Applying multi-label and multi-class classification to enhance K-anonymity in sequential releases

Privacy-preserving data mining is gaining prominence due to increased accumulation of data containing personal information. Data holders in healthcare, finance and other sectors collecting person-specific information are challenged to publish useful data, while meeting ever-increasing demands of privacy protection for data subjects. K-anonymity is a popular technique used to preserve data privacy for data publishing by anonymizing quasi identifiers (QI) (e.g., race, gender, age). However, K-anonymized data can be at risk of temporal attacks that target multiple versions of released data, also called sequential releases. The objective of this study is to develop a model that uses multi-class and multi-label classifiers to evaluate risk in re-identifying QI information in previous data releases through learning from current data release. In our empirical study, we use five healthcare and financial data sets to compare performance of binary relationship and label powerset problem transformations and Naïve Bayes, C4.5, random tree and kNN learning algorithms. Our empirical results show that multi-label classification is a powerful tool in enhancing K-anonymity of sequential data release. Statistical analysis of the classification results shows that RAkEL outperforms other transformation methods in predicting demographics information, hence, can be useful in assessing risks of QI re-identification.