A Generative Steganography Method Based on WGAN-GP

Abstract

With the development of Generative Adversarial Networks (GAN), GAN-based steganography and steganalysis techniques have attracted much attention from researchers. In this paper, we propose a novel image steganography method without modification based on Wasserstein GAN Gradient Penalty (WGAN-GP). The proposed architecture has a generative network, a discriminative network, and an extractor network. The Generator is used to generate the cover image (also is the stego image), and the Extractor is used to extract secret information. During the process of stego image generation, no modification operations are required. To make full use of the learning ability of convolutional neural networks and GAN, we synchronized the training of Generator and Extractor. Experiment results show that the proposed method has the advantages of higher recovery accuracy and higher training efficiency.