Abstract
Digital watermarking is an effective covert communication technique in the military field, which can solve the problems in traditional encryption. In general, the Spread Spectrum (SS) algorithm can well satisfy the requirements of concealment and anti-interference in the covert communication process. However, it has a natural defect in resisting desynchronization attacks. In this paper, we propose a Robust Feature Points Scheme (RFPS) against desynchronization attacks by calculating the maximum response value of the second-order derivative of the original audio signal. The maximum response value will be used as a feature point and the audio segments centering at the detected feature points are extracted for both watermark embedding and extraction process. This study mainly focuses on cropping attacks and pitch invariant Time-Scale Modification (TSM) attacks. The experimental results show that under non-uniform cropping attacks, the proposed method can extract the watermark information without any bit error rate. Even in the case of the pitch-invariant TSM attacks, the proposed scheme has great improvement compared with the existing works.
Introduction
Digital watermarking is essential for covert communication and copyright protection. However, the most common technique—Spread Spectrum (SS)—has a significant vulnerability: it is highly position-dependent. If an attacker performs a desynchronization attack, such as cropping a few milliseconds from the start of a file or slightly changing the playback speed (Time-Scale Modification or TSM), the receiver loses synchronization. Because the watermark is no longer where the algorithm expects it to be, the information becomes unrecoverable. Existing solutions often rely on complex synchronization codes that increase computational overhead and are themselves prone to failure.
Methods
The researchers developed RFPS (Robust Feature Points Scheme), a method that uses second-order derivatives to identify stable "anchor points" within an audio signal. By embedding watermarks relative to these points rather than fixed timestamps, the watermark remains recoverable even if the audio is cropped or time-stretched.
Feature Point Detection and Embedding
The core innovation is using the second-order derivative of the audio signal. By calculating how the "slope of the slope" of the waveform changes, the algorithm identifies points of maximum response. These points are physically tied to the audio content, making them remarkably stable even if the file is edited. The system selects the top $n$ feature points and extracts audio segments centered around them.
To ensure the watermark is inaudible and robust, the team applies a Stationary Wavelet Transform (SWT) to these segments. Unlike the standard Discrete Wavelet Transform, SWT is "shift-invariant," meaning a small error in finding the exact center of a feature point won't destroy the data. They use Pseudo-Noise (PN) sequences to represent the watermark bits, which are then embedded into the SWT coefficients.
Watermark Extraction
During extraction, the watermark is recovered by calculating the correlation between the extracted signal and the known PN sequences. A sharp correlation peak indicates the presence of a specific bit, allowing for reliable detection even after signal manipulation.
Key Results
The RFPS method was tested against two major attack types and compared to state-of-the-art methods:
- Non-Uniform Cropping: When 10% of the audio was cut from the front, middle, or back, RFPS achieved a 0% Bit Error Rate (BER). Competing methods saw error rates as high as 50%.
- Jittering Attacks: Under "uniform cropping" (deleting 1 out of every 500 samples), RFPS maintained a BER of 12.5%, while previous methods failed completely with error rates over 60%.
- Time-Scale Modification (TSM): In pitch-invariant stretching (±5% to ±15%), RFPS showed significant improvement. For example, at -5% TSM, RFPS had a BER of 17.86%, compared to ~46% for the baseline SS method.
- Efficiency: The embedding process takes roughly 4.4 seconds, which is considered an acceptable trade-off for the massive jump in reliability.
Conclusion
While RFPS is a major step forward, it isn't invincible. As the intensity of time-stretching increases, the accuracy of feature point detection begins to degrade. The researchers plan to look for even more "steady" features—characteristics of audio that remain unchanged regardless of the environment or attack type—to push the Bit Error Rate even lower in extreme TSM scenarios.
Summary updated