Layerwise Adversarial Learning for Image Steganography

Image steganography is a subfield of pattern recognition. It involves hiding secret data in a cover image and extracting the secret data from the stego image (described as a container image) when needed. Existing image steganography methods based on Deep Neural Networks (DNN) usually have a strong embedding capacity, but the appearance of container images is easily altered by visual watermarks of secret data. One of the reasons for this is that, during the end-to-end training process of their Hiding Network, the location information of the visual watermarks has changed. In this paper, we proposed a layerwise adversarial training method to solve the constraint. Specifically, unlike other methods, we added a single-layer subnetwork and a discriminator behind each layer to capture their representational power. The representational power serves two purposes: first, it can update the weights of each layer which alleviates memory requirements; second, it can update the weights of the same discriminator which guarantees that the location information of the visual watermarks remains unchanged. Experiments on two datasets show that the proposed method significantly outperforms the most advanced methods.