Binary steganography based on generative adversarial nets

Some of the most advanced steganographic methods for binary images are to manually extract the features of binary images. And state-of-the-art binary image steganography techniques need to be promoted in the human visual system. This paper proposes a secure binary image steganography method by a generative adversarial network (GAN). The generator part of GAN simulates stego images, and the discriminator is designed to discriminate between the stego image produced by the generator and the cover image. The proposed GAN can automatically learn the most suitable flipped pixels in a binary image. Firstly, we learn the probability of embedded change from each pixel in the binary image, which can be converted into an embedded distortion map. Then we design an embedded function to simulate the steganography of the binary image. Experimental results show that the proposed method can find more suitable texture areas to embed secret information under the premise of ensuring security with fewer pixels flipped and better visual effects. The proposed network structure is different from the traditional binary image steganography by achieving more advanced content-adaptive embedding. Meanwhile, the proposed method is the first to apply GAN structure to the field of binary image steganography.