An Experimental Study of Matrix-Based Data Distortion Methods

Abstract

A number of matrix-based data distortion methods are presented and experimentally studied in this paper. The performances of seven methods are compared in terms of utility, privacy and computational cost. We find that left multiplication based random projection methods are useless in data privacy protection. Even though there is no application-free solution in data privacy protection, the nonnegative matrix factorization (NMF) based method has an appealing privacy performance under the promise of a reasonable utility and computational cost. While the random projection method with a right multiplication of an orthogonal random matrix does well in support vector machine classification, its computational disadvantages may make it less attractive for an online analysis and processing application.