven business model based on monetising corporate data pools. In this work, we offer three main contributions: First, we contribute an algorithm that vectorises the QID search. This QID discovery is based on Bayesian inference detection, which usually suffers a state-space explosion for large-scale datasets. By utilising GPU acceleration to execute the vectorised algorithm, we counter the state-space-explosion issue raised by Bayesian networks. Second, we show its applicability to anonymising high-dimensional data which suffers high information-loss when using standard anonymisation approaches. Third, we offer an empirical model that compares multiple optimisations to discover all QIDs in near real-time, even in large-scale datasets. The latter becomes extremely useful for instances in digital health settings where algorithmic execution time can influence life-and-death triage. Finally, we point out that the same approach can foster de-anonymisation attacks on already published datasets. A demonstration is enclosed to re-identify individuals from Mount Vernon, NY and Southern California in a published Twitter dataset on US Presidential Election 2020.