Skip to Main Content
Data mining is a powerful tool for information discovery from huge datasets. Various sectors, including commercial, government, financial, medical, and scientific, are applying data mining techniques on their datasets that typically contain sensitive individual information. During this process the datasets get exposed to several parties, which can potentially lead to disclosure of sensitive information and thus to breaches of privacy. Several data mining privacy preserving techniques have been recently proposed. In this paper we focus on data perturbation techniques, i.e., those that add noise to the data in order to prevent exact disclosure of confidential values. Some of these techniques were designed for datasets having only numerical non-class attributes and a categorical class attribute. However, natural datasets are more likely to have both numerical and categorical non-class attributes, and occasionally they contain only categorical attributes. Noise addition techniques developed for numerical attributes are not suitable for such datasets, due to the absence of natural ordering among categorical values. In this paper we propose a new method for adding noise to categorical values, which makes use of the clusters that exist among these values. We first discuss several existing categorical clustering methods and point out the limitations they exhibit in our context. Then we present a novel approach towards clustering of categorical values and use it to perturb data while maintaining the patterns in the dataset. Our clustering approach partitions the values of a given categorical attribute rather than the records of the datasets; additionally, our approach operates on the horizontally partitioned dataset and it is possible for two values to belong to the same cluster in one horizontal partition of the dataset, and to two distinct clusters in another partition. Finally, we provide some experimental results in order to evaluate our perturbation technique and to compare our clustering approach with an existing method, the so-called CACTUS.