Digital Cardan Grille: A Modern Approach for Information Hiding

Jia Liu, Tanping Zhou, Zhuo Zhang, Yan Ke, Lei Yu, Minqing Zhang, Xiaoyuan Yang

Abstract

In this paper, a new framework for construction of Cardan grille for information hiding is proposed. Based on the semantic image inpainting technique, the stego image are driven by secret messages directly. A mask called Digital Cardan Grille (DCG) for determining the hidden location is introduced to hide the message. The message is written to the corrupted region that needs to be filled in the corrupted image in advance. Then the corrupted image with secret message is feeded into a Generative Adversarial Network (GAN) for semantic completion. The adversarial game not only reconstruct the corrupted image, but also generate a stego image which contains the logic rationality of image content. The experimental results verify the feasibility of the proposed method.

Introduction

Traditional steganography—the art of hiding messages in plain sight—usually works by slightly modifying the "cover" data, such as changing the least significant bits of an image's pixels. However, these modifications create statistical anomalies that modern steganalysis tools can easily detect.

The authors identify a major challenge: how do we move away from "information modifying" to "information generating"? If a message is part of the natural generation process of an image, there is no "original" to compare it against, and no modification artifacts to find. The goal is to create a stego image that possesses "logic rationality"—meaning the image content looks perfectly natural and consistent to both humans and machines.

The Digital Cardan Grille Framework

The researchers propose a framework called the Digital Cardan Grille (DCG), inspired by Girolamo Cardano’s 1550 invention of a physical template with holes used to reveal hidden text in a letter.

Overview of the Digital Cardan Grille steganography framework, showing the encoding and decoding process.

The process follows several key steps:

  1. The Mask (DCG): The sender defines a binary mask (the Grille) that determines exactly which pixel locations in a specific region will hold the secret bits.
  2. Image Corruption: A real image is "corrupted" by removing a central square (e.g., 32x32 pixels).
  3. Message Writing: The secret message is written into the corrupted region at the locations specified by the DCG.
  4. Semantic Inpainting via GAN: The corrupted image, now containing the raw message bits, is fed into a Deep Convolutional GAN (DCGAN). The GAN’s job is to fill in the missing pixels so the image looks whole again.

To ensure the result is both a high-quality image and a reliable carrier, the authors employ a Triple Loss Function:

  • Contextual Loss: Ensures the new pixels match the surrounding "un-corrupted" parts of the face.
  • Perceptual Loss: Forces the GAN to create a "realistic" looking face that can fool a discriminator.
  • Message Loss: Penalizes the GAN if it attempts to change the specific pixels where the secret message was written.

To handle the inherent "noise" of GAN generation, they introduce a Stability Index (SI), which uses the lower significant bits of a pixel as redundancy to ensure the message remains extractable even after the GAN completes the image.

Experimental Setup and Results

The authors evaluated their approach using the Labeled Faces in the Wild (LFW) dataset, which contains over 13,000 face photographs. For this study, images were aligned and resized to 64x64 pixels, with 12,000 samples used to train the DCGAN model.

Sample face images from the LFW dataset used for training and evaluating the semantic inpainting GAN.

Key Findings

  • Visual Fidelity: The system successfully reconstructed faces that look logically consistent. However, as the size of the Cardan Grille (the amount of hidden data) increases, visual distortion becomes more apparent, especially at the edges of the inpainted region.
  • Extraction Accuracy: There is a direct relationship between the "Stability Index" and the error rate. By using 5 to 7 bits of redundancy, the message could be stabilized. The best error rate achieved for message extraction was 0.42 when using a stability index of 7. While this indicates the message is being recovered, it also suggests that error-correction coding would be necessary for practical use.

Discussion and Limitations

The most profound shift in this work is that the secret message drives the image generation. The message isn't hidden in the image; the image is built around the message. This acts as an automated creator of "meaningful covers," similar to the "Tibetan poem" style of steganography.

However, a clear "tug-of-war" exists between image quality and message stability. If the message loss is prioritized too heavily, the GAN produces "heavy distortion" (artifacts) because it is forced to work around fixed pixel values that might not naturally fit a human face. High-capacity hiding currently results in noticeable visual noise.

Conclusion

The Digital Cardan Grille framework represents a significant step toward generative steganography. By leveraging GANs for semantic inpainting, the method avoids the statistical pitfalls of traditional additive steganography. Future work will focus on balancing message loss with contextual loss and extending the framework to other media such as text and video.

Summary updated

Abstract

In this paper, a new framework for construction of Cardan grille for information hiding is proposed. Based on the semantic image inpainting technique, the stego image are driven by secret messages directly. A mask called Digital Cardan Grille (DCG) for determining the hidden location is introduced to hide the message. The message is written to the corrupted region that needs to be filled in the corrupted image in advance. Then the corrupted image with secret message is feeded into a Generative Adversarial Network (GAN) for semantic completion. The adversarial game not only reconstruct the corrupted image , but also generate a stego image which contains the logic rationality of image content. The experimental results verify the feasibility of the proposed method.
BibTeX
@inproceedings{liu2018digital,
  author = {Jia Liu and Tanping Zhou and Zhuo Zhang and Yan Ke and Lei Yu and Minqing Zhang and Xiaoyuan Yang},
  title = {{Digital Cardan Grille: A Modern Approach for Information Hiding}},
  booktitle = {Proceedings of the 2018 2nd International Conference on Computer Science and Artificial Intelligence},
  year = {2018},
  doi = {10.1145/3297156.3297255},
}
Figure