Deep learning semantic image synthesis: a novel method for unlimited capacity, high noise resistance coverless video steganography

Nowadays coverless steganography has fascinated huge attentiveness in the field of information security as it does not modify the transmission carrier of the secret information so it won’t be detected by steganalysis algorithms constructively. Yet, the present techniques have a limited capacity for transferring bits and a few of them use deep neural networks, plus most coverless techniques don’t use videos as carriers and benefit from the privilege of having more abundant content. To address these problems, a coverless video steganography technique is proposed that has unlimited hiding capacity and provides effective resistance against different noises. This paper utilizes two different deep neural networks to alter the video frames to new ones and extract information. The first one is a deep high-resolution network that segments the input frame and converts it to a label map. Then with a generative adversarial network, the semantic label is synthesized into a new frame from which the receiver acquires information in a way that makes it independent of using other frames. The sender should also dispatch the receiver some auxiliary info about each frame. The method is evaluated on two popular datasets: common objects in context (COCO), and the densely annotated video segmentation (DAVIS) dataset. Ostensibly, this is the first method with infinite capacity per each video frame. The proposed has great noise robustness of more than 29% compared to its rival on the DAVIS dataset, and also about 12% stronger against common noise compared to some state of art methods. Besides this technique can effectively resist steganalysis tools. Some methods have also been introduced to limit the effect of noise in the proposed method.