A coverless image steganography method based on invertible neural network

Recently, coverless image steganography (CIS) has become a hot topic in the field of image steganography. However, currently most CIS methods do not build an invertible two-way mapping between the secret information and image, so that the steganography capacity of CIS is limited. Moreover, the generated stego-images have a visible distinction between real-images, and it is difficult to deceive the deep-learning-based steganalysis, which makes CIS unsafe. To address the above issues, we propose a coverless image steganography method based on invertible neural networks, we call it as CIS-INN. Our method uses an invertible flow-based model, Glow model, and we combine it with adversarial generative network (GAN) to improve generated images quality. In encryption phase, secret information is encoded into latent variables by proposed Gray-code-based coding method, then the Glow model takes these latent variables as a prior and generates stego-images. To against steganalysis, we make CIS-INN more secure by introducing adversarial examples when generating stego-images. In decryption phase, secret information can be recovered through the reverse process of the glow model. Experiments demonstrate the CIS-INN achieves significant improvement of the steganography capacity (4 BPP) and maintains reliable security when confronted with multiple steganalysis methods.