Implicit Steganography Beyond the Constraints of Modality

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

TL;DR

INRSteg is a novel steganography framework that leverages Implicit Neural Representations (INRs) to hide secret data of any modality (image, audio, video, or 3D shape) inside a cover data container of any other modality. By operating entirely within the weight space of coordinate-based neural networks rather than discrete pixels or waveforms, it achieves lossless secret data recovery and perfect invisibility against modern steganalysis tools—all without requiring heavy deep learning model training.

What Problem Does This Paper Address?

Traditional steganography typically operates within a single modality, such as hiding a secret image inside another cover image. Cross-modal steganography (e.g., hiding an audio clip inside an image) introduces severe complications due to structural incompatibilities, dimensional mismatches, and temporal constraints.

While recent deep learning methods (like GANs, Invertible Neural Networks, and diffusion models) have improved embedding capacity, they suffer from significant drawbacks:

  1. High Computational Cost: They require training massive networks on large datasets.
  2. Capacity-Security Trade-offs: Modifying discrete data arrays (like pixels) leaves statistical footprints that traditional and deep-learning-based steganalysis tools can easily detect.
  3. Modality Rigidity: Existing frameworks are usually hardcoded for specific cover-secret pairs and cannot flexibly adapt to diverse data types like 3D shapes or video.

Methods

INRSteg bypasses these limitations by shifting data representation from discrete arrays to continuous functions via Implicit Neural Representations (INRs), specifically using SIREN (Sinusoidal Representation Networks).

The framework consists of three main stages:

  1. Data Representation Transformation: The secret data types are first parameterized into individual Multi-Layer Perceptrons (MLPs) with sine activation functions, yielding the weights $\theta_{secret}$.
  2. Hide and Reveal (Weight Allocation): The weights of the secret INRs are mapped into the diagonal blocks of a larger, newly initialized "stego" MLP network ($\theta_{stego}$). Using a binary mask, these secret weights are frozen. The remaining un-frozen, non-diagonal parameters of the stego network are then optimized ("fitted") to represent the cover data.
  3. Layer-wise Permutation Encoding: Before sharing, the nodes of each layer in the stego network are randomly shuffled using a secret key derived from a 128-bit private key. Because MLPs are inherently permutation-invariant, this scrambling does not alter the network's output (the cover data remains perfectly reconstructible), but it completely obfuscates the structural boundaries of the hidden weights.
Fig. 1: A general framework of INRSteg for hiding two types of secret data. All four data types can be transformed into INRs, and the red box shows which of the four data types is selected in this case. After the representation transformation of each secret data, the weights of the two INRs are allocated in a new MLP network. With the weight freezing technique, the weights from the secret data are fixed while the rest are fitted on the cover data. The existence of the secret data is concealed by permutation encoding via private key, which does not affect the reconstruction performance. The private key is then used to decode the permuted stego network, and the secret data are revealed by separating each MLP network.
An example of allocating two distinct secret networks into a single unified stego network structure.

Datasets

The authors evaluated INRSteg across a diverse suite of standard datasets to prove its cross-modal capabilities:

  • Images: ImageNet and mini-ImageNet.
  • Audio/Video: VoxCeleb2 (for simultaneous audio-video embedding) and GTZAN (music genre dataset).
  • Video: scikit-video dataset.
  • 3D Shapes: Stanford 3D Scanning Repository (specifically the Asian Dragon and Armadillo models).

Key Insights and Ideas

  • The Weight Space as a Universal Modality: By converting images, audio, video, and 3D shapes into network weights, the modality barrier completely vanishes. Every data type becomes just a collection of continuous MLP parameters.
  • Lossless Recovery via Freezing: Because the secret weights are explicitly masked and frozen during the cover-fitting process, the retrieval mechanism is mathematically lossless.
  • Exploiting Permutation Invariance for Security: Without permutation, the weight matrices of the stego network display visible "borderlines" or block structures under visualization. Shuffling the hidden nodes leverages a fundamental property of neural networks to distribute the secret parameters uniformly, rendering them visually and statistically undetectable.

Key Results

  • Perfect Steganalysis Immunity: When tested against state-of-the-art image steganalysis tools like SiaStegNet and XuNet, INRSteg achieved exactly 50.00% detection accuracy. This means the detectors performed no better than a random guess, rendering the hidden data completely undetectable.
  • Extreme Compactness: INRSteg requires only 0.40M parameters. Compared to prior art like DeepSteg (42.58M) or the diffusion-based CRoSS (1,066.24M), INRSteg is between 100$\times$ to 2,600$\times$ smaller, making it highly suitable for edge devices like smartphones.
  • High Fidelity: In intra-modal image tasks, it achieved a cover PSNR of 62.34 dB (outperforming DeepMIH's 36.05 dB). In cross-modal tasks (e.g., hiding VoxCeleb2 audio and video inside an ImageNet cover), it maintained excellent reconstruction metrics (Audio SNR: 37.55 dB, Video PSNR: 41.44 dB).
  • Robustness: The hidden data successfully survived model compression techniques, showing minimal perceptual degradation even after undergoing int8 quantization and iterative magnitude pruning up to 10%.

Limitations and Future Work

  • SIREN Instability: The underlying SIREN architecture can occasionally suffer from optimization instability. For instance, when using 3D shapes as the cover modality, the stego network sometimes carries minor surface defects and fails to recover perfectly.
  • Metadata Dependency: While INRSteg can easily accommodate diverse secret formats, it relies on metadata (network dimensions, layer counts) to properly extract the sub-networks. The authors suggest this metadata could also be hidden within specific weight regions in future iterations.

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{song2024implicit,
  author = {Sojeong Song and Seoyun Yang 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