A Framework for Transformation Network Training in Coordination with Semi-trusted Cloud Provider for Privacy-Preserving Deep Neural Networks

摘要

We propose a framework for transformation network training in coordination with a semi-trusted cloud provider for privacy-preserving DNNs. In the framework, a user trains a transformation network using a model that a cloud provider has for transforming plain images into visually protected ones. Conventional perceptual encryption methods have a weak visualprotection performance and some accuracy degradation in image classification. In contrast, the proposed framework overcomes the two issues. In an image classification experiment, the transformation network trained under the framework is demonstrated to strongly protect the visual information of plain images, without any performance degradation under the use of two typical classification networks: ResNet and VGG. In addition, it is shown that the visually protected images are robust against a DNN-based attack.