Constraints Inference Channels and Secure Databases

Abstract

This paper investigates the problem of confidentiality violations via illegal data inferences that occur when arithmetic constraints are combined with non-confidential numeric data to infer confidential information. The database is represented as a point in an (n + k)-dimensional constraint space, where n is the number of numerical data items stored in the database (extensional database) and k is the number of derivable attributes (intensional database). Database constraints over both extensional and intensional databases form an (n + k)-dimensional constraint object. A query answer over a data item x is an interval I of values along the x axis of the database such that I is correct (i.e., the actual data value is within I) and safe (i.e., users cannot infer which point within I is the actual data value). The security requirements are expressed by the accuracy with which users are allowed to disclose data items. More specifically, we develop two classification methods: (1) volume-based classification, where the entire volume of the disclosed constraint object that contains the data item is considered and (2) interval based classification, where the length of the interval that contains the data item is considered. We develop correct and safe inference algorithms for both cases.