Highly Robust Image Watermarking Using Complementary Modulations

Chun-Shien Lu, Hong-Yuan Mark Liao, Shih-Kun Huang, Chwen-Jye SzeISW 1999 · DOI

TL;DR

This paper introduces a novel image watermarking technique called "complementary modulations" that embeds two watermarks simultaneously to ensure at least one survives various attacks. By analyzing the statistical behavior of attacks, they derive a theoretical guarantee for watermark extraction, leading to a remarkably robust and effective image protection scheme.

What Problem Does This Paper Address?

The increasing prevalence of digital media on the internet has created a critical need for robust security measures. Watermarking is a promising technique for protecting intellectual property by embedding hidden information within an image. However, existing watermarking methods often struggle with robustness, meaning they can be easily degraded or removed by common image processing operations or malicious attacks. This paper aims to address this limitation by developing a watermarking scheme that can withstand a wider range of attacks, ensuring the survival of the embedded watermark.

Methods

The core innovation of this paper is the "complementary modulations" technique. Instead of embedding a single watermark, the authors embed two complementary watermarks into the host image. This is achieved by using two distinct modulation rules: positive modulation (increasing the magnitude of transformed coefficients) and negative modulation (decreasing the magnitude). These two rules are applied simultaneously to different sets of coefficients, ensuring that if one watermark is destroyed by an attack, the other has a high probability of surviving.

The authors also introduce a statistical analysis to derive a lower bound for the likelihood of successfully extracting the "better" of the two watermarks. This theoretical guarantee helps ensure that a robust watermark is always obtained. The paper utilizes the wavelet transform for embedding and employs a human visual model to guide the process, aiming for perceptual invisibility. Additionally, a "relocation strategy" is proposed to handle attacks that cause asynchronous phenomena, by reordering the attacked wavelet coefficients to match their original positions.

Datasets

The paper uses the "Kids" image (128x128 pixels) as a host image for their experiments. No new datasets are introduced.

Key Insights and Ideas

  • Complementary Watermarks are Key: The fundamental insight is that embedding two watermarks with opposing modulation strategies significantly increases robustness. If an attack favors one type of modulation (e.g., increasing coefficients), the other watermark (which decreased coefficients) is more likely to survive.
  • Statistical Guarantees Matter: Deriving a statistical lower bound for watermark extraction provides a theoretical basis for the robustness of the scheme, moving beyond empirical observations.
  • Attack Behavior is Predictable (to an extent): The authors analyze how different attacks affect the magnitudes of transformed coefficients. This understanding allows them to design modulations that are complementary to common attack behaviors.
  • Relocation for Asynchronous Attacks: Attacks that subtly shift or distort image data (like jitter) can disrupt watermark detection. The relocation strategy, by re-aligning coefficients, effectively counters these asynchronous phenomena.
  • Bipolar Watermarks Enhance Robustness: Using bipolar watermark values (e.g., +1 and -1) is shown to be more advantageous than real-valued watermarks, especially when dealing with attacks that alter coefficient magnitudes.

Key Results

  • The proposed complementary modulation scheme demonstrates remarkable robustness against a wide variety of attacks, including blurring, median filtering, JPEG compression, StirMark, jitter, and Gaussian noise.
  • Experimental results show that even when one watermark is destroyed, the other can still be extracted with a high detector response, often exceeding 0.9.
  • The relocation strategy significantly improves detection performance, particularly for attacks that introduce asynchronous distortions.
  • Compared to existing methods like Cox et al. and Podilchuk and Zeng, the proposed scheme achieves higher detector responses under similar attack conditions, even with a weaker watermark embedding strength.
  • The scheme maintains good image quality, with a PSNR of 40 dB for the watermarked image.

Limitations and Future Work

  • Parameter Tuning: While the paper mentions using a wavelet-based visual model, the heuristic weights assignment for different frequency channels suggests that some parameter tuning might still be necessary for optimal performance.
  • Capacity: The paper focuses on robustness and doesn't explicitly detail the maximum capacity of the watermark. Future work is mentioned to focus on capacity problems.
  • Deadlock and Registration: The authors acknowledge that future work will address issues like deadlock (ownership disputes) and registration (aligning watermarks in case of transformations like rotation or scaling).
  • Complexity: While not explicitly stated as a limitation, the use of wavelet transforms and complementary modulations likely adds computational complexity to both embedding and detection.

Summary updated

BibTeX
@inproceedings{lu1999highly,
  author = {Chun-Shien Lu and Hong-Yuan Mark Liao and Shih-Kun Huang and Chwen-Jye Sze},
  title = {{Highly Robust Image Watermarking Using Complementary Modulations}},
  booktitle = {ISW},
  year = {1999},
  doi = {10.1007/3-540-47790-x\_12},
}
Figure