Cloud-based differentially private image classification

Abstract

In this paper, our aim is to design and develop an anonymous full-duplex image classification framework under Differential Privacy. We work under the assumption that both, the cloud and the querier are semi-trusted entities, thus their data should remain safe and confidential. That is, neither the querier nor the cloud should be able to link a particular individual from the other party to an image while maintaining, to a certain extent, suitable classification accuracy. We use Principal Component Analysis (PCA) to transform sample images into anonymized vectors; differentially private synopsis of PCA vectors, and we ensure that the individuals in these vectors remain unidentifiable. © 2018, Springer Science+Business Media, LLC, part of Springer Nature.