Selection of the optimal embedding positions of digital audio watermarking in wavelet domain

Yangxia Hu, Maode Ma, Wenhuan Lu, Neal N. Xiong, Jianguo WeiarXiv 2020 · OpenReview

Abstract

This study investigates the optimal embedding positions for digital audio watermarking within the wavelet domain to provide a clear understanding of watermarking characteristics. By analyzing the statistical distributions and features of audio signals across various transform levels, we identify that the first four levels of the Discrete Wavelet Transform (DWT) are most suitable for watermark embedding. Unlike traditional methods that replace coefficient values, this paper proposes a self-adaptive interpolation algorithm to enhance imperceptibility. Experimental results across diverse audio styles demonstrate that this approach maintains high signal quality while offering varying degrees of robustness and fragility depending on the chosen sub-band, providing a practical framework for different watermarking applications.

Introduction

Digital audio watermarking involves embedding hidden data into an audio signal such that it remains imperceptible to the human ear while being robust enough to survive common signal processing operations. A critical challenge in this field is determining the optimal "where" of embedding. While many studies focus on complex embedding mechanisms, the selection of specific wavelet coefficients significantly impacts the final performance. This paper systematically explores the DWT domain to establish empirical rules for selecting embedding positions that balance imperceptibility and robustness.

Methodology

The core of the proposed approach is the Discrete Wavelet Transform (DWT), which decomposes an audio signal into multiple resolution levels, separating high-frequency (detail) and low-frequency (approximation) components.

Comparison between a stationary sine wave (a) and a localized wavelet (b).

Wavelets are preferred over traditional Fourier transforms because they provide both time and frequency localization, making them ideal for non-stationary audio signals. The decomposition process can be applied iteratively to focus on specific frequency bands.

The 3-level DWT decomposition process, showing the separation into low-frequency and high-frequency coefficients.

Self-Adaptive Interpolation Algorithm

To minimize audible artifacts, we introduce a self-adaptive interpolation algorithm. Traditional "replacement" methods, which substitute wavelet coefficients with watermark values, often disrupt the signal's statistical properties. Our method instead interpolates a 31-bit pseudo-random sequence into the coefficients. This approach aims to preserve the original statistical distribution, thereby achieving superior imperceptibility.

Analysis of Embedding Positions

The study evaluated several factors influencing watermark performance:

  1. Wavelet Base Selection: Analysis showed that the choice of wavelet base (e.g., Haar vs. Daubechies) has a negligible effect on the statistical distribution of coefficients, suggesting that simpler bases like Haar are often sufficient.
  2. Transform Levels: The first four levels of DWT decomposition were found to be the most effective. Beyond level five, the coefficients lose their resemblance to the original signal's features, increasing the risk of audible distortion.
  3. Frequency Sub-bands:
    • Low-Frequency (LF) Coefficients: These carry the bulk of the signal's energy. Embedding here results in high robustness, allowing the watermark to survive lossy compression like MP3.
  4. High-Frequency (HF) Coefficients: These represent fine details. Embedding here creates fragile watermarks, which are highly sensitive to any signal tampering, making them ideal for integrity authentication.

Experimental Results

The algorithm was tested on six audio genres: Speech, Blues, Classic, Rock, Jazz, and Pop. Key findings include:

  • Imperceptibility: The interpolation method consistently outperformed traditional replacement algorithms in terms of energy difference metrics.
  • Robustness: Watermarks in the first two DWT levels showed strong resistance to MP3 compression at 16kbps, with Bit Error Rates (BER) typically below 0.3.
  • Optimal Regions: High-energy regions of the audio signal generally offered better imperceptibility for embedded data compared to low-energy regions.

Conclusion

This research provides a foundational guide for selecting watermark embedding positions. We conclude that for most applications, the second level of DWT transformation offers an ideal "sweet spot" for balancing transparency and data survival. Future work will explore hybrid transform domains and resistance to more sophisticated malicious attacks.

Summary updated

Abstract

This work studied embedding positions of digital audio watermarking in wavelet domain, to make beginners understand the nature of watermarking in a short time. Based on the theory of wavelet transform, this paper analyzed statistical distributions of each level after transformation and the features of watermark embedded in different transform levels. Through comparison and analysis, we found that watermark was suitable for embedding into the coefficients of the first four levels of wavelet transform. In current state-of-art approaches, the embedding algorithms were always to replace the coefficient values of the embedded positions. In contrast this paper proposed an embedding algorithm of selfadaptive interpolation to achieve a better imperceptibility. In order to reduce the computational complexity, we took a pseudo random sequence with a length of 31 bits as the watermark. In the experiments, watermark was embedded in different locations, including different transform levels, high-frequency coefficients and low-frequency coefficients, high-energy regions and low-frequency regions. Results showed that the imperceptibility was better than traditional embedding algorithms. The bit error rates of the extracted watermark were calculated and we analyzed the robustness and fragility of each embedded signal. At last we concluded the best embedding positions of watermark for different applications and our future work.
BibTeX
@misc{hu2020selection,
  author = {Yangxia Hu and Maode Ma and Wenhuan Lu and Neal N. Xiong and Jianguo Wei},
  title = {{Selection of the optimal embedding positions of digital audio watermarking in wavelet domain}},
  year = {2020},
  eprint = {2010.11461},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2010.11461},
}
Figure