Towards k-Anonymous Non-numerical Data via Semantic Resampling

Abstract

Privacy should be carefully considered during the publication of data (e.g. database records) collected from individuals to avoid disclosing identities or revealing confidential information. Anonymisation methods aim at achieving a certain degree of privacy by performing transformations over non-anonymous data while minimising, as much as possible, the distortion (i.e. information loss) derived from these transformations. k-anonymity is a property typically considered when masking data, stating that each record (corresponding to an individual) is indistinguishable from at least k-1 other records in the anonymised dataset. Many methods have been developed to anonymise data, but most of them are focused solely on numerical attributes. Non-numerical values (e.g. categorical attributes like job or country-of-birth or unbounded textual ones like user preferences) are more challenging because arithmetic operations cannot be applied. To properly manage and interpret this kind of data, it is required to have operators that are able to deal with data semantics. In this paper, we propose an anonymisation method based on a classic data re-sampling algorithm that guarantees the fulfilment of the k-anonymity property and is able to deal with non-numerical data from a semantic perspective. Our method has been applied to anonymise the well-known Adult Census dataset, showing that a semantic interpretation of non-numerical values better minimises the information loss of the masked data file.