Two-Step Image-in-Image Steganography via GAN

Recently, convolutional neural network has been introduced to information hiding and deep net- work has shown great potential in steganography. However, one drawback of deep network is that it’s sensitive to small fluctuations. In previous works, the encoder-decoder structure is trained end-to-end, but in practice, encoder and decoder should be used separately. Therefore, end-to-end trained steganography networks are vulnerable to fluctuations and the secret decoded from those networks suffers from unpleasant noise. In this work, we present an image-in-image steganog- raphy method called TISGAN to achieve better results, both in image quality and security. In particular, we divide the training process into two parts. Moreover, perceptual loss is applied to encoder, to improve security in our work. We also append a denoising structure to the end of de- coder to achieve better image quality. Finally, the adversarial structure with useful techniques employed is also used in secret revealed process.