Differentially private Naive Bayes learning over multiple data sources.

For meeting diverse requirements of data analysis, the machine learning classifier has been provided as a tool to evaluate data in many applications. Due to privacy concerns of preventing disclosing sensitive information, data owners often suppress their data for an untrusted trainer to train a classifier. Some existing work proposed privacy-preserving solutions for learning algorithms, which allow a trainer to build a classifier over the data from a single owner. However, they cannot be directly used in the multi-owner setting where each owner is not totally trusted for each other. In this paper, we propose a novel privacy-preserving Naive Bayes learning scheme with multiple data sources. The proposed scheme enables a trainer to train a Naive Bayes classifier over the dataset provided jointly by different data owners, without the help of a trusted curator. The training result can achieve ϵ-differential privacy while the training will not break the privacy of each owner. We implement the prototype of the scheme and conduct corresponding experiment.