ACGIS: Adversarial Cover Generator for Image Steganography with Noise Residuals Features-Preserving

Abstract

Recent works with the technique of adversarial example have been bringing the possibility of effectively resisting the machine learning-based steganalyzers. Nevertheless, these methods likely introduce unexpected artifacts and destroy the statistics when adding adversarial perturbations. In this paper, under the assumption of similarity between the noise residuals of normal image sub-regions, we propose a Siamese generator to learn and preserve sub-regions noise residuals features for minimizing the impact of adversarial perturbations on similarity. The cover and stego sub-regions pairs are used as the input, which incorporates steganography domain knowledge to further encourage the generator to yield the more favorable adversarial covers. Moreover, during interactive training with steganalyzer, using a random embedding strategy to replace the specific steganographic algorithm saves training time and improves the generalization. We can employ the trained generator to produce numerous adversarial covers, cooperating with the existing steganographic methods to embed secret messages achieving much safer steganography. Security analysis and experiments show that the generated adversarial covers are superior in terms of quality and security.