Neural Cover Selection for Image Steganography

TL;DR

This work introduces a novel cover selection framework, which involves optimizing within the latent space of pretrained generative models to identify the most suitable cover images, distinguishing itself from traditional exhaustive search methods.

Abstract

In steganography, selecting an optimal cover image, referred to as cover selection, is pivotal for effective message concealment. Traditional methods have typically employed exhaustive searches to identify images that conform to specific perceptual or complexity metrics. However, the relationship between these metrics and the actual message hiding efficacy of an image is unclear, often yielding less-than-ideal steganographic outcomes. Inspired by recent advancements in generative models, we introduce a novel cover selection framework, which involves optimizing within the latent space of pretrained generative models to identify the most suitable cover images, distinguishing itself from traditional exhaustive search methods. Our method shows significant advantages in message recovery and image quality. We also conduct an information-theoretic analysis of the generated cover images, revealing that message hiding predominantly occurs in low-variance pixels, reflecting the waterfilling algorithm's principles in parallel Gaussian channels. Our code can be found at: https://github.com/karlchahine/Neural-Cover-Selection-for-Image-Steganography.

Figures (14)

• Figure 1: Left: Image steganography framework: the encoder takes as input the cover image $\mathbf{x}$ and a secret binary message $\mathbf{m}$ and outputs the steganographic image $\mathbf{s}$. The decoder then estimates $\mathbf{\hat{m}}$ from s. Right: Randomly sampled cover images from the ImageNet dataset before and after optimization using our framework (described in Section \ref{['sec:method']}). These optimized images demonstrate a significantly reduced error $||\mathbf{m} - \hat{\mathbf{m}}||$ while maintaining high image quality.
• Figure 2: DDIM-based cover selection framework overview. The input cover image $\textbf{x}_0$ is first converted to the latent space $\textbf{x}_T$ via forward diffusion. Then, guided the message recovery loss, the latent space is fine-tuned, and the updated cover image is generated via the reverse diffusion process. The DDIM model as well as the steganographic encoder-decoder pair are pretrained.
• Figure 3: Normalized pixel variances (top) and residuals (bottom) calculated across a batch of 500 Robin images for each color channel, before optimization.
• Figure 4: Power coefficients $\gamma_i$ for each color channel, calculated using a batch of 500 Robin images.
• Figure 5: Normalized pixel variances across a batch of 500 Robin images for each color channel, after optimization.