Privacy and Utility Effects of k-anonymity on Association Rule Hiding

Abstract

In recent years, privacy preservation has attracted much interest due to concerns regarding breaches of privacy when data are published and analyzed. Private information can be observed directly from published data or inferred through data mining techniques. The k-anonymity concept was first proposed to hide sensitive attribute values that could be discovered using a linking attack. Association rule hiding techniques have been proposed to hide sensitive patterns in mining results. However, these association rule hiding techniques have side effects such as hiding failure, creation of new rules, and lost rules. In addition, the k-anonymity approach does not consider hiding association rules. In this work, we extend the k-anonymity concept to hide sensitive association rules and compare it with the association rule hiding approach. We propose a novel concept of measuring privacy gain and utility loss of anonymized association rules. Numerical experiments comparing the two approaches show that the k-anonymity for association rule mining approach achieves higher privacy gain, while the direct anonymization approach of association rule hiding achieves lower utility loss. The results obtained here provide a guideline for adopting anonymization techniques under different requirements and suggests a direction for the development of new association rule hiding techniques.