ASIR: Steganography for Diffusion Models via Antipodal Sampling and Iterative Recovery

Yaofei Wang, Yufeng Zheng, Han Fang, Wenzhao Cao, Donghui Hu

Abstract: Messages embedded in diffusion generation noise suffer from severe attenuation due to denoising and VAE decoding, creating a persistent capacity–robustness trade-off. Identifying that extraction accuracy strictly correlates with the distance between candidate hypothesis images, we propose ASIR, a training-free and provably secure steganography framework for both pixel and latent diffusion models. ASIR introduces two key innovations: (i) Antipodal Sampling, which maximizes signal separation in probability space to enhance distinguishability, and (ii) Iterative Recovery, a paradigm shift that treats extraction as a gradient-based optimization problem to reverse non-linear distortions. Extensive experiments demonstrate that ASIR achieves state-of-the-art performance, embedding up to 65,536 bits (pixel-space) and 16,384 bits (latent-space) with 99\% accuracy, while remaining statistically undetectable to deep steganalyzers.

Lay Summary: AI image generators can create realistic pictures, but they can also be used to carry hidden messages for private communication, copyright protection, or authentication. The challenge is that when too much information is embedded, it becomes difficult to recover reliably from the final image. We introduce ASIR, a new method for embedding hidden information into AI-generated images without retraining the model. Our approach makes different message candidates easier to distinguish and then uses an iterative recovery process to better reconstruct the hidden message from the final image. This allows much larger messages to be embedded than in previous methods while still achieving very high recovery accuracy. At the same time, the generated images remain visually natural and difficult for detection systems to distinguish from ordinary AI-generated images. These results could support future applications in privacy protection, copyright verification, and secure communication.

Primary Area: Deep Learning->Generative Models and Autoencoders

Keywords: Generative Steganography, Diffusion Models

Submission Number: 21221