A Review of Deep Learning Based Image Steganography Methods
Wencui Yang, Zhu Ting-ge, Ying Liu · 2024
Traditional image steganography methods are mainly based on the null or transform domain to embed secret images, which have the problems of limited hiding capacity, low visual quality and weak resistance to steganalysis. By introducing deep learning into image steganography, the performance of image steganography is greatly improved. This thesis mainly reviews the image steganography methods based on deep learning in recent years. Firstly, according to whether the cover image needs to be modified during steganography, the deep learning-based image steganography methods are divided into two major categories: cover-modified image steganography and cover-less image steganography, and existing deep learning-based steganography methods are categorised and summarised from two aspects. Then the commonly used datasets and evaluation indexes for image steganography are introduced, and the performance of deep learning-based image steganography methods is compared and analysed. Image steganography methods based on cover modification are superior to coverless image steganography in terms of hiding capacity, but relatively weak in terms of resistance to steganography analysis. Finally, the challenges faced by deep learning-based image steganography methods are discussed, and future research directions are pointed out.