Privacy-Preserving Linear Fisher Discriminant Analysis

Abstract

Privacy-preserving data mining enables two or more parties to collaboratively perform data mining while preserving the data privacy of the participating parties. So far, various data mining and machine learning algorithms have been enhanced to incorporate privacy preservation. In this paper, we propose privacy-preserving solutions for Fisher Discriminant Analysis (FDA) over horizontally and vertically partitioned data. FDA is one of the widely used discriminant algorithms that seeks to separate different classes as much as possible for discriminant analysis or dimension reduction. It has been applied to face recognition, speech recognition, and handwriting recognition. The secure solutions are designed based on two basic secure building blocks that we have proposed—the Secure Matrix Multiplication protocol and the Secure Inverse of Matrix Sum protocol—which are in turn based on cryptographic techniques. We conducted experiments to evaluate the scalability of the proposed secure building blocks and overheads to achieve privacy when performing FDA.