Combined data distortion strategies for privacy-preserving data mining

Abstract

The problem of privacy-preserving data mining has become more and more important in recent years. Many successful and efficient techniques have been developed. However, in collaborative data analysis, part of the datasets may come from different data owners and may be processed using different data distortion methods. Thus, combinations of datasets processed using different methods are of practical interests. In this paper, a class of novel data distortion strategies is proposed. Four schemes via attribute partition, with different combinations of singular value decomposition (SVD), nonnegative matrix factorization (NMF), discrete wavelet transformation (DWT), are designed to perturb submatrix of the original datasets for privacy protection. We use some metrics to measure the performance of the proposed new strategies. Data utility is examined by using a binary classification based on the support vector machine. Our experimental results indicate that, in comparison with the individual data distortion techniques, the proposed schemes are very efficient in achieving a good trade-off between data privacy and data utility, and provide a feasible solution for collaborative data analysis.