Homomorphically encrypted k-means on cloud-hosted servers with low client-side load

Abstract

The significance of data analytics has been acknowledged in many scientific and business domains. However, the required processing power and memory capacity is a prohibiting factor for performing data analytics on proprietary platforms. An obvious solution is the outsourcing of data analytics to cloud storage and cloud computing providers but this entails that privacy and security issues are raised, given the fact that data can be valuable and/or personal. The aim of this paper is the development of a server-side k-means algorithm over encrypted data using homomorphic encryption in order to overcome both the lack of resources of the data owner and the security concerns. Current solutions that deal with homomorphic encryption impose a heavy load on the side of the data owner; this limitation is now addressed in this work. More specifically, in this paper, we present a framework for the implementation of an homomorphic version of k-means, we discuss the capabilities of the current state-of-the-art homomorphic encryption schemes, and we propose a novel approach to server-side computation of k-means assuming a new adversary model tailored to modern settings. We instantiate our framework in two different versions in terms of operation assignment each coming in three flavors of operation implementation. All alternatives are evaluated thoroughly using both real experiments and analytic cost models.