[Paper Review] Secret-to-Image Reversible Transformation for Generative Steganography

Zhili Zhou, Yuecheng Su+3 more|arXiv (Cornell University)|Mar 13, 2022

Advanced Steganography and Watermarking Techniques4 citations

TL;DR

This paper proposes a secret-to-image reversible transformation (S2IRT) scheme using the Glow flow-based generative model to enable high-capacity, accurate, and reversible information hiding in generative steganography. By leveraging bijective mappings between latent vectors and images, the method achieves up to 4 bpp hiding capacity with nearly 100% extraction accuracy while maintaining strong anti-detectability and imperceptibility.

ABSTRACT

Recently, generative steganography that transforms secret information to a generated image has been a promising technique to resist steganalysis detection. However, due to the inefficiency and irreversibility of the secret-to-image transformation, it is hard to find a good trade-off between the information hiding capacity and extraction accuracy. To address this issue, we propose a secret-to-image reversible transformation (S2IRT) scheme for generative steganography. The proposed S2IRT scheme is based on a generative model, i.e., Glow model, which enables a bijective-mapping between latent space with multivariate Gaussian distribution and image space with a complex distribution. In the process of S2I transformation, guided by a given secret message, we construct a latent vector and then map it to a generated image by the Glow model, so that the secret message is finally transformed to the generated image. Owing to good efficiency and reversibility of S2IRT scheme, the proposed steganographic approach achieves both high hiding capacity and accurate extraction of secret message from generated image. Furthermore, a separate encoding-based S2IRT (SE-S2IRT) scheme is also proposed to improve the robustness to common image attacks. The experiments demonstrate the proposed steganographic approaches can achieve high hiding capacity (up to 4 bpp) and accurate information extraction (almost 100% accuracy rate) simultaneously, while maintaining desirable anti-detectability and imperceptibility.

Motivation & Objective

• Address the inefficiency and irreversibility of existing secret-to-image transformations in generative steganography.
• Overcome the limited hiding capacity and low extraction accuracy in current GAN-based approaches.
• Enable a bijective, reversible mapping between secret messages and generated images for secure and reliable information hiding.
• Improve robustness against common image processing attacks through a separate encoding-based variant (SE-S2IRT).

Proposed method

• Utilizes the Glow model’s bijective mapping between a multivariate Gaussian latent space and complex image distributions for reversible transformation.
• Encodes secret messages into high-dimensional latent vectors using a structured, reversible encoding scheme.
• Maps the latent vector to a generated image via the trained Glow model, embedding the secret information in the image’s distribution.
• Employs a separate encoding-based S2IRT (SE-S2IRT) variant that reorganizes element positioning to enhance robustness against image distortions.
• Leverages the invertibility of the Glow model to enable perfect reconstruction of the original secret from the generated image.
• Applies statistical analysis and entropy-based security modeling to evaluate resistance against random guessing attacks.

Experimental results

Research questions

• RQ1 Can a flow-based generative model enable efficient and reversible secret-to-image transformation for generative steganography?
• RQ2 How does the proposed S2IRT scheme balance hiding capacity, extraction accuracy, and imperceptibility?
• RQ3 To what extent does the SE-S2IRT variant improve robustness against common image processing attacks compared to S2IRT?
• RQ4 What is the security strength of the proposed schemes against random key-guessing attacks?
• RQ5 Can the method achieve high hiding capacity (e.g., >4 bpp) while maintaining near-perfect extraction accuracy?

Key findings

• The proposed S2IRT scheme achieves a hiding capacity of up to 4 bpp with nearly 100% information extraction accuracy.
• SE-S2IRT significantly improves robustness against intensity change, contrast enhancement, and image sterilization, achieving 100% accuracy under image sterilization.
• Under salt & pepper noise (3%) and Gaussian noise (σ=10⁻³), SE-S2IRT maintains 91% and 84% accuracy respectively, outperforming S2IRT.
• The cracking probability for SE-S2IRT is reduced to 0.107×10⁻²⁸² for K=30, n=10, indicating strong resistance to random key guessing.
• Both S2IRT and SE-S2IRT maintain high imperceptibility and anti-detectability, as confirmed by visual and statistical evaluations.
• The bijective nature of the Glow model enables perfect reversibility, ensuring lossless recovery of the original secret message.