Geavanceerde tactieken voor het anonimiseren van datasets

The volume of data that is collected each year grows rapidly and has led to the development of data-driven ecosystems, where information - often of sensitive nature - is leveraged to improve productivity and services. Although companies traditionally relied solely on internal data sources, there is a tendency to combine internally collected data with external data sources. This means that data processors enrich internal data with data that is kept at the premises of external data owners. However, uncontrolled release of sensitive data is often undesirable. Releasing fine-grained sensitive data can undermine the competitiveness of companies, potentially leading indirectly to loss of consumers or reputational damage. Moreover, privacy regulations stipulate constraints on the release and processing of sensitive personal data. Hence, controlled release of data is a necessity in many scenarios. Statistical dataset anonymization is a promising privacy-enhancing approach to support this and the core research field of this thesis. It consists of controlling data flows by transforming sensitive datasets - in which the identity of individuals can be linked to sensitive information - to anonymized ones, before the data is actually released to data processors. However, thoughtless dataset transformations can undermine the utility of the data that is actually exchanged. This PhD dissertation exactly focuses on the tension between the conflicting concerns of data owners and data processors. More specifically, this PhD dissertation introduces novel tactics to improve the privacy-utility balance during dataset anonymization, thereby mainly relying on machine learning mechanisms. We demonstrate their feasibility by applying the proposed tactics in settings where data processors want to rely on external sensitive data sources to train machine learning algorithms. We further propose tactics to reconcile the machine learning pipeline in the presence of sensitive yet valuable data from external data owners, and compare mechanisms for statistical dataset anonymization with synthetic data approaches. This dissertation further results in guidelines that can be applied by data owners and data processors. By increasing the privacy-utility balance of released sensitive datasets, this PhD dissertation facilitates collaborations between data owners and data processors ultimately increasing the competitiveness of both stakeholders. The former increases revenues while being compliant with privacy regulation while the latter can build more effective machine learning models due to external data sources. The use-cases that were tackled and experiments that were elaborated in the context of this thesis have shown that intelligent anonymization tactics can effectively improve the privacy-utility balance and that revising the machine learning pipeline can be a real added value when dealing with sensitive data. We have also shown that partial automation of the anonymization pipeline is possible, but manual steps remain necessary for some tasks. Finally, synthetic data gains high attraction by industry today. However, experiments across various datasets and application domains have shown that synthetic datasets not necessarily outperform traditional statistical anonymization techniques.