Adaptive spatial steganography based on adversarial examples

Recently, the researchers start to apply adversarial attack to enhance the security of steganographic algorithms. The typical deep learning model is vulnerable to adversarial attack. Such attack is generating special instance via neural network. The generated instance can increase the detection error of the steganalyzer. In this paper, we propose a practical adversarial method to enhance the security of typical distortion-minimizing steganographic algorithms. The proposed method is an adaptation of the Fast Gradient Sign Method in the steganography. We utilize the gradients back-propagated from the deep-learning steganalyzer to control the changing direction of the pixels. This kind of steganaographic modification in the image helps to improve the security towards the steganalysis. The experimental results prove that the proposed method can enhance the security of typical distortion-minimizing steganaographic algorithms.