Secure Steganography Scheme Based on Steganography Generative Adversarial Network

Taking into consideration the security concerns surrounding steganography due to the infiltration of dense images during the image stealth process, this article proposes the addition of a noise layer between the embedded information and decoding information of the SteGAN model in order to train the dense image to resist noise attacks. To improve the stabilization of the discriminator's training, the principle of spectral normalization is applied to mitigate the convergence speed of the discriminator. This will enable the networks to continue training and enhance the confrontation between stealth and stealth analysis. Finally, the model is trained on the COCO dataset. The experimental results reveal that the dense image obtained through noise layer training performs well in terms of imperceptibility, with a decoding accuracy rate of more than 90% for all confidential images trained by the noise, except for JPEG compression. Moreover, it can effectively protect against the detection of stealth analysis tools.