Layerwise Adversarial Learning for Image Steganography

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