Generative Steganography Diffusion

Ping Wei, Qing Zhou, Zichi Wang, Zhenxing Qian, Xinpeng Zhang, Sheng Li

Abstract

Generative steganography (GS) is an emerging technique that generates stego images directly from secret data. While GAN-based methods produce realistic images, they often fail at perfect data recovery due to a lack of invertibility. Conversely, Flow-based methods offer reversibility but suffer from poor image quality. To address this, we propose Generative Steganography Diffusion (GSD), which utilizes an invertible diffusion model called "StegoDiffusion." By modeling the generation process as an Ordinary Differential Equation (ODE), GSD enables 100% recovery of hidden data while maintaining high-fidelity image synthesis. Our method achieves a massive 1 bit-per-pixel (bpp) payload and significantly outperforms existing GS methods across all security and quality metrics.

Introduction

Steganography aims to hide information within cover media such as images. Traditional methods modify existing images, leaving statistical traces that steganalysis tools can detect. Generative Steganography (GS) avoids this by synthesizing new "stego images" directly from secret data.

Overview of the Generative Steganography Diffusion (GSD) framework.

However, previous GS frameworks faced a critical trade-off: GANs provided realism but lacked the mathematical invertibility required for perfect data extraction, while Flow-based models were reversible but produced low-quality images due to architectural constraints. GSD bridges this gap by leveraging the expressive power of diffusion models within a reversible ODE framework.

Methodology: The StegoDiffusion Model

The core innovation of GSD is the StegoDiffusion model, which adapts Denoising Diffusion Implicit Models (DDIM) for secure data embedding.

ODE Formulation and Reversibility

The transition between Gaussian noise and a clear image in diffusion models can be modeled as a discrete Ordinary Differential Equation (ODE). By utilizing an Euler iteration formula, the path becomes deterministic. This allows GSD to use a standard, highly expressive U-Net architecture—which is not inherently invertible—while maintaining a mathematically reversible process for data extraction. By setting the noise parameter ($\eta$) to zero, the path from noise to image becomes a fixed trajectory that can be traversed backward to recover the original latent state.

Frequency Domain Hiding

Rather than hiding data in pixel space, GSD embeds binary secret data into the Discrete Cosine Transform (DCT) coefficients of the initial latent noise vector ($z$). This approach makes the hidden data resilient to quantization noise—the slight errors introduced when converting floating-point values to 8-bit integers (0-255).

Detailed architecture of the StegoDiffusion model, illustrating the data hiding and extraction processes.

Deterministic and Fast Sampling

To overcome the computational cost of standard diffusion, GSD employs a fast sampling technique using a sub-sequence of steps (e.g., 50 steps). This maintains the deterministic nature of the ODE solver, allowing for efficient generation and extraction without sacrificing accuracy.

Experimental Results

The authors evaluated GSD on CIFAR-10, CelebA, and LSUN-Bedroom datasets. The results demonstrate a significant leap over state-of-the-art (SOTA) methods:

  • Extraction Accuracy: GSD achieved near 100% accuracy (99.9% to 100%) across all benchmarks, solving the reliability issues of GAN-based models.
  • Image Quality: On the CelebA dataset, GSD reached a Fréchet Inception Distance (FID) of 8.10, a massive improvement over the previous best reversible method (GSF), which scored 22.45.
  • Security: Steganalysis tests using Ye-net yielded a security score ($P_e$) near 0.5, indicating that the generated stego images are statistically indistinguishable from clean synthetic images.
  • Efficiency: Using fast sampling, the model processes images in approximately 0.13 seconds on modern hardware.

Discussion and Limitations

GSD successfully breaks the invertibility constraint, proving that powerful, non-invertible neural networks can be used for perfectly reversible steganography. However, some challenges remain. The current experiments focus on low-resolution images (up to 64×64); scaling to high-definition resolutions may introduce new complexities in maintaining reversibility. Additionally, while GSD is secure against passive detection, it is not yet designed to be robust against active attacks like JPEG compression or cropping, which would disrupt the deterministic ODE path.

Conclusion

Generative Steganography Diffusion (GSD) represents a new frontier in covert communication. By combining the high-fidelity generation of diffusion models with the mathematical rigor of ODE-based reversibility, GSD achieves the "holy grail" of steganography: high capacity, perfect recovery, and superior visual quality.

Summary updated

Abstract

Generative steganography (GS) is an emerging technique that generates stego images directly from secret data. Various GS methods based on GANs or Flow have been developed recently. However, existing GAN-based GS methods cannot completely recover the hidden secret data due to the lack of network invertibility, while Flow-based methods produce poor image quality due to the stringent reversibility restriction in each module. To address this issue, we propose a novel GS scheme called "Generative Steganography Diffusion" (GSD) by devising an invertible diffusion model named "StegoDiffusion". It not only generates realistic stego images but also allows for 100\% recovery of the hidden secret data. The proposed StegoDiffusion model leverages a non-Markov chain with a fast sampling technique to achieve efficient stego image generation. By constructing an ordinary differential equation (ODE) based on the transition probability of the generation process in StegoDiffusion, secret data and stego images can be converted to each other through the approximate solver of ODE -- Euler iteration formula, enabling the use of irreversible but more expressive network structures to achieve model invertibility. Our proposed GSD has the advantages of both reversibility and high performance, significantly outperforming existing GS methods in all metrics.
BibTeX
@misc{wei2023generative,
  author = {Ping Wei and Qing Zhou and Zichi Wang and Zhenxing Qian and Xinpeng Zhang and Sheng Li},
  title = {{Generative Steganography Diffusion}},
  year = {2023},
  eprint = {2305.03472},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2305.03472},
}
Figure