Privacy-Preserving Data Publishing Based on Utility Specification

Abstract

Most existing privacy-preserving data publishing methods anonymize data based on some general utility measures. However, the anonymized data may not be useful to applications that have specific requirements for the data they use. In this paper, we propose a method for data users to describe some characteristics of the anonymized data, as a special requirement of some classification applications, and a heuristic anonymization algorithm that incorporates the user-specified requirements into a generalization technique. Our preliminary results show that the specification format and the anonymization algorithm can significantly improve the utility of the anonymized data for a number of data mining applications that learn decision trees, Naive Bayes Classifier and Classification Rules.