Personal data privacy is considered to be a fundamental right. It forms a part of our highest ethical standards and is anchored in legislation and various best practices from the technical perspective. Yet, protecting against personal data exposure is a challenging problem from the perspective of generating privacy-preserving datasets to support machine learning and data mining operations. The issue is further compounded by the fact that devices such as consumer wearables and sensors track user behaviours on such a fine-grained level, thereby accelerating the formation of multi-attribute and large-scale high-dimensional datasets.
In recent years, increasing news coverage regarding de-anonymisation incidents, including but not limited to the telecommunication, transportation, financial transaction, and healthcare sectors, have resulted in the exposure of sensitive private information. These incidents indicate that releasing privacy-preserving datasets requires serious consideration from the pre-processing perspective. A critical problem that appears in this regard is the time complexity issue in applying syntactic anonymisation methods, such as k-anonymity, l-diversity, or t-closeness to generating privacy-preserving data. Previous studies have shown that this problem is NP-hard.
This thesis focuses on large high-dimensional datasets as an example of a special case of data that is characteristically challenging to anonymise using syntactic methods. In essence, large high-dimensional data contains a proportionately large number of attributes in proportion to the population of attribute values. Applying standard syntactic data anonymisation approaches to generating privacy-preserving data based on such methods results in high-information loss, thereby rendering the data useless for analytics operations or in low privacy due to inferences based on the data when information loss is minimised.
We postulate that this problem can be resolved effectively by searching for and eliminating all the quasi-identifiers present in a high-dimensional dataset. Essentially, we quantify the privacy-preserving data sharing problem as the Find-QID problem by formalising a sequence of reductions to the Hitting-Set problem. This formally derives the Find-QID problem as W-complete containing a fixed-parameter tractable problem (FPT). FPT implies a super-polynomial time complexity unless the W-hierarchy partially collapses. Further, we show that despite their complex nature, the discovery of QID can be achieved reliably for large datasets. The risk of private data exposure through inferences can be circumvented, and both can be practicably achieved without the need for high-performance computers.
For this purpose, we present, implement, and empirically assess both mathematical and engineering optimisation methods for deterministic discovery of privacy-violating inferences. This includes a greedy search scheme by efficiently queuing QID candidates based on their tuple characteristics, projecting QIDs on Bayesian inferences, and countering Bayesian network’s state-space-explosion with multigrid solver and vectorised GPU acceleration. Part of this work showcases magnitudes of processing acceleration, particularly in high dimensions. We even achieve near real-time runtime for currently impractical applications. At the same time, we demonstrate how such contributions could be abused to de-anonymise Kristine A. and Cameron R. in a public Twitter dataset addressing the US Presidential Election 2020.
Finally, this work contributes, implements, and evaluates a novel syntactic anonymisation methodology, attribute compartmentation. Attribute compartmentation promises sanitised datasets without remaining quasi-identifiers while minimising information loss. To prove its functionality in the real world, we partner with digital health experts to conduct a medical use case study. As part of the experiments, we illustrate that attribute compartmentation is suitable for everyday use and, as a positive side effect, even circumvents a common domain issue of base rate neglect. For this purpose, we present, implement and empirically assess both mathematical and engineering optimisation methods for deterministic discovery of privacy-violating inferences. This includes a greedy search scheme by eﬃciently queuing QID candidates based on their tuple characteristics, projecting QIDs on Bayesian inferences and countering Bayesian network’s state-space-explosion with multigrid solver and vectorised GPU acceleration.