Attribute Segregation based on Feature Ranking Framework for Privacy Preserving Data Mining

Abstract

Attributes in macro-data have to be segregating based on their sensitivity for privacy preservation purposes. Automating this attribute segregation becomes complicated in high dimensional datasets and data streams. In this work, information or correlation of the attribute on the target class attribute is measured using Information Gain [IG], Gain Ratio [GR] and Pearson Correlation [PC] ranker based feature selection methods and this values are used to segregate them as Sensitive Attributes [SA], Quasi Identifiers [QI] and Non-Sensitive [NS] Attributes. Segregated attributes are subjected to various levels of privacy preservation using both the proposed Double layer Perturbation [DLP] and Single Layer Perturbation [SLP] algorithms to form the level-1 perturbed datasets. The level-1 perturbed dataset is further perturbed by applying SLP algorithm to form level-2 and level-3 privacy preserved datasets. Thus, the multiple versions of Adult dataset created are distributed to data seekers based on their trust levels in Multi Trust Level [MTL] environment. The privacy preserved dataset versions created using the proposed algorithms are evaluated based on their utility, distortion and purity metrics. The results show that the ranker methods are able to identify attributes which had sensitive content as either SA or QI automatically and the proposed perturbed datasets have good utility on selected classification and clustering algorithms when compared to original and L-diversified datasets. Also, the distortion values of these datasets signify that they can prevent diversity attacks.