Watermarking Based on Compressive Sensing for Digital Speech Detection and Recovery

Wenhuan Lu, Zonglei Chen, Ling Li, Xiaochun Cao, Jianguo Wei, Naixue Xiong, Jian Li, Jianwu Dang

TL;DR

This paper introduces a novel watermarking scheme for digital speech that is imperceptible, fragile, and blind, enabling both tampering detection and self-recovery. By embedding watermark data derived from DCT coefficients and leveraging compressive sensing, the method can accurately locate tampered areas and recover the original speech, even with significant tampering (up to 47.6%). A deep learning enhancement step further improves the quality of the recovered audio.

What Problem Does This Paper Address?

In today's digital world, multimedia content, including speech, is easily transmitted and can be susceptible to malicious tampering or accidental corruption during transmission. This poses a significant challenge for verifying the integrity of speech data and recovering it to its original state. While digital watermarking offers a solution for authentication and recovery, existing methods often struggle with a trade-off between efficiently using watermark data and preventing "tampering coincidence" (where both the original data and its watermark are corrupted, rendering recovery impossible). This paper aims to address these limitations by proposing a robust and effective speech watermarking scheme that can detect tampering, localize the damaged areas, and recover the original speech content with high fidelity.

Methods

The proposed watermarking scheme involves two main stages: embedding and recovery.

Embedding: 1. Framing and DCT: The speech signal is divided into non-overlapping frames. The Discrete Cosine Transform (DCT) is applied to each frame to obtain its coefficients. 2. Watermark Generation: Watermark data is generated from these DCT coefficients. Instead of embedding information in a single frame, the watermark data is shared across a "frames-group." This approach aims to balance the "data waste problem" (where watermark data is unused if the host signal isn't tampered) and the "tampering coincidence problem." 3. Permutation and Quantization: DCT coefficients are pseudo-randomly permuted within frames. Reference values are calculated and then quantized to generate integers. 4. Check-bit Calculation: Hash-bits are generated from frame information and reference values, then XORed with identical "label-bits" to create "check-bits." These check-bits, along with reference values, form the watermark. 5. LSB Embedding: The watermark is embedded into the Least Significant Bits (LSBs) of the speech signal.

Recovery: 1. Tampered Area Localization: The received signal is divided into frames, and watermark bits (reference values and check-bits) are extracted. By comparing calculated label-bits (derived from hash-bits and extracted check-bits) across frames, tampered areas are identified. Frames with identical label-bits to a majority are considered "reserved." 2. Data Recovery using Compressive Sensing: For tampered frames, the DCT coefficients are recovered using compressive sensing. This technique exploits the inherent sparsity of DCT coefficients in the frequency domain. The method uses the extracted reference values from reserved frames to reconstruct the DCT coefficients of the tampered frames. 3. Inverse DCT and Enhancement: An Inverse DCT is applied to recover the speech signal. Additionally, a deep learning-based enhancement method is employed to further improve the Signal-to-Noise Ratio (SNR) of the recovered speech.

Datasets

The paper uses the CASIA-863-speech synthesis database for its experiments. Five sentences from this database were selected, resampled to 16-bit 8-kHz, and used for both subjective and objective evaluations. No new benchmarks were introduced; existing methods were compared against the proposed approach.

Key Insights and Ideas

  • Frames-Group Watermarking: Sharing watermark information across multiple frames, rather than embedding it within a single frame, is a clever way to mitigate both data waste and the risk of complete watermark destruction due to tampering.
  • Compressive Sensing for Speech Recovery: Applying compressive sensing, a technique traditionally used in signal processing for sparse signals, to recover tampered DCT coefficients in speech is a novel approach. It leverages the inherent sparsity of speech in the DCT domain for efficient reconstruction.
  • Trade-off Management: The scheme explicitly addresses the trade-off between the "data waste problem" and the "tampering coincidence problem," which are common challenges in watermark design.
  • Deep Learning for Enhancement: Integrating a deep learning model for post-processing the recovered speech is a significant step towards achieving high-quality audio restoration, going beyond basic recovery.

Key Results

  • Imperceptibility: Subjective tests (SDG and ABX methods) indicated that the watermarked speech signals were perceptually indistinguishable from the original, with average SDG scores close to zero and correct detection rates in the ABX test suggesting minimal perceivability.
  • Tampering Recovery: The proposed method demonstrated effective recovery of tampered speech signals, even at high tampering rates. The paper reports intelligible recovered signals for tampering rates up to 47.6%.
  • SNR Improvement: The deep learning enhancement significantly improved the SNR of the recovered speech. For instance, one experiment showed an SNR improvement from -0.20249 (without processing) to 2.0729 with a specific network configuration.
  • Robustness to Tampering: The SNR of recovered signals dropped gradually with increasing tampering percentages, indicating a stable recovery performance. The smaller the tampered area, the better the quality of the recovered signal.
  • Comparison with Existing Methods: The proposed method showed competitive or superior performance compared to several recent methods in terms of SNR and Mean Opinion Score (MOS), particularly in imperceptibility.

Limitations and Future Work

  • Fragility and Robustness: The paper acknowledges that the proposed watermark is fragile, meaning it is easily detectable and can be attacked. This fragility, while good for tamper detection, limits its robustness against intentional attacks.
  • Noise in Recovery: Despite the deep learning enhancement, the paper notes that some noise remains in the recovered speech signal.
  • Future Work: The authors suggest future research directions including:
  • Investigating post-processing techniques for recovered speech to further improve its quality and reduce residual noise.
  • Developing semi-fragile watermarking schemes to enhance robustness against attacks while still allowing for tamper detection.
  • Exploring different deep learning architectures and training strategies for more effective speech enhancement.

Summary updated

Abstract

In this paper, a novel imperceptible, fragile and blind watermark scheme is proposed for speech tampering detection and self-recovery. The embedded watermark data for content recovery is calculated from the original discrete cosine transform (DCT) coefficients of host speech. The watermark information is shared in a frames-group instead of stored in one frame. The scheme trades off between the data waste problem and the tampering coincidence problem. When a part of a watermarked speech signal is tampered with, one can accurately localize the tampered area, the watermark data in the area without any modification still can be extracted. Then, a compressive sensing technique is employed to retrieve the coefficients by exploiting the sparseness in the DCT domain. The smaller the tampered the area, the better quality of the recovered signal is. Experimental results show that the watermarked signal is imperceptible, and the recovered signal is intelligible for high tampering rates of up to 47.6%. A deep learning-based enhancement method is also proposed and implemented to increase the SNR of recovered speech signal.
BibTeX
@inproceedings{lu2018watermarking,
  author = {Wenhuan Lu and Zonglei Chen and Ling Li and Xiaochun Cao and Jianguo Wei and Naixue Xiong and Jian Li and Jianwu Dang},
  title = {{Watermarking Based on Compressive Sensing for Digital Speech Detection and Recovery}},
  booktitle = {Sensors},
  year = {2018},
  url = {https://openreview.net/forum?id=FLzeEH4k5c},
}
Figure