Convolutional Network for Image Steganography With Redundant Embedding

Abstract

Image steganography is one of the secure methods of information communication. Based on deep learning, the steganography models have obtained better performance than those based on handcraft features, but they can not ensure absolute correctness when extracting bit message. To improve extracting accuracy, we propose a new redundant embedding method. Also, to improve robustness and security against steganalysis, we introduce generative adversarial training into our model. From the experimental results, our proposed methods reduce the extracting inaccuracy significantly while maintaining the capability of resisting steganalysis attack. What's more, our proposed methods can be easily generalized to cover images with different size and embedding capacity of different bit message length. Therefore, the proposed model can have a wider application in real use.