Large capacity generative image steganography via image style transfer and feature-wise deep fusion

Compared with modification-based steganography, the coverless steganography models have stronger anti-detection performance. However, the limitations of low embedding capacity and image quality are existing in current coverless steganographic models. The image style transfer is a generation task that translates a style to another while maintaining the structure and semantics of the original image. The stable structures that can be used as cover to hide secret for steganography. In this paper, a generative steganographic model based on style transfer and feature-wise deep fusion is proposed, which can achieve a large embedding capacity and high generation quality. In embedding stage, the channel reduction net is designed to distill structural features from image, and the fusion net is proposed to fuse the secret matrix and distilled structural features. The embedded structural features are participated in style transfer to finish steganography. In secret recovery stage, the two-level extraction model is adopted. The structural features are extracted from stego image by residual network, then the embedded secret message is distilled from the extracted features. Compared with existing steganographic model, the proposed model is convenient applied without code-book or database, and the experimental results show that the model achieves a larger capacity, higher security and visual quality.