Privacy and the Dimensionality Curse

Most privacy-transformation methods such as k-anonymity or randomization use some kind of transformation on the data for privacy-preservation purposes. In many cases, the data can be indirectly identified with the use of a combination of attributes. Such attributes may be available from public records and they may be used to link the sensitive records to the target of interest. Thus, the sensitive attributes of the record may be inferred as well with the use of publicly available attributes. In many cases, the target of interest may be known to the adversary, which results in a large number of combinations of attributes being known to the adversary. This is a reasonable assumption, since privacy attacks will often be mounted by an adversary with some knowledge of the target. As a result, the number of attributes for identification increases, and results in almost unique identification of the target. In this paper, we will examine a number of privacypreservation methods and show that in each case the privacy-preservation approach becomes either ineffective or infeasible.