We propose an image transformation network to generate visually-protected images for privacy-preserving deep neural networks (DNNs). The transformation network inspired by adversarial examples is considered not only to generate visually-protected images but also to achieve the performance of DNNs that using plain images has. Moreover, the proposed transformation network can be public to users, so there is no need to manage secret keys. In an image classification experiment, the proposed transformation network is confirmed to protect the visual information without any performance degradation under the use of two image classification networks: ResNet and VGG.