Contrastive Noise-Guided Invertible Network for Image Steganography

Image steganography aims to produce stego images through hiding secret images in the cover images to achieve covert communication. To simultaneously improve the invisibility and revealing quality of covert communication, this paper proposes a Contrastive Noise-Guided Invertible Network (CNGI-Net). Specifically, the Noised-Guided Invertible (NGI) mechanism is first proposed to actively and gradually generate adaptive noise in the forward hiding process using the interaction mechanism. The generated adaptive noise effectively guides and adjusts the cover-secret fusion process, as well as blurs the secret information in the stego images, which can significantly improve the transmission security. Besides, benefiting from the reversible property of NGI, high-quality revealed secret images can also be progressively decoupled from the stego images along the backward flow of NGI. To further ensure the invisibility and security of communication, Contrastive Hiding Learning (CHL) is proposed to improve the cover-stego similarity using the contrastive information. Extensive experiments demonstrate that the proposed CNGI-Net significantly promotes concealment security of transmission process and the quality of revealing for covert communication. Especially, the quality of stego and revealed secret images on $\{1,2,3\}$ image hiding has been respectively promoted by $\{1.53,0.43,0.38\}$ dB and $\{4.26,3.45,4.98\}$ dB on ImageNet dataset, compared to SOTA steganography methods.