A Novel Hybrid Image Synthesis-Mapping Framework for Steganography Without Embedding

Rong Huang, Chunyan Lian, Zhen Dai, Zhaoying Li, Ziping MaIEEE Access 2023 · DOI

Abstract

Researchers have developed SMH-SWE, a hybrid framework that achieves 100% secret message recovery in "steganography without embedding" (SWE). By combining deep-learning image synthesis with a mapping technique to handle extraction errors, the system bypasses traditional steganalysis while avoiding the "image explosion" problem typical of previous mapping-based methods.

Introduction

Traditional steganography hides data by subtly modifying the pixels of a "cover" image. However, this creates statistical distortions that modern steganalysis tools can detect. To counter this, researchers developed Steganography Without Embedding (SWE), which either generates a new image from scratch (synthesis-based) or selects an existing image that already "matches" the secret data (mapping-based).

Both SWE approaches have historically faced significant trade-offs. Synthesis-based methods, often utilizing Generative Adversarial Networks (GANs), offer high payload capacity but suffer from "fuzzy" reconstruction, where small errors prevent 100% message recovery. Conversely, mapping-based methods provide perfect accuracy but extremely low capacity; hiding a long message might require transmitting thousands of container images, a phenomenon known as "image explosion" that can itself trigger suspicion.

The SMH-SWE Framework

This paper introduces a hybrid approach to get the best of both worlds: the high capacity of synthesis and the perfect accuracy of mapping. The proposed Synthesis-Mapping Hybrid SWE (SMH-SWE) operates in two distinct stages.

Overview of the SMH-SWE framework showing the Sender (a) and Receiver (b) processes.

Stage 1: Image Synthesis Module

The system utilizes a disentanglement auto-encoder to separate an image into "structure" and "texture" features. The "principal component" of the secret message is encoded into the structure of a synthesized image. The generator, inspired by StyleGAN2, ensures high visual fidelity, making the synthesized image appear as a natural scene or face.

Stage 2: Image Mapping Module

Because the synthesis stage is not inherently perfect, the sender extracts the message from the synthesized image and compares it to the original secret data. The resulting difference—the residual error—is then compressed. Since this error is significantly smaller than the original message, the system only requires a few additional "container" images to hide it. These images are selected from a candidate dataset by matching their block statistical hashes to the compressed error.

At the receiver's end, the principal message is extracted from the synthesized image, and the residual is recovered from the mapped images. An XOR operation then combines these components to restore the original message perfectly.

Experimental Evaluation

The framework was evaluated using four widely-used image datasets: LSUN Bedrooms, LSUN Churches, FFHQ (Flickr-Faces-HQ), and CelebA. These datasets allowed the authors to demonstrate the framework's versatility across diverse image types, from architectural scenes to human faces.

Key Results

  • 100% Extraction Accuracy: SMH-SWE achieved perfect recovery across all datasets, overcoming the reliability issues of previous synthesis-only methods.
  • Superior Image Quality: The method achieved significantly lower Fréchet Inception Distance (FID) scores than state-of-the-art (SOTA) methods. For example, on the Bedrooms dataset, SMH-SWE achieved an FID of 13.69, compared to 54.32 for the next best SOTA (CIS-Net).
  • Steganalysis Immunity: When tested against SOTA steganalysis tools like SRNet and XuNet, the detection rate was approximately 0.5, indicating the method is theoretically undetectable by current statistical analysis.

Discussion and Conclusion

The primary innovation of SMH-SWE lies in the realization that synthesis and mapping are complementary. Synthesis handles the "heavy lifting" of hiding the bulk of the data, while mapping serves as an efficient error-correction layer. By only mapping the residuals, the authors drastically reduced the number of images required—from over 22 images in pure mapping methods to approximately 4 in SMH-SWE for a 512-bit payload.

Despite these advancements, the system remains sensitive to "image attacks" such as aggressive JPEG compression, blurring, or resizing, which can destroy the high-frequency features used for data hiding. Future work will focus on enhancing robustness against such common image processing operations through adversarial learning and exploring full-image-to-image hiding techniques.

Summary updated

BibTeX
@article{huang2023novel,
  author = {Rong Huang and Chunyan Lian and Zhen Dai and Zhaoying Li and Ziping Ma},
  title = {{A Novel Hybrid Image Synthesis-Mapping Framework for Steganography Without Embedding}},
  journal = {IEEE Access},
  year = {2023},
  doi = {10.1109/access.2023.3324050},
}
Figure