[Paper Review] Generative Steganography Diffusion

Wei Ping, Qing Zhou+4 more|arXiv (Cornell University)|May 5, 2023

Advanced Steganography and Watermarking Techniques6 citations

TL;DR

This paper proposes Generative Steganography Diffusion (GSD), a novel generative steganography scheme that uses an invertible diffusion model called StegoDiffusion to generate high-quality stego images while enabling 100% secret data recovery. By modeling the image generation process as an ordinary differential equation (ODE) solvable via Euler iteration, GSD achieves full reversibility and superior performance over existing GAN- and Flow-based methods.

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.

Motivation & Objective

• Address the limitations of existing generative steganography methods, particularly the trade-off between image quality, payload capacity, and data recovery accuracy.
• Overcome the irreversibility of GAN-based methods and the quality degradation in Flow-based methods due to strict invertibility constraints.
• Develop a generative steganography framework that ensures perfect secret data recovery while maintaining high image fidelity.
• Enable the use of expressive, non-invertible neural networks by introducing an ODE-based invertible transformation mechanism.

Proposed method

• Design a non-Markovian diffusion process in StegoDiffusion to accelerate image generation and reduce sampling steps.
• Construct an ordinary differential equation (ODE) based on the transition probability of the diffusion process to enable model invertibility.
• Use the Euler iteration formula as an approximate solver for the ODE to perform bidirectional mapping between stego images and secret data.
• Integrate secret data into the latent space of the diffusion model during the forward diffusion process, enabling end-to-end generation and extraction.
• Train the generator and extractor jointly using a diffusion-based architecture that supports both high-fidelity image synthesis and exact data recovery.
• Leverage fast sampling techniques to reduce inference time while preserving high image quality and perfect recovery accuracy.

Experimental results

Research questions

• RQ1 Can a diffusion-based generative steganography model achieve 100% secret data recovery while maintaining high image quality?
• RQ2 Can an ODE-based formulation enable invertibility in a generative steganography model without requiring strictly invertible network modules?
• RQ3 Does the proposed GSD scheme outperform existing GAN- and Flow-based generative steganography methods across all key metrics: payload, accuracy, FID, and security?
• RQ4 How does the proposed method perform under standard steganalysis attacks, particularly in terms of detection resistance?

Key findings

• GSD achieves 100% secret data extraction accuracy at a payload of 1 bpp, surpassing all prior methods, including GSF (99.96%) and GSN (98.15%).
• The FID score of GSD is 8.10, significantly lower than GSF (22.45) and GSN (12.88), indicating superior image quality.
• The ROC AUC value of GSD on CelebA and Bedroom datasets is close to 0.5 (0.4960), demonstrating strong resistance to steganalysis.
• The absolute mean error between the original and recovered stego latent is less than 1e-2, confirming near-perfect reversibility of the ODE-based transformation.
• Using 50 sampling steps (S=50), GSD generates high-quality stego images efficiently, with the ability to scale to higher S for improved performance.
• The method successfully enables the use of expressive, non-invertible networks by embedding invertibility in the ODE solver rather than the network architecture.