STNet: A Style Transformation Network for Deep Image Steganography

摘要

Image steganography is the technique of hiding information within images in plain sight. With the rapid development of deep learning in the field of steganalysis, it becomes a tremendous challenge to design a secure steganographic algorithm. To this end, we propose a novel stegano-graphic network based on style transfer, named STNet. This network accepts the content and style images as input to synthesize art image with content from the former and style from the latter and embeds the secret information in style features. It can effectively resist most steganalysis tools. Steganalysis can identify stego images from cover images, but they cannot, distinguish our stego images from other art images. Meanwhile, our method produces stego images of arbitrary size with 0.06 bit per pixel, improving over other deep steganographic models which only can embed fixed-length secret. Experiment results demonstrate that our STNet can achieve great visual effect, security, and reliability.