Texture adaptive steganography via Convolutional Neural Networks

For the investigation of the steganography, most of the current algorithms focus on how to describe the noise of an image. This paper proposes a new paradigm for steganography to automatically evaluating the modification priority of each pixel in image by using Convolutional Neural Networks (CNN). An image is decomposed into several sub-images which center pixel is modified, steganalysis method is used to judge which sub-image is secure, this suggests the center pixel suitable to embed message or not and obtain an appropriate priority. Corresponding stenographic cost assigned to each pixel on the foundation of priority. Experiments show that the proposed method achieves the state-of-the art performance on resisting advanced steganalysis.