A security machanism for statistical database

Abstract

The problem of user inference in statistical databases is discussed and illustrated with several examples. It is assumed that the database allows “total,” “average,” “count,” and “percentile” queries; a query may refer to any arbitrary subset of the database. Methods for protecting the security of such a database are considered; it is shown that any scheme which gives “statistically correct” answers is vulnerable to penetration. A precise definition of compromisability (in a statistical sense) is given. A general model of user inference is proposed; two special cases of this model appear to contain all previously published strategies for compromising a statistical database. A method for protecting the security of such a statistical database against these types of user inference is presented and discussed. It is shown that the number of queries required to compromise the database can be made arbitrarily large by accepting moderate increases in the variance of responses to queries. A numerical example is presented to illustrate the application of the techniques discussed.