A stable GAN for image steganography with multi-order feature fusion

Content-adaptive automatic cost learning frameworks for image steganography based on deep learning can generate a more exquisite embedding probability map within a short time; such methods have reached remarkable security performance compared with conventional handcraft-based methods. However, some issues in deep steganography are not discussed in spatial domain: (1) the key point to the design of generator has not reached clearly; and (2) existing methods are unstable of model training due to vanishing gradient problem. To investigate these issues, this paper proposes a stable GAN (generative adversarial network) for image steganography called UMC-GAN, which presents a redesigned and adjustable nested U-Shape generator and utilizes deep supervision to fuse multiple embedding probability maps to improve security performance. A novel linear-clipped embedding simulator is designed to alleviate vanishing gradient problem at the staircase regions. Extensive experiments and ablation studies show that the proposed method outperforms existing GAN-based automatic cost learning embedding frameworks, and it can be applied at high resolution through the flexible adjustment of the generator. Further investigation on the design of generator is explored by model pruning which shows that in-depth features should be captured for deep steganography to ensure the security performance.