Generative Adversarial Network based Steganography with different Color Spaces

Abstract

Traditional steganographic uses an approach where various steps of a steganographic algorithm are devised by human experts. This process can be automated with Generative Adversarial Networks. The use of Generative Adversarial Networks (GAN) in the field of steganography helps in generating suitable and secure covers for steganography with no need for human based algorithms. The network learns and evolves to replace the role played by a steganographic algorithm in generating robust steganalyzers without human intervention. All the GAN based steganographic models use RGB images for hiding the secret data. In this work, the impact of various color spaces on GAN based steganography application is explored. Steganographic images in different color formats such as RGB, YCrCb, YIQ, YUV, CIEXYZ, YDbDr, HED and HSV are generated using DCGAN based model to study the importance of color spaces in steganography. The results of the experimentation on CelebA dataset show that the color spaces play an important role in GAN based steganography. The error rate and the message extraction accuracy of a model vary significantly with different color spaces. The experimental analysis depicts that color spaces such as HED, YUV and YCrCb perform better than RGB and other color spaces in terms of distortion, extraction accuracy and convergence for the same number of epochs.