Towards value disclosure analysis in modeling general databases

The issue of confidentiality and privacy in general databases has become increasingly prominent in recent years. A key element in preserving privacy and confidentiality of sensitive data is the ability to evaluate the extent of all potential disclosure for such data. This is one major challenge for all existing perturbation or transformation based approaches as they conduct disclosure analysis on the perturbed or transformed data, which is too large, considering many organizational databases typically contain a huge amount of data with a large number of categorical and numerical attributes. Instead of conducting disclosure analysis on perturbed or transformed data, our approach is to build an approximate statistical model first and analyze various potential disclosure in terms of parameters of the model built. As the model learned is the only means to generate data for release, all confidential information which snoopers can derive is contained in those parameters.