Further Developments with Perturbation Techniques to Protect Tabular Data

Abstract

Statistical agencies collect input data from individuals and deliver output information to the society based on these data. A fundamental feature of output information is the “protection” of sensitive information, since too many details could disseminate privacy information from individuals and therefore violate their rights. Another feature of output information is the “utility” to data users, as a scientific may use this output for research or a politician for making decisions. Clearly more details are in the output, more useful it is, but it is also less protected. There are several methodologies based on Mathematical Optimization to solve the problem of finding “good” protected and useful solutions. While the literature on algorithms to apply them is extensive, statisticians have major concerns to use them in practice because these algorithms may have numeral troubles on frequency tables and may produce biased solutions. This article discusses these observations and describes how to overcome them using a modern technique called Enhanced Controlled Tabular Adjustment. Computational experiments show the effectiveness of the approach on benchmark instances.