Power Spectrum Compliant QIM Watermarking for Autoregressive Host Signals

Bin Yan, Ya-Fei Wang, Lingyun Song, Hong-Mei Yang

Introduction

Digital watermarking faces a significant challenge from sophisticated attackers who use optimal filters, such as Wiener filters, to estimate and remove watermarks. To counter this, a watermark must be "Power Spectrum Density (PSD) compliant," meaning its spectral characteristics must mirror those of the host signal. While Spread Spectrum (SS) watermarking has achieved this, Quantization Index Modulation (QIM)—valued for its high capacity and blind detection—has lacked a robust PSD-compliant framework, especially for signals modeled as autoregressive (AR) processes, such as speech or EEG data. This paper introduces the Autoregressive Spectrum Shaped Dither Modulation (AR-SSDM) scheme to bridge this gap.

Proposed Methodology: AR-SSDM

The core innovation of the AR-SSDM framework lies in embedding the watermark into the AR residual signal rather than the host signal directly. This approach leverages the properties of AR models to ensure spectral compliance naturally.

Figure 1: The AR-SSDM watermarking encoder pipeline.

The process follows a four-step pipeline:

  1. Inverse Filtering: The host signal $\{X_n\}$ is passed through an inverse AR filter to extract the residual signal $\{U_n\}$, which resembles white Gaussian noise.
  2. Embedding: Watermark bits $m$ are embedded into the residual using Dither Modulation (DM) or Spread-Transform Dither Modulation (STDM), producing the watermarked residual $\{\hat{U}_n\}$.
  3. Synthesis (Filtering): The watermarked residual is passed through an AR synthesis filter. This step is crucial; it shapes the watermark's spectrum to match the host signal's PSD perfectly.
  4. Decoding: At the receiver, inverse filtering is applied to the received signal to recover a noisy version of the residual, followed by minimum distance decoding to extract the message.

Key Insights and Theoretical Performance

The "residual embedding" strategy simplifies spectral shaping. By embedding a white-noise-like watermark into a white residual, the synthesis filter automatically imposes the host's spectral envelope on the watermark. To enhance robustness against channel noise, the authors utilize STDM, which spreads bits across multiple residual samples, effectively increasing the quantization step size without violating PSD compliance.

Furthermore, the authors derived an analytical formula for the bit error rate ($P_e$), allowing for predictable performance based on the signal-to-noise ratio and AR parameters.

Experimental Results

The proposed method was validated using Monte Carlo simulations on synthetic AR signals modeled after real-world speech data.

Figure 6: Comparison of the Power Spectrum Density (PSD) of the host signal and the embedded watermark, demonstrating perfect spectral matching.

Key findings include:

  • Spectral Compliance: As shown in Figure 6, the watermark's PSD perfectly tracks the host signal's PSD, confirming its resistance to optimal filtering attacks.
  • Predictability: Empirical bit error rates closely matched theoretical predictions across various Watermark-to-Noise Ratios (WNR).
  • Robustness: Even when AR parameters were unknown and had to be estimated from noisy signals, the AR-SSDM method outperformed traditional SS watermarking.

Conclusion and Future Directions

The AR-SSDM framework provides a mathematically sound and practical approach to PSD-compliant QIM watermarking. While the current decoding approach is sub-optimal due to noise coloration during inverse filtering, the system's predictable performance makes it a strong candidate for real-world applications in speech and medical signal processing. Future work may focus on optimizing the decoder for non-white noise environments and testing against a broader range of signal processing attacks.

Summary updated

BibTeX
@article{yan2015power,
  author = {Bin Yan and Ya-Fei Wang and Lingyun Song and Hong-Mei Yang},
  title = {{Power Spectrum Compliant QIM Watermarking for Autoregressive Host Signals}},
  journal = {J. Inf. Hiding Multim. Signal Process.},
  year = {2015},
}
Figure