A Compact Neural Network-based Algorithm for Robust Image Watermarking

Hongbo Xü, Rong Wang, Jia Wei, Lu Shao-PingarXiv 2021 · DOI · arXiv

TL;DR

Researchers have developed the Invertible Watermarking Network (IWN), a compact architecture that treats image watermark embedding and extraction as a single, bijective inverse problem. By using an Invertible Neural Network (INN) and a novel bit normalization module, the system achieves significantly higher robustness against attacks like JPEG compression and cropping while maintaining superior image quality compared to traditional separate encoder-decoder pipelines.

What Problem Does This Paper Address?

Digital image watermarking faces a "conflict triangle" between three key factors: robustness (surviving attacks), imperceptibility (staying invisible to humans), and capacity (how much data can be hidden).

Traditional methods often rely on hand-crafted rules in the spatial or frequency domains, which struggle to generalize to new types of image distortion. Recent deep learning approaches typically use an "Encoder-Noiser-Decoder" framework. However, because the encoder and decoder are separate pipelines, they often suffer from information loss during the forward pass, making it harder for the decoder to perfectly reconstruct the message. Furthermore, training two independent networks requires complex hyperparameter tuning to balance the conflicting goals of invisibility and data recovery.

Methods

The core innovation is the Invertible Watermarking Network (IWN). Instead of two separate networks, the authors use a single Invertible Neural Network (INN).

  1. Bijective Architecture: In an INN, the network is mathematically designed to be bijective. The forward process takes a cover image and a message to produce a watermarked image. The reverse process takes that same watermarked image (potentially with noise) and runs the math backward to extract the message and the original cover image. This ensures that the embedding and extraction share the exact same parameters.
  2. Bit Message Normalization: To improve robustness, the authors don't just feed raw bits into the network. They introduce a module that groups bits, converts them to decimal values, and "left-shifts" them into higher-order bit positions. This makes the message more resilient to small pixel perturbations (noise) that usually destroy low-order bit information.
  3. Differentiable Noise Layer: To train the network to survive real-world distortions, the authors include a "Noise Layer" between the forward and reverse passes. This layer simulates attacks like Gaussian blur, cropping, and JPEG compression. Since JPEG is normally non-differentiable (making it hard for neural networks to learn from), they use a differentiable approximation to allow end-to-end training.

Datasets

The model was trained and evaluated using standard high-resolution image datasets:

  • Training: DIV2K and Flickr2K, which provide high-quality, diverse imagery.
  • Testing: A subset of ImageNet (1,000 images) to verify how well the model generalizes to unseen data.
  • Benchmarks: The results were compared primarily against HiDDeN, a well-known baseline in deep-learning-based watermarking.

Key Insights and Ideas

  • Embedding as Inversion: The most powerful idea here is treating watermarking as a pair of inverse problems. If you can learn a stable, invertible mapping between "Image + Message" and "Watermarked Image," the extraction process becomes a natural mathematical consequence of the embedding process.
  • Condensing Information: Unlike previous models that might use one image channel to represent a single bit, the IWN's normalization module allows a single channel to represent multiple bits by treating them as a normalized tensor. This significantly increases capacity without sacrificing robustness.
  • Parameter Efficiency: Because the encoder and decoder are essentially the same set of weights run in different directions, the model is more "compact" and easier to train than dual-network systems.

Key Results

The IWN outperformed the baseline (HiDDeN) across almost all metrics:

  • Imperceptibility: Under JPEG compression (Q=50), IWN achieved a PSNR of 36.16 dB, a +3.52 dB gain over the baseline, meaning the watermark is much harder for the human eye to detect.
  • Robustness: In terms of bit accuracy (the percentage of correctly recovered bits), IWN showed a massive improvement under JPEG attacks, reaching 18.4% higher accuracy than the baseline.
  • Ablation Success: The researchers found that increasing the number of invertible blocks (up to 16) consistently improved both image quality and recovery accuracy, proving the scalability of the INN approach.

Limitations and Future Work

Despite the strong performance, the authors noted a specific limitation: visual artifacts in smooth regions. In images with large areas of flat color (like a clear blue sky or a plain wall), the watermark can sometimes manifest as subtle textures or "noise" that is perceptible to a sharp-eyed observer.

Future Work includes:

  • Developing better loss functions (like SSIM or Fourier transform loss) to specifically target and eliminate artifacts in smooth image areas.
  • Expanding the framework to "cross-media" channels, such as watermarks that can survive being printed on paper and then photographed by a smartphone.

Summary updated

Abstract

Digital image watermarking seeks to protect the digital media information from unauthorized access, where the message is embedded into the digital image and extracted from it, even some noises or distortions are applied under various data processing including lossy image compression and interactive content editing. Traditional image watermarking solutions easily suffer from robustness when specified with some prior constraints, while recent deep learning-based watermarking methods could not tackle the information loss problem well under various separate pipelines of feature encoder and decoder. In this paper, we propose a novel digital image watermarking solution with a compact neural network, named Invertible Watermarking Network (IWN). Our IWN architecture is based on a single Invertible Neural Network (INN), this bijective propagation framework enables us to effectively solve the challenge of message embedding and extraction simultaneously, by taking them as a pair of inverse problems for each other and learning a stable invertible mapping. In order to enhance the robustness of our watermarking solution, we specifically introduce a simple but effective bit message normalization module to condense the bit message to be embedded, and a noise layer is designed to simulate various practical attacks under our IWN framework. Extensive experiments demonstrate the superiority of our solution under various distortions.
BibTeX
@misc{xu2021compact,
  author = {Hongbo Xü and Rong Wang and Jia Wei and Lu Shao-Ping},
  title = {{A Compact Neural Network-based Algorithm for Robust Image Watermarking}},
  year = {2021},
  eprint = {2112.13491},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2112.13491},
}
Figure