Datamining and Disclosure Limitation for Categorical Statistical Databases

There are many distinctions between statistical research databases and those arising in commercial or administrative settings, and thus different issues regarding confidentiality and privacy protection on the one hand and and data access and the use of databases on the other. Data integration across multiple databases raises issues in both domains, especially with regard to protection against intruders. This paper highlights some methods developed to limit possible disclosure of confidential information from statistical databases while at the same time publicly releasing sufficient information to allow users, whether dataminers or other more traditional statistical analysts, sufficient data to reach proper statistical conclusions from their analyses. The disclosure limitation tools discussed include: data perturbation and simulation, partial releases, and sampling, with a special focus on partial release of data from multidimensional cross-classifications or contingency tables.