Normalizing Flows-Based Latent Variable Rearrangement for Generative Image Steganography

Image steganography is a technology that hides secret messages within images. Traditional steganography methods typically rely on modifying the cover image to conceal secret messages, which makes them susceptible to detection and attack. Recently, generative image steganography, which conceals messages during image generation, provides a more effective approach to address the challenge. However, the existing generative image steganography methods often impact the quality of the generated image and the data distribution due to embedding messages. Additionally, they are often criticized for the poor steganographic capacity and extraction accuracy. To overcome the challenges, we propose a Normalizing flows-based generative steganography method. Our method rearranges the values of the Gaussian latent variables which are used to generate the image based on the secret messages and the mapping rules. Secret messages are embedded within the adjusted Gaussian latent variables. The generated image is not affected by embedding messages because the values of the latent variables and the distribution are not changed. Experimental results demonstrate the effectiveness of our method in terms of imperceptibility, steganographic capacity and extraction accuracy.