Generative steganography uses generative models to create images that carry hidden information. In StegoRank, this family includes both earlier GAN-based approaches, such as SteganoGAN, and more recent methods based on diffusion models.
Quick Summary
- Domain:
- generated images, especially scenarios involving GANs and general-purpose AI image models.
- Type:
- generation of stego images or use of generated images as carriers.
- Tools and methods:
- SteganoGAN and mas_GRDH.
- Main reading:
- detecting a specialized steganographic generator is not the same as distinguishing cover/stego images from a generator used for many types of images.
Why it matters
Generative approaches are conceptually different from classical embedding methods. Instead of modifying an existing cover image, they can generate an image as part of the hiding process.
This changes the detection problem. A detector may learn differences between generated and natural images rather than traces caused only by the hidden message. Diffusion-based approaches are especially relevant because they can use high-quality image generators such as Stable Diffusion.
Steganographic generators vs general-purpose generators
SteganoGAN and Stable Diffusion represent very different detection scenarios. SteganoGAN is a steganographic generator: its practical purpose is to generate images carrying hidden information. In that setting, detecting that an image comes from SteganoGAN may already be enough to flag it as stego, because the generator itself is tied to the hiding process.
Stable Diffusion is different. It is a general-purpose image generator used to create images of many kinds, most of them unrelated to steganography. In this case, detecting that an image was generated by Stable Diffusion is not sufficient. The relevant question is whether a Stable Diffusion image is a cover image or a stego image.
This distinction is important when reading the results. Detecting a specialized steganographic generator and detecting a hidden message inside images from a widely used image generator are not the same problem.
Typical tools and methods
The current comparison includes:
- SteganoGAN;
- mas_GRDH, a Stable Diffusion based generative steganography method described in the IEEE paper linked below.
Detectability results
In the Aletheia comparison, SteganoGAN is highly detectable across the evaluated payloads. This result should be read carefully: for SteganoGAN, high detectability may reflect that the image comes from a specialized steganographic generator, not only traces left by the hidden payload.
mas_GRDH is represented separately in the Stable Diffusion chart with one point at 0.02 bpp and accuracy 0.50, corresponding to chance-level detection in the reported setting.
Limitations
The results do not fully answer the broader question of detecting hidden messages in generated images. They describe specific methods, payloads, generators, and evaluation protocols.