Measuring privacy/utility tradeoffs of format-preserving strategies for data release

In this paper, we introduce a novel approach to evaluate the risk of re-identification of individuals associated with format-preserving data release strategies, focusing on three strategies: data minimization (i.e. through data removal using random sampling and data Shapley values), data anonymization (i.e. through $k$k-anonymity), and data synthesis (i.e. through CTGAN and TVAE generative models). More precisely, our approach consists in simulating a security game in which (1) an attacker performs singling-out attacks as outlined in data protection regulations and (2) an evaluator scores attacks based on the linkability of records and the information gain obtained by the attacker. In addition, we further enhance our approach by simulating attacks as a cooperative game, in which the value of the attackers’ information resources is determined using the Shapley value borrowed from game theory. Re-identification Shapley value is proposed as a method to measure the level of re-identification potential of each feature in a dataset when combined with other features. We demonstrate the effectiveness of our approach using three datasets commonly used in the privacy literature. Overall, our work contributes to a better understanding of the inherent trade-offs that exist between data privacy and data utility in organizations.