Steganography is the process of concealing a secret message in ordinary digital media by making small modifications that try to preserve cover statistics. In this paper, a novel unified steganographic framework is introduced inspired by image-to-image translation networks and adversarial learning. We consider image steganography as an image-to-image translation which is a transformation from cover distribution to stego distribution. The image-to-image translation models can exploit imperceptible perturbations to embed useful input information into the target image so that the input image can be retrieved more accurately. So, we adapt the CycleGAN as a popular translator to conceal the secret message in the form of an imperceptible, high-frequency signal in the container image called stego, while the cyclic consistency constraint helps to extract the secret message. As deep neural networks are sensitive to small adversarial perturbations, adversarial learning is exploited to embed the secret message more semantically to preserve the statistical characteristics of the cover as much as possible. According to the competitive relation of steganography and steganalysis, the steganographic methods act as an adversarial attack to mislead the steganalysis network toward determining the input images as the cover images. Therefore, in the steganographic concept, a CNN-based steganalyzer is considered as the discriminator such that the steganographer and the steganalyzer alternately adjust their weights according to the updated knowledge of one another. On the other hand, we append the smoothness constraint into the objective function to encourage the model to encode the secret message as smooth embedding modifications to improve the secrecy of generated image. This constraint limits the range of local differences and enforces a stronger correlation between the cover and the corresponding stego image. Smoothing modifications mitigate the statistical impact and enhance the robustness against steganalyzers. Experiment results show the efficiency of our proposed framework which achieves an appropriate trade-off between secrecy and capacity and outperforms the existing works.
Read full abstract