Privacy-preserving Naive Bayes classification in semi-fully distributed data model

Abstract

In recent years, issues of privacy preservation in data mining and machine learning have received more and more attention from the research community. Privacy-preserving data mining and machine learning solutions enable data holders to jointly discover knowledge and valuable information, as well as construct machine learning models without privacy concerns. In this paper, we address the distressing problem of privacy-preservation for a novel data model called the semi-fully distributed setting. Differently from the existing scenarios, each record of the dataset in this data model is composed of three parts, in which the first part is privately kept by a data user, the second one is securely stored by the miner, and the rest is publicly known by both the miner and the data user. For this new data model, we propose a privacy-preserving Naive Bayes classification solution based on secure multi-party computation. Our proposed solution not only achieves a high level of privacy but also guarantees the accuracy of the classification model. The experimental results show that the new proposal is efficient in real-life applications. Furthermore, our pioneering study paves the way for new researches into privacy preservation issues for the semi-fully distributed data model.