Distributed privacy-preserving minimal distance classification

Abstract

The paper focuses on the problem of preserving privacy for a minimal distance classifier working in the distributed environment. On the basis of the study of available works devoted to privacy aspects of machine learning methods, we propose the novel definition and taxonomy of privacy. This taxonomy was used to develop new effective classification algorithms which can work in distributed computational environment and assure a chosen privacy level. Instead of using additional algorithms for secure computing, the privacy assurance is embedded in the classification process itself. This lead to a significant reduction of the overall computational complexity what was confirmed by the computer experiments which were carried out on diverse benchmark datasets. © 2013 Springer-Verlag.