A PRIVACY-PRESERVING IMAGE CLASSIFICATION FRAMEWORK WITH A LEARNABLE OBFUSCATOR

Xiangyi Meng, Zixuan Huang, Yuefeng Du, Antoni Chan, Cong Wang

Abstract: Real world images often contain large amounts of private / sensitive information that should be carefully protected without reducing their utilities. In this paper, we propose a privacy-preserving deep learning framework with a learnable ob- fuscator for the image classification task. Our framework consists of three mod- els: learnable obfuscator, classifier and reconstructor. The learnable obfuscator is used to remove the sensitive information in the images and extract the feature maps from them. The reconstructor plays the role as an attacker, which tries to recover the image from the feature maps extracted by the obfuscator. In order to best protect users’ privacy in images, we design an adversarial training methodol- ogy for our framework to optimize the obfuscator. Through extensive evaluations on real world datasets, both the numerical metrics and the visualization results demonstrate that our framework is qualified to protect users’ privacy and achieve a relatively high accuracy on the image classification task.

Keywords: privacy-preserving, image classification, adversarial training, learnable obfuscator

TL;DR: We proposed a novel deep learning image classification framework that can both accurately classify images and protect users' privacy.