Learning quasi-identifiers for privacy-preserving exchanges: a rough set theory approach

摘要

The challenging and pervasive issue associated with information exchange is inferential disclosure. It occurs in the following three situations: (1) the exchanged data correlate with publicly available information, (2) the exchanged data comprise patterns similar to those in a sharing partner's datum, and (3) the shared data's attributes are interdependent. In this work, we provide and implement new algorithms that impede the third type of inferential attack. They rely on rough set theory to undermine the deductive route from nonsensitive to sensitive features. Our approach comprises three steps which include learning quasi-identifiers, computing a granulation of the underlying information system that maximizes the distribution of sensitive attributes in each granule, and masking the deductive route from nonsensitive to sensitive features. Our routine for learning quasi-identifiers achieves both the largest distinction and separation without an exhaustive search among tuples of features. The learned quasi-identifiers are employed to find a granulation of the information system that strikes a balance between the anonymity of quasi-identifiers and the diversity of sensitive attributes, without solving a difficult optimization problem. We employ this granulation in a strategy similar to that used in k-anonymity to de-identify private information systems.