One-off Disclosure Control by Heterogeneous Generalization

How can we orchestrate an one-off sharing of informative data about individuals, while bounding the risk of disclosing sensitive information to an adversary who has access to the global distribution of such information and to personal identifiers? Despite intensive efforts, current privacy protection techniques fall short of this objective. Differential privacy provides strong guarantees regarding the privacy risk incurred by one's participation in the data at the cost of high information loss and is vulnerable to learning-based attacks exploiting correlations among data. Syntactic anonymization bounds the risk on specific sensitive information incurred by data publication, yet typically resorts to a superfluous clustering of individuals into groups that forfeits data utility.

In this paper, we develop algorithms for disclosure control that abide to sensitive-information-oriented syntactic privacy guarantees and gain up to 77% in utility against current methods. We achieve this feat by recasting data heterogeneously, via bipartite matching, rather than homogeneously via clustering. We show that our methods resist adversaries who know the employed algorithm and its parameters. Our experimental study featuring synthetic and real data, as well as real learning and data analysis tasks, shows that these methods enhance data utility with a runtime overhead that is small and reducible by data partitioning, while the 𝛽-likeness guarantee with heterogeneous generalization staunchly resists machine-learning-based attacks, hence offers practical value.