Replay attack detection with auditory filter-based relative phase features

Zeyan Oo, Longbiao Wang, Khomdet Phapatanaburi, Meng Liu, Seiichi Nakagawa, Masahiro Iwahashi, Jianwu DangEURASIP J. Audio Speech Music. Process 2019 · DOI

Abstract

Automatic Speaker Verification (ASV) systems are increasingly vulnerable to replay attacks, where a recorded voice is used to spoof biometric security. This paper introduces a novel detection approach focusing on the often-ignored phase information of speech signals. By applying auditory filter banks—specifically Mel and Gammatone scales—to Relative Phase (RP) features, the researchers significantly improved the ability of security systems to distinguish between genuine human speech and replayed recordings, especially in noisy, real-world conditions.

Introduction

Automatic Speaker Verification (ASV) systems are widely used for biometric security, yet they remain vulnerable to "spoofing" attacks. One of the most common and easiest to execute is the replay attack: an attacker records a target's voice using a smartphone or hidden microphone and plays it back to the system.

Most existing countermeasures rely on magnitude-based features, such as Mel-frequency cepstral coefficients (MFCCs) or Constant Q Transform Cepstral Coefficients (CQCCs). However, replayed speech often contains subtle artifacts introduced by the recording and playback hardware that magnitude analysis might miss. This paper seeks to harness phase information—the timing and alignment of sound waves—to build more robust replay attack detectors.

Proposed Methodology

The authors build upon the concept of Relative Phase (RP), which normalizes phase variations to make them usable for machine learning. Their core innovation is the integration of auditory filter banks into the RP extraction process.

Auditory Filter-Based RP

The researchers proposed two primary enhancements to conventional linear RP:

  1. Mel-scale RP: Conventional RP uses a linear frequency scale. The authors applied a Mel-filter bank, which mimics the human ear's non-linear perception of pitch, focusing more on lower frequencies.
  2. Gammatone-scale RP: They also tested a Gammatone filter bank, which provides even finer frequency resolution at lower ranges than the Mel scale, further simulating biological processing.
Comparison of conventional, Mel-scale, and Gammatone-scale Relative Phase (RP) extraction processes.

Classification and Fusion

The classification is handled by a standard Gaussian Mixture Model (GMM). The system extracts features from the input speech and calculates a likelihood ratio between genuine and replayed speech models.

Flowchart of the GMM-based replay attack detection system.

Recognizing that magnitude and phase provide complementary information, the authors combined the scores from their new phase features with a standard magnitude feature (CQCC) using a weighted likelihood ratio.

Experimental Setup and Results

The study utilized the ASVspoof 2017 challenge dataset, which includes recordings made in diverse real-world acoustic environments. The effectiveness of the features was measured using the Equal Error Rate (EER), where a lower percentage indicates better accuracy.

On the difficult Evaluation set:

  • Baseline Magnitude (CQCC): 21.61% EER
  • Conventional Linear RP: 13.15% EER
  • Proposed Mel-scale RP: 10.98% EER
  • Proposed Gammatone-scale RP: 11.07% EER
  • Best Fusion (CQCC + Gammatone-RP): 9.48% EER

Key Insights

  • Phase Robustness: While magnitude features performed well on clean data, their performance dropped significantly on real-world evaluation data. In contrast, phase features remained stable, suggesting that phase artifacts from replaying are more universal.
  • Bio-Inspiration: Moving from a linear scale to biologically-inspired scales (Mel or Gammatone) improved the EER, suggesting that critical phase information is concentrated in frequency bands used for human hearing.
  • Complementarity: The best results came from the fusion of magnitude (CQCC) and phase (Gammatone-RP), indicating that magnitude captures vocal tract shape while phase captures the physical characteristics of the sound source.

Conclusion

The proposed auditory filter-based phase features offer a significant improvement in replay attack detection. By mimicking human hearing, these features provide a more robust and reliable defense for ASV systems. Future work will explore modern Deep Neural Network (DNN) classifiers and additional feature refinement techniques to further enhance security.

Summary updated

BibTeX
@article{oo2019replay,
  author = {Zeyan Oo and Longbiao Wang and Khomdet Phapatanaburi and Meng Liu and Seiichi Nakagawa and Masahiro Iwahashi and Jianwu Dang},
  title = {{Replay attack detection with auditory filter-based relative phase features}},
  journal = {EURASIP J. Audio Speech Music. Process.},
  year = {2019},
  doi = {10.1186/s13636-019-0151-2},
}
Figure