Enhancing image steganography security via universal adversarial perturbations

Image steganography refers to embedding secret information into a cover image without drawing perceptible distortions. Nevertheless, steganalyzers are potentially reveal steganography by detecting subtle modifications, especially with the introduction of deep learning into image steganalysis. Recent researches show that adversarial examples can greatly enhance the security of image steganography works. In this work, a new terminology of Universal Adversarial Perturbations (UAPs) is presented to further improve the security of image steganography. Specifically, we introduce a generator within the framework of generative adversarial networks (GAN) that learns to generate UAPs, where the UAPs can be applied to universal images without the need to design perturbation specific to an individual image. The UAPs are directly added to the embedding probability map of the image, which can make the generated stego image more deceptive. Experimental results show that the proposed UAPs can effectively improve the security of image steganography.