Steganography with Convincing Normal Image from A Joint Generative Adversarial Framework

Image steganography conceals secret message into digital image without influencing human perception. Recently, a steganographic method based on generative adversarial networks (GANs) has been tentatively applied to human face dataset in order that the generated images can evade being detected by steganalytic methods. In this paper, we propose a more effective GAN-based steganographic framework, named VAE-SGAN, which combines together several deep learning based network structures: the encoder, the decoder/generator, the discriminator, and the steganalyser. This proposed model can generate better visually convincing images with less model collapse. Through comparative experiments, it has been proved that the generated images are more secure against steganalysis than those generated by the previously established GAN-based methods when working under some popular steganography schemes, such as LSB-matching, WOW and S-UNIWARD.