Saliency Map Construction for Adversarial Image Steganography

Adversarial image steganography can fool the targeted convolutional neural network (CNN)-based steganalyzers, thereby improving the security performance. Despite the fact that existing works have achieved great success, there are still some limitations that make it difficult to exploit their potentiality, including the issue that selecting a final stego from the candidate stegos cannot perfectly help them fool the targeted steganalyzers. Since the trade-off between gradient and embedding cost has not been thoroughly investigated, this may simplify the design of more effective methods. In this article, we design a new model to score each image element in a cover by utilizing this trade-off, and a saliency map is constructed to represent the scores of the image. Based on the above, a simple and efficient scheme called SAL is presented. It selects the elements from the map according to the amplitudes of the scores, and their costs are updated based on the signs of the corresponding gradients. Finally, data embedding is accomplished with the new costs to generate an adversarial stego. Extensive experiments illustrate that SAL can achieve better security performance than state-of-the-art methods under different targeted CNN-based steganalyzers in both spatial and JPEG domains.