Somewhat Reversible Data Hiding by Image to Image Translation.

The traditional reversible data hiding technique is based on image modification which is more easily analyzed and attacked. With the appearance of generative adversarial networks (GANs), a novel type of image steganography methods without image modification have received a great deal of attention since they have the ability to resist detection by traditional steganalysis methods. In this paper, a novel somewhat reversible data hiding scheme (SRDH) by image to image translate is proposed. This scheme consists of two processes. In prepare process, two GAN models and an extractor are trained. In implementation process, data extraction and image recovery are separately performed by the trained extractor and restorer. We call the scheme somewhat because it is not completely reversible. Experimental results have verified the correctness of the scheme.