Saliency Map Construction for Adversarial Image Steganography

Adversarial image steganography can fool the targeted CNN (convolutional neural network)-based steganalyzers to improve the security performance. Despite existing works have achieved great success, there are still some limitations make it difficult to exploit their potentiality, including that selecting a final stego image from the candidate stego images cannot perfectly help them fool the targeted steganalyzers. In this article, we design a new model to score each image element in a cover image by the trade-off between its absolute value of gradient and the original embedding cost, and saliency map is constructed to represent the score of the image. Based on the proposed method, a simple and efficient adversarial steganographic scheme SAL is presented. It selects the elements from the map according to the amplitudes of scores, and their costs are updated based on the sign of corresponding gradients. Finally, data embedding is accomplished with the new costs to get an adversarial stego image. 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.