A novel gradient-guided post-processing method for adaptive image steganography

Designing an effective cost function has always been the key in image steganography after the development of the near-optimal encoders. To learn the cost maps automatically, the Generative Adversarial Networks (GAN) are often trained from the given cover images. However, this needs to train two Convolutional Neural Networks (CNN) in theory and is thus very time-consuming. In this paper, without modifying the original stego image and the associated cost function of the steganography, and no need to train a GAN, we proposed a novel post-processing method for adaptive image steganography. The post-processing method aims at the embedding cost, hence it is called Post-cost-optimization in this paper. Given a cover image, its gradient map is learned from a pre-trained CNN, which is further smoothed by a low-pass filter. The elements of the cost map derived from the original steganography are projected to 0,1 for separating embeddable and non-embeddable areas. For embeddable areas, the elements will be further screened by the gradient map, according to the magnitudes of the gradients, to produce a new cost map. Finally, the new cost map is used to generate new stego images. Comprehensive experiments have validated the efficacy of the proposed method, which has outperformed several state-of-the-art approaches, whilst the computational cost is also significantly reduced.