SI-HILL is a side-informed variant of HILL for spatial-domain steganography in uncompressed images generated by Stable Diffusion. It uses side information from the image synthesis pipeline, specifically the quantization residuals available before rounding the continuous VAE decoder output to 8-bit pixels.
Use in tools
SI-HILL is currently represented in StegoRank as an experimental research method, not as an end-user tool. It extends the standard HILL cost function by reducing the cost of modifications whose direction is consistent with the quantization residual of the generated image.
Detectability results
In the evaluated Stable Diffusion setting, SI-HILL substantially reduces detectability compared with conventional HILL. 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.