Scaling Payload Capacity in Neural Information Hiding

This direction is merged into another direction: steganographic trilemma. Its last version stays readable and citable here.

Version 1 of 195 papers112 conceptsfrozen 2026-10-04

The fundamental challenge of steganography is a three-way tug-of-war between capacity (how much data can we hide?), fidelity (how much does the carrier change?), and security (can an adversary detect the message?). Historically, this was a zero-sum game: increasing the payload meant inevitably introducing "salt-and-pepper" noise or statistical anomalies that a keen observer or a steganalysis tool could easily spot.

Recent research has shifted this frontier by moving away from simple pixel-level tweaks toward embedding information within the high-dimensional latent spaces of neural networks. By leveraging generative models and implicit representations, researchers are finding ways to pack massive amounts of data—sometimes multiple full-resolution images or minutes of audio—into a single carrier without leaving a visible trace.

The Capacity Fidelity Tradeoff

In classical steganography, the "slack" used to hide data is found in the least significant bits of an image. As Guangjie Liu, Yuewei Dai, and Zhiquan Wang noted in Adaptive Image Steganography by Reversible Data Hiding, the goal is to hide data in "busy" or high-texture areas where the human eye is less sensitive to change. However, even with clever grouping of pixels—as explored by Zhihui Wang and Chin-Chen Chang in their work on Exploiting Modification Direction—there is a hard ceiling. Once you exceed roughly 2 bits per pixel (bpp), the structural integrity of the image begins to collapse.

Neural networks break this ceiling by treating the carrier not as a grid of independent pixels, but as a sample from a complex distribution. Japsimar Singh Wahi and colleagues demonstrated in Multi-Image Steganography Using Deep Neural Networks that a "prep" network can condense multiple secret images into a single cover. This reached a new extreme with InvMIHNet, where Zihan Chen and Tianrui Liu used invertible neural networks to hide up to 16 secret color images within one, treating the secrets as a "mosaic" that the network can unfold.

From Modification to Synthesis

A more radical approach to maximizing capacity is to stop modifying existing covers and start generating them from scratch. If the secret message is used to guide the generation process, there is no "original" image to compare against, making traditional steganalysis ineffective.

Jia Liu and Yan Ke have explored this through Generative Adversarial Networks (GANs), where the network learns to match the distribution of natural images while simultaneously encoding bits. In Digital Cardan Grille, they showed that semantic inpainting can weave messages into the very fabric of a generated scene.

More recently, diffusion models have provided even higher fidelity. Daegyu Kim and Jooyoung Choi introduced Diffusion-Stego, which projects messages into the denoising process. This allows for high-capacity hiding without the "training-free" limitations of earlier models, ensuring that the resulting image remains a perfectly plausible sample of the target distribution.

Hiding in the Weights

The most recent shift in the field moves the hiding place from the output of a model to the parameters of the model itself. This is known as steganography via Implicit Neural Representations (INRs).

Instead of a pixel grid, an image can be represented as a function $f(x, y) = \text{color}$. By training a network to represent a cover image, researchers like Sojeong Song and Junmo Kim in INRSteg found they could hide entire videos or audio clips within the continuous weight space of the network.

This "format-agnostic" approach was further refined by Qi Song and Sheng Li in U-INR. By using a private key to determine which neurons store "cover" information and which store "secret" information, they achieved a unified framework that works across images, 3D scenes (like Gaussian Splatting), and audio. The intuition here is that neural networks are naturally over-parameterized; steganography simply harvests this "excess capacity" to store hidden data.

Why It Works: The High-Dimensional Slack

The reason neural methods can achieve such high payloads (often exceeding 4 bpp, as seen in LISO by Xiangyu Chen) is that they exploit the discrepancy between what the human visual system perceives and what a high-capacity model can distinguish.

While a human might see a standard portrait, a "reveal" network can be trained to extract high-frequency patterns that are invisible to us but carry dense information. As Chaoning Zhang and Philipp Benz argued in Universal Deep Hiding, deep steganography essentially functions by hiding secret signals in frequency bands that the cover image's primary features do not occupy, effectively "disentangling" the secret from the cover.

Papers in this version

Showing 60 of 95 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_scaling_payload_capacity_in_neural_infor_c9c4ed27,
  title = {Scaling Payload Capacity in Neural Information Hiding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/scaling-payload-capacity-in-neural-information-hiding/rdr_c9c4ed2753c04bb7ad01e6fbbf7e3497},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_scaling_payload_capacity_in_neural_infor_c9c4ed27_v1,
  title = {Scaling Payload Capacity in Neural Information Hiding},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/scaling-payload-capacity-in-neural-information-hiding/rdr_c9c4ed2753c04bb7ad01e6fbbf7e3497/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}