Message Drives Image: A Coverless Image Steganography Framework Using Multi-Domain Image Translation

The coverless steganography methods based on the GAN satisfy the need of large image set which is a shortcoming for traditional coverless methods. But most of GAN based methods have either low hiding capacity or weak robustness. In this paper, to solve these problems we design a coverless image steganography framework based on multi-domain image translation called MDI. The target label used to translation is built a relationship with secret message. This framework includes a generator and a multi-label classifier. Generator is used to translate cover image into stego image according to target label. Classifier is adopted to extract label in the image and obtain secret message. Moreover, common types of image attacks are considered in advance to improve the robustness of classifier. Finally, our experiments and discussion show our method has more robustness and higher hiding capacity than state of art coverless image steganography methods.