💻 computer science

High-Capacity Robust Watermarking Technology for High-Resolution Images

This paper proposes a high-capacity, robust watermarking method for high-resolution images that utilizes a block-wise strategy and a reversible symmetric encoder-decoder architecture to embed 4 KB of data into 1024×1024 images while maintaining strong visual imperceptibility and resistance to various noise attacks.

Original authors: Shaowu Wu, Wei Lu, Qing Qian, Mingsen Deng

Published 2026-07-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shaowu Wu, Wei Lu, Qing Qian, Mingsen Deng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are walking through a massive digital library where every book, painting, and song is being created by a super-smart robot artist. These robots are amazing; they can churn out masterpieces in seconds, helping humans work faster and cheaper. But there's a catch: because the robots make so many things so easily, it's getting hard to tell who actually owns a piece of art or where it came from. It's like if a million people started copying your favorite song and selling it as their own. To solve this, scientists have been using "digital watermarks." Think of these not as wet ink, but as invisible, tiny secret messages hidden inside the pixels of an image. Just like a real watermark on a banknote, these digital codes prove ownership without ruining the picture.

For a long time, these secret messages were very small—like a tiny note saying "Copyright 2024" or a short ID number. They worked great on small, low-resolution pictures, like the size of a postage stamp. But today, we are dealing with huge, high-definition images that look like they could be printed on a billboard. Trying to hide a massive amount of information (like a whole paragraph of text or a long ID code) inside a giant, high-quality picture is incredibly difficult. If you try to stuff too much into a small space, the picture gets blurry, or the secret message gets lost when the image is tweaked. This paper tackles the challenge of hiding a lot of secret information inside these giant, high-resolution images without ruining the picture or losing the message.

The researchers, a team from Sun Yat-sen University and Guizhou University of Finance and Economics, have built a new system to solve this puzzle. Their main idea is to stop trying to hide the whole secret message in one giant lump. Instead, they use a "block-wise" strategy. Imagine you have a giant, high-resolution photo of a city. Instead of trying to hide a secret message across the whole city at once, they chop the city into thousands of small, manageable neighborhoods (blocks). They then hide a tiny piece of the secret message in each neighborhood.

This approach is like a team of spies working in a large building. Instead of one spy trying to hide a massive suitcase in the main lobby (which would be obvious and hard to move), they have many spies, each hiding a small, lightweight package in a different room. This makes the job much easier, even if the spies don't have a lot of resources (like powerful computers) to work with. By breaking the problem down, their system can hide a massive amount of data—up to 4 KB, which is 32,768 bits—inside a single 1024×1024 pixel image. That's a huge jump from the usual 30 or 100 bits that older systems could handle.

The system works like a two-step magic trick. First, an "Encoder" takes the original image and the secret message, chops them up, and weaves the message into the tiny blocks of the image. It's designed to be reversible and symmetrical, meaning the process is perfectly balanced. Then, even if the image gets attacked—like being squashed by JPEG compression, covered in static noise, or having parts of it cropped out—a "Decoder" can still find the pieces. The Decoder reverses the process, gathering the tiny message fragments from each block and stitching them back together to reveal the full secret.

To make sure the magic trick works perfectly, the team trained their computer brain using a special set of rules called a "loss function." Think of this as a strict teacher grading the system on two things: how invisible the secret message is (visual quality) and how well the message survives attacks (robustness). They didn't just grade the whole image; they graded every single neighborhood (block) individually. This ensures that no part of the image gets ruined, and the message stays safe everywhere.

The results are impressive. In their tests, the system managed to hide that massive 4 KB message into high-resolution images while keeping the picture looking almost identical to the original. The quality scores (PSNR) stayed above 36 dB, and the structural similarity (SSIM) was over 0.92, which means the human eye can barely tell the difference. Even when they simulated tough attacks like heavy JPEG compression or adding random noise, the system recovered the secret message with over 98% accuracy in most cases.

The paper suggests that while older methods were great for small, low-resolution images, they struggled when asked to handle the huge data needs of modern high-resolution images. By breaking the image into blocks and treating each one as a small, independent puzzle, this new method proves that you can have both high capacity and high quality. It's a significant step forward, showing that we can protect the copyright of our giant, high-definition digital creations without sacrificing their beauty or losing the secret message inside.

Technical Summary: High-Capacity Robust Watermarking Technology for High-Resolution Images

Problem Statement
Existing digital watermarking techniques are predominantly designed for low-resolution images (e.g., 128×128 to 512×512) and are limited to fixed, low-capacity watermarks (typically 30 to 256 bits). As the demand for AIGC traceability grows—requiring the embedding of substantial metadata such as C2PA manifests—current methods struggle to handle high-resolution carriers (e.g., 1024×1024) and high-capacity payloads (up to 4KB). Furthermore, training deep learning models for high-resolution, high-capacity scenarios often leads to a surge in network parameters, making optimization difficult under resource-constrained conditions.

Methodology
The proposed method addresses these challenges through a block-wise strategy and a symmetric reversible network architecture:

Block-wise Partitioning: To manage high-resolution images and large payloads, the method partitions the image into local blocks and segments the watermark into corresponding independent parts. A high-resolution image is divided into $k$ blocks, and the high-capacity watermark is split into $k$ segments. This transforms the problem of embedding a large watermark into a high-resolution image into multiple sub-problems of embedding low-capacity watermarks into low-resolution blocks. This approach significantly reduces the number of learnable parameters, enabling training even with limited GPU resources.
Encoder and Decoder Architecture:
- Encoder: Utilizes a reversible symmetric architecture with three convolutional layers. It processes each image block and its corresponding watermark segment. The watermark tensor is coupled with the image features at each convolutional layer via concatenation and upsampling.
- Decoder: Employs a symmetric structure with three deconvolutional layers to extract the watermark from the noisy image. It decouples watermark features from each convolutional layer and fuses them to reconstruct the final watermark.
- Noise Layer: A cascaded noise layer is inserted between the encoder and decoder to simulate various attacks sequentially (e.g., JPEG compression followed by Gaussian noise), effectively emulating combined noise attacks during training.
Loss Function: The optimization objective combines four components with weighted contributions:
1. Global Image Loss ( $L_{global}$ ): Mean Squared Error (MSE) between the original and watermarked images to ensure overall visual quality.
2. Local Image Loss ( $L_{local}$ ): Sum of MSEs for individual blocks to constrain visual quality in local regions.
3. Global Watermark Loss ( $L_{W}$ ): Binary Cross-Entropy (BCE) between the original and extracted watermarks.
4. Local Watermark Loss ( $L_{merge}$ ): Sum of BCEs for individual watermark segments.
  This dual-level (global and local) loss strategy ensures a balance between imperceptibility and robustness.

Key Contributions

High-Capacity Method for High-Resolution Images: The paper proposes a method capable of embedding a 4KB (32,768 bits) watermark into 1024×1024 images, achieving an embedding rate of 0.0313 bpp. This is achieved by decomposing the high-capacity task into local low-capacity tasks.
Resource-Efficient Training: By utilizing a block-wise strategy, the method reduces network complexity, allowing for normal training and learning even under low-resource conditions where full high-resolution training would be infeasible.
Enhanced Robustness and Imperceptibility: The adoption of a weighted loss function combining global and local constraints improves the overall visual quality and robustness of the image against various noise attacks.
Arbitrary Length Support: The scheme supports the embedding and extraction of watermarks of arbitrary lengths, provided they do not exceed the 4KB limit.

Experimental Results
Extensive experiments were conducted on three datasets (Mirflickr, COCO, and BOSSBase) with images rescaled to 1024×1024.

Visual Quality: At an embedding strength ( $\alpha$ ) of 1.0, the method achieved a PSNR exceeding 36 dB and an SSIM above 0.92 across all datasets. Reducing the embedding strength to 0.6 improved PSNR to over 40 dB and SSIM to over 0.96, demonstrating the ability to trade capacity for quality.
Robustness: The method demonstrated strong robustness against various attacks:
- JPEG Compression: Accuracy exceeded 98.51% even at a quality factor (QF) of 50.
- Gaussian Noise: Accuracy exceeded 99.66% with variance 0.02.
- Other Attacks: The method achieved near-perfect accuracy (>99.99%) against Identity, Dropout, Cropout, and Gaussian filtering attacks. Accuracy remained above 96.75% against Crop and Scaling attacks.
Comparative Performance: Compared to existing methods (e.g., HiDDeN, StegaStamp, MBRS), the proposed method offers the highest embedding rate (0.0313 bpp) while maintaining competitive visual quality and robustness, particularly in high-resolution scenarios where other methods are not directly applicable or require significant adaptation.

Significance and Claims
The paper claims to provide an effective and feasible solution for transitioning digital watermarking from low-resolution, low-capacity scenarios to high-resolution, high-capacity scenarios. It specifically addresses the challenges of embedding large metadata (such as C2PA manifests) into high-resolution images generated by AIGC systems. The authors assert that their block-wise strategy and symmetric reversible network design successfully overcome the limitations of parameter surges and fixed-length constraints found in prior art, ensuring that the watermarked images maintain high visual quality and strong robustness against noise attacks. The work validates that high-capacity watermarking is achievable in high-resolution contexts without sacrificing the network's ability to train under standard resource constraints.

Technical Summary: High-Capacity Robust Watermarking Technology for High-Resolution Images

More like this