Neural style transfer for image steganography and destylization with supervised image to image translation

Mallika Garg, Jagpal Singh Ubhi, Ashwani Kumar Aggarwal

Abstract: Today, a lot of information is being shared electronically in a way or another. Despite the advancements in technology used for data transfer, the reliable transmission of sensitive data is still a major challenge that need to be addressed. In this paper, we propose a steganographic technique to generate a stego image using the Neural Style Transfer (NST) algorithm that maintain the perceptual quality of the stego image with maximum capacity payload. Along with this, we propose to recover the secret content from generated stego image with minimal distortion. So, to recover the secret image from the generated stego image, destylization is performed using conditional Generative Adversarial Networks (cGANs). The proposed destyling GAN is forced to learn the embedded secret information using a loss function that learns the same representation as in the embedding NST algorithm. This whole framework of embedding and extraction of secret image is evaluated for Imagenet dataset and later tested on the PASCAL VOC12 dataset. The algorithm outperforms in terms of Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Visual Information Fidelity (VIF) with 44.175 dB, 0.9958, and 0.954 respectively for the Imagenet dataset. Also, the proposed algorithm is more robust against StegExpose with 0.529, area under the curve (AUC).