Deep learning based image steganography: A review

Abstract

A review of the deep learning based image steganography techniques is presented in this paper. For completeness, the recent traditional steganography techniques are also discussed briefly. The three key parameters (security, embedding capacity, and invisibility) for measuring the quality of an image steganographic technique are described. Various steganography techniques, with emphasis on the above three key parameters, are reviewed. The steganography techniques are classified here into three main categories: Traditional, Hybrid, and fully Deep Learning. The hybrid techniques are further divided into three sub‐categories: Cover Generation, Distortion Learning, and Adversarial Embedding. The fully Deep Learning techniques, based on the nature of the input, are further divided into three sub‐categories: GAN Embedding, Embedding Less, and Category Label. The main ideas of the important deep learning based steganography techniques are described. The strong and weak features of these techniques are outlined. The results reported by researchers on benchmark data sets CelebA, Bossbase, PASCAL‐VOC12, CIFAR‐100, ImageNet, and USC‐SIPI are used to evaluate the performance of various steganography techniques. Analysis of the results shows that there is scope for new suitable deep learning architectures that can improve the capacity and invisibility of image steganography.