Abstract: Driven by noise, the generator of the generative adversarial network (GAN) can generate high-quality digital images and provide a new data carrier for hiding information. In this paper, a coverless information hiding model combining orthogonal GAN and Ushape network is proposed on the fact that the orthogonal GAN discriminator can extract feature codes of the generated image. While hiding information, the binary sequence of the information to be hidden is mapped into a noise vector according to a group quantization rule, and the generator of the orthogonal GAN is driven by the noise vector to generate a hidden digital image. While extracting information, the feature code of the hidden image is extracted by the discriminator of the orthogonal GAN, and then the U-shape network is used to realize the mapping from the feature code to the driving noise, thereby recovering the secret information. By performing adversarial training of the built-in model with the CelebA dataset, the generator can generate high-quality hidden images and the discriminator can be combined with the U-shaped network to extract secret information from the hidden image. Compared with the latest similar algorithms, the proposed model performs better information extraction accuracy and security under the same information hiding capacity with reduced the training overhead and improved practicability.

Key words: information hiding, deep learning, generative adversarial network(GAN), Unet