Perturbative Data Protection of Multivariate Nominal Datasets

Many of the potentially sensitive personal data produced and compiled in electronic sources are nominal and multi-attribute (e.g., personal interests, healthcare diagnoses, commercial transactions, etc.). For such data, which are discrete, finite and non-ordinal, privacy-protection methods should mask original values to prevent disclosure while preserving the underlying semantics of nominal attributes and the (potential) correlation between them. In this paper we tackle this challenge by proposing a semantically-grounded version of numerical correlated noise addition that, by relying on structured knowledge sources (ontologies), is capable of perturbing/masking multivariate nominal attributes while reasonably preserving their semantics and correlations.