Privacy-Preserving LDA Classification over Horizontally Distributed Data

Abstract

This paper presents a framework for constructing Linear Discriminate Analysis (LDA) classifier, securely, over distributed data. It is assumed that data is partitioned among several parties such that for obtaining higher benefits from a greater extent of data, all participants are willing to model the LDA classifier on whole data, but for privacy concerns, they refuse to share the original datasets. To this end, we propose an algorithm based on secure computation protocols, which provides the data owners the possibility of LDA classifier construction over all datasets without revealing any sensitive information. In our experimental analysis, applying data-packing and tree communication model, has shown the reduction of computation and communication costs 35 and 15 times, respectively.