Encrypted Classification Using Secure K-Nearest Neighbour Computation

Abstract

Machine learning (ML) is one of the growing areas of engineering with sweeping applications. Executing machine learning algorithms on vast amount of data raises demand of huge resources and large data set handling. Thus, machine learning was too costly for many enterprise budgets. However, cloud service suppliers are making this technology reasonable to enterprises by offering massive shared resources. Machine learning as a service (MlaaS) is a category of cloud computing services that provides machine learning tools to allow customers to run, develop and manage applications in cloud without the complexity of building and maintaining. However, ascent of machine learning as a service procreates scenarios where one faces concealment dilemma, where the model must be revealed to the outsourced platform. Hence, cloud data security is an important issue where users can fancy the ability of executing applications by outsourcing sensitive data. Fully Homomorphic Encryption (FHE) offers a refined way to accommodate these conflicting interests in the cloud scenario by preserving data confidentiality as well as applying Mlaas in secure domain. However, processing on FHE data can not be directly performed on traditional instruction execution flow, but requires special circuit based representation of algorithms. In this paper, we focus on realizing K-Nearest Neighbour (KNN) computation on encrypted data, where data is stored using a generalized encrypted representation. Such representation will be suitable for easily extending to encrypted ensemble learning framework supporting multiple encrypted learners for higher accuracy. Extensive performance studies are carried out to evaluate the timing overhead of the encrypted KNN computation.