GAN-based image steganography for enhancing security via adversarial attack and pixel-wise deep fusion

In recent years, the development of steganalysis based on convolutional neural networks (CNN) has brought new challenges to the security of image steganography. However, the current steganographic methods are difficult to resist the detection of CNN-based steganalyzers. To solve this problem, we propose an end-to-end image steganographic scheme based on generative adversarial networks (GAN) with adversarial attack and pixel-wise deep fusion. There are mainly four modules in the proposed scheme: the universal adversarial network is utilized in Attack module to fool CNN-based steganalyzers for enhancing security; Encoder module is seen as the generator to implement the pixel-wise deep fusion for imperceptible information embedding with high payload; Decoder module is responsible for the process of recovering embedded information; Critic module is designed for the discriminator to provide objective scores and conduct adversarial training. Besides, multiple loss functions together with Wasserstein GAN strategy are applied to enhance the stability and availability of the proposed scheme. Experiments on different datasets have verified the advantages of adding universal adversarial perturbations for higher security against CNN-based steganalyzers without compromising imperceptibility. Compared with state-of-the-art methods, the proposed scheme has achieved better performance in security.