Reversible Data Hiding Using GAN Based Image-Image Transformation
R. Geetha
Abstract
Traditional Reversible Data Hiding (RDH) techniques modify the cover image, unavoidably leaving traces which can be easily analyzed and exploited by an adversary. Drawing inspiration from modifying cover steganography with Generative Adversarial Networks (GAN), this work presents a innovative generative RDH scheme (GRDH) based on image transformation. Initially, a realistic image is generated using an image generator, which is then processed through an image-to-image transformation model using CycleGAN. This process produces a marked image with different semantic data. Both the secret data and the original cover image can be independently recovered using a well-trained message extractor and the inverse transformation of the original image transformation. Experimental results confirm the effectiveness of this scheme resulting in a peak signal to noise ration (PSNR) of 25.78 dB and 31.65 dB for GAN models.