SI-UNIWARD is a side-informed version of UNIWARD. It is a general strategy that can be applied in different settings, such as JPEG, Stable Diffusion images, or other scenarios where useful side information is available.
In the StegoRank evaluation described here, SI-UNIWARD is applied to images generated with Stable Diffusion. In that setting, it uses quantization residuals from the image synthesis pipeline as side information to guide embedding changes.
Quick Summary
- Method:
- side-informed variant of UNIWARD.
- Evaluated setting:
- AI-generated images from Stable Diffusion, evaluated as uncompressed images.
- Side information:
- quantization residuals from the synthesis pipeline in the evaluated protocol.
- Use:
- experimental research method, not an end-user tool.
- Main reading:
- it reduces the detectability of conventional UNIWARD by using side information available before rounding to 8-bit pixels.
Use in tools
SI-UNIWARD is currently represented in StegoRank as an experimental research method, not as an end-user tool. In the evaluated Stable Diffusion protocol, it extends the standard UNIWARD cost function by reducing the cost of modifications whose direction agrees with the residual between the continuous VAE decoder output and the rounded 8-bit image.
Detectability results
In the evaluated Stable Diffusion setting, SI-UNIWARD substantially reduces detectability compared with conventional UNIWARD. At 0.10 bits per pixel, the reported detection accuracy is close to chance level against both EfficientNet-B0 and SRNet under the same-source supervised steganalysis protocol.
The result should be interpreted in the context of the evaluated protocol: Stable Diffusion v2.1, 512×512 RGB images, side information from the VAE decoder output, and detectors trained and tested on images generated under the same configuration. At higher payloads, especially 0.20 and 0.40 bpp, the method becomes more detectable.
Reference paper
- On the Effectiveness of Side-Informed Steganography in Diffusion-Generated Images. Daniel Lerch-Hostalot and David Megías. In Proceedings of the 21st International Conference on Availability, Reliability and Security (ARES ‘26). To appear, August 2026.