Implicit Steganography Beyond the Constraints of Modality

Seoyun Yang, Sojeong Song, Chang D. Yoo, Junmo Kim

Abstract

Cross-modal steganography is committed to hiding secret information of one modality in another modality. Despite the advancement in the field of steganography by the introduction of deep learning, cross-modal steganography still remains to be a challenge to the field. The incompatibility between different modalities not only complicate the hiding process but also results in increased vulnerability to detection. To rectify these limitations, we present INRSteg, an innovative cross-modal steganography framework based on Implicit Neural Representations (INRs). We introduce a novel network allocating framework with a masked parameter update which facilitates hiding multiple data and enables cross modality across image, audio, video and 3D shape. Moreover, we eliminate the necessity of training a deep neural network and therefore substantially reduce the memory and computational cost and avoid domain adaptation issues. Detailed experiments in extreme modality settings demonstrate the flexibility, security, and robustness of INRSteg.

Introduction

Traditional steganography typically operates within a single modality—hiding an image within another image or audio within audio. Cross-modal steganography, such as embedding a 3D mesh inside a JPEG, presents significant challenges due to the structural and dimensional incompatibilities between different data types. Furthermore, existing deep learning approaches often require massive training datasets, suffer from domain shift, and remain vulnerable to sophisticated steganalysis tools that detect mathematical artifacts in raw pixel or bit domains.

Methodology: The INRSteg Framework

The core of INRSteg is the use of Implicit Neural Representations (INRs) to shift steganography from the data domain to the neural weight domain. By representing data as continuous functions (Multi-Layer Perceptrons or MLPs), the framework becomes modality-agnostic.

The overall INRSteg pipeline, including data transformation, weight allocation, and layer-wise permutation.

Data as Functions

Instead of discrete grids or samples, an INR maps coordinates (e.g., $(x, y)$ for images or $(x, y, z)$ for 3D shapes) to values (e.g., RGB or occupancy). This transformation converts any data type into a set of neural weights $\theta$.

Weight Allocation and Masked Update

To hide a secret, its weights $\theta_{secret}$ are "allocated" into a larger MLP designed for the cover data. During the hiding process, these secret weights are frozen. A masked parameter update allows the remaining "non-secret" weights to be trained to represent the cover data, effectively "wrapping" the cover information around the fixed secret.

Detailed view of the weight allocation process where secret sub-networks are embedded into a larger stego network.

Layer-wise Permutation

To ensure security, INRSteg employs a 128-bit private key to randomly permute nodes in each layer. Since MLPs are permutation-invariant, this shuffling distributes secret information throughout the network without affecting reconstruction quality, making the weights mathematically indistinguishable from standard neural weights.

Experimental Evaluation

The researchers evaluated INRSteg across diverse datasets, including ImageNet (images), VoxCeleb2 and GTZAN (audio/video), the Stanford 3D Scanning Repository (3D shapes), and scikit-video.

Security and Efficiency

  • Undetectability: INRSteg achieved a 50.00% detection accuracy against state-of-the-art steganalysis tools (SiaStegNet, XuNet), indicating perfect security where detectors are reduced to random guessing.
  • Computational Efficiency: The model is remarkably compact, with 0.40M parameters—107x smaller than DeepSteg and 2690x smaller than diffusion-based models like CRoSS.
  • High Fidelity: Reconstructed data maintained high quality, with PSNR values reaching ~62dB for images.
  • Robustness: The system is robust to weight quantization (e.g., 32-bit to 8-bit), facilitating deployment on resource-constrained devices.

Conclusion

INRSteg represents a significant leap in cross-modal steganography by leveraging the modality-agnostic nature of neural weights. By eliminating the need for large training datasets and providing perfect security through permutation, it offers a lightweight and robust solution for secure data transmission across disparate media types. Future work will focus on improving the stability of 3D reconstructions and embedding architectural metadata directly within the weights.

Summary updated

Abstract

Cross-modal steganography is committed to hiding secret information of one modality in another modality. Despite the advancement in the field of steganography by the introduction of deep learning, cross-modal steganography still remains to be a challenge to the field. The incompatibility between different modalities not only complicate the hiding process but also results in increased vulnerability to detection. To rectify these limitations, we present INRSteg, an innovative cross-modal steganography framework based on Implicit Neural Representations (INRs). We introduce a novel network allocating framework with a masked parameter update which facilitates hiding multiple data and enables cross modality across image, audio, video and 3D shape. Moreover, we eliminate the necessity of training a deep neural network and therefore substantially reduce the memory and computational cost and avoid domain adaptation issues. To the best of our knowledge, in the field of steganography, this is the first to introduce diverse modalities to both the secret and cover data. Detailed experiments in extreme modality settings demonstrate the flexibility, security, and robustness of INRSteg.
BibTeX
@inproceedings{yang2024implicit,
  author = {Seoyun Yang and Sojeong Song and Chang D. Yoo and Junmo Kim},
  title = {{Implicit Steganography Beyond the Constraints of Modality}},
  booktitle = {European Conference on Computer Vision},
  year = {2024},
  doi = {10.1007/978-3-031-73016-0\_17},
}
Figure