Abstract:

Image information hiding is one of the important methods to ensure information security. With the growth of deep learning, numerous deep learning-based image to image steganography models have been presented. Most of them are deficient in terms of image quality, hiding security, or embedding capability balance. So, this paper proposed an image information hiding algorithm based on cross-domain adversarial adaptation to address the above problems. First, a super-resolution network was built to embed the secret information into the image content unaffected by zooming in and out, to increase the secret information’s embedding capability. Then, an attention mechanism was introduced to the encoding network to enable the network to focus on the primary features and suppress superfluous features, so enhancing the image’s resolution. Finally, a domain adaption loss was introduced to the generator network to guide the production of the stego image, and the model was trained in a generative adversarial way to reduce the cross-domain difference between the carrier image and the stego image. The experimental results demonstrate that, compared to other steganography techniques, the proposed algorithm improves the security and embedding capability of information hiding while maintaining image quality.

Key words: information hiding, generating adversarial network, domain adaptation, self-attention mechanism