A New Steganography Method Based on Generative Adversarial Networks

In this paper, we propose a new steganography method based on generative adversarial networks (GAN). In using steganography, the third party can extract the embedded secret easily by comparing the cover data and the hidden message if the cover data are publicly available (e.g. accessible in the internet). To avoid this problem, we need a unique cover datum for each piece of secret messages. As the cover data, we focus on digital images in this study. To make a lot of unique and natural looking images, we can use GAN. Actually, training of GAN results in two neural networks namely the generator and the discriminator. The generator makes virtual images, and the discriminator evaluates the naturalness of the virtual images. In the proposed method, we use both the generator and the discriminator to guarantee the naturalness of the cover data on the sender side, and to filter out stego data sent from malicious third party. In this experiments, we first confirmed the capability of the generator for producing unlimited number of cover data, and then investigated the possibility of naturalness checking using the discriminator. We believe that the proposed method can provide a better way for information hiding.