Diffusion-Based Hierarchical Image Steganography

TL;DR

HIS addresses robust, high-capacity multi-image steganography by differentiating content importance and embedding $\,\mathbf{x}_{s-\text{Tier1}}$ with higher protection than $\,\mathbf{x}_{s-\text{Tier2}}$ within a single container. It fuses diffusion-based container generation (via DDIM inversion) with reversible flow-based modules (Embed-Flow and Enhance-Flow) to achieve tiered robustness and efficient recovery. The approach relies on private/public keys to control hiding and revealing, and employs a two-stage enhancement to boost reconstruction quality while maintaining security against common distortions and steganalysis. Empirical results show improved robustness under noise and compression, strong anti-analysis security, and scalable capacity for concealing multiple images and text with practical privacy benefits.

Abstract

This paper introduces Hierarchical Image Steganography, a novel method that enhances the security and capacity of embedding multiple images into a single container using diffusion models. HIS assigns varying levels of robustness to images based on their importance, ensuring enhanced protection against manipulation. It adaptively exploits the robustness of the Diffusion Model alongside the reversibility of the Flow Model. The integration of Embed-Flow and Enhance-Flow improves embedding efficiency and image recovery quality, respectively, setting HIS apart from conventional multi-image steganography techniques. This innovative structure can autonomously generate a container image, thereby securely and efficiently concealing multiple images and text. Rigorous subjective and objective evaluations underscore our advantage in analytical resistance, robustness, and capacity, illustrating its expansive applicability in content safeguarding and privacy fortification.

Figures (10)

• Figure 1: Framework of our diffusion-based Hierarchical Image Steganography (HIS), which takes $\mathbf{x}_{s-Tier1}$ and multiple $\mathbf{x}_{s-Tier2}$ as input to produced $\mathbf{y}_{c-Add}$ as container. Inversely, $\mathbf{x}^{\prime}_{s-Tier1}$ is recovered via Enhance-Flow and $\mathbf{x}^{\prime}_{s-Tier2}$ is recovered via Embed-Flow from container.
• Figure 2: The workflow of Diffusion-based Inversion: $\mathbf{x}_{s-Tier1}$ is firstly transformed to intermedia noise $\mathbf{y}_{n-Diff}$ with private key $\mathbf{k}_{pri}$, and then transformed to container $\mathbf{y}_{c-Diff}$ with public key $\mathbf{k}_{pub}$ using the DDIM. Reversely, public key $\mathbf{k}_{pub}$ is initially ultilized to transfer $\mathbf{y}^{\prime}_{c-Diff}$ to $\mathbf{y}^{\prime}_{n-Diff}$. Finally, $\mathbf{x}^{}_{Diff}$ is reconstructed under the condition of $\mathbf{k}_{pri}$.
• Figure 3: Framework of Enhance-flow: $\mathbf{x}_{s-Tier1}$ is firstly transfered to Gaussian-like $\mathbf{y}_{n-Flow}$ via Enh-Flow-S1 with the feature of $\mathbf{x}_{s-Diff}$, followed by the conversion to $\mathbf{y}_{n-Diff}$ via Enh-Flow-S2. The reversion from $\mathbf{y}^{\prime}_{n-Diff}$ to $\mathbf{x}^{\prime}_{s-Tier1}$ is symmetrical to the forward process.
• Figure 4: Framework of Embed-Flow: the Invertible Blocks take the $\mathbf{y}_{c-Diff}$ and $\mathbf{x}_{s-Tier2}$ as input to produce container $\mathbf{y}_{c-Add}$ and redundant Gaussian Noise. In the backward pass, Gaussian sampling datapoint and distorted $\mathbf{y}^{\prime}_{c-Add}$ will produce recovered secret $\mathbf{x}^{\prime}_{s-Tier2}$.
• Figure 5: Visual results of HIS and other methods xu2022robustguan2022deepmih under the same degradations of JPEG QF-80. Our method can vividly reconstruct the secret images, while other methods exhibit significant color distortion or have completely failed.