Interpretable Spectrum Transformation Attacks to Speaker Recognition

Jiadi Yao, Hong Luo, Xiao-Lei ZhangarXiv 2023 · DOI · arXiv

Abstract

Adversarial attacks on speaker recognition systems often struggle with transferability in black-box settings. This paper proposes the Spectral Transformation Attack based on Modified Discrete Cosine Transform (STA-MDCT). By shifting the attack from the time domain to the time-frequency domain and applying spectral modifications, STA-MDCT significantly improves attack success rates against unknown victim models. Additionally, the framework utilizes Layer-Class Activation Maps (Layer-CAM) to provide visual evidence of how adversarial perturbations shift a model's focus, offering a new level of interpretability in audio adversarial research.

Introduction

Speaker recognition systems are increasingly vulnerable to adversarial attacks—subtle perturbations designed to fool AI models. While white-box attacks, where the attacker has full access to the model, are highly effective, real-world scenarios often involve black-box models. In these cases, attackers rely on the transferability of adversarial examples from a known surrogate model to an unknown victim. Traditional time-domain attacks often fail to generalize across different architectures because they ignore the frequency-dependent nature of how neural networks process audio.

Methodology

The STA-MDCT framework addresses these limitations by operating in the time-frequency domain. The process begins by transforming raw audio using the Modified Discrete Cosine Transform (MDCT). Unlike standard time-domain modifications, STA-MDCT applies spectral transformations, including Gaussian noise addition and random scaling of spectral coefficients via a scaling matrix (Hadamard product). These transformations help capture salient features that are more likely to be shared across diverse model architectures, enhancing the transferability of the generated adversarial voice.

The STA-MDCT framework compared to traditional transfer-based black-box attacks.

Interpretability via Saliency Maps

A key contribution of this work is the introduction of interpretability through Layer-CAM. By generating saliency maps, researchers can visualize the regions of the spectrogram that a model prioritizes for identification. Successful STA-MDCT attacks are shown to physically shift the model's attention away from critical speaker features toward irrelevant noise regions, leading to misclassification. This visual evidence confirms that the spectral modifications effectively disrupt the model's internal representation of the speaker.

Saliency maps generated via Layer-CAM showing model attention on different architectures.

Key Results

The effectiveness of STA-MDCT was evaluated against several baseline attackers (e.g., FGSM, MI-FGSM, ACG) across multiple victim architectures, including ResNetSE34L, ECAPA-TDNN, and RawNet3.

  • Superior Transferability: In black-box Automatic Speaker Verification (ASV) tasks, STA-MDCT achieved relative Equal Error Rate (EER) improvements of 42.56%, 44.18%, and 31.1% over MI-FGSM, NI-FGSM, and ACG, respectively.
  • High Success Rates: In targeted attack scenarios, the Targeted Attack Success Rate (TASR) reached as high as 70.5%.
  • Stealthiness: Despite its potency, the attack maintains high stealthiness. Perceptual Evaluation of Speech Quality (PESQ) scores indicate that STA-MDCT adversarial examples are more imperceptible to human listeners than those produced by traditional time-domain methods.

Conclusion

STA-MDCT represents a significant advancement in transferable adversarial attacks for speaker recognition. By leveraging the frequency domain and providing visual interpretability, the framework not only improves attack efficacy but also offers insights into the specific vulnerabilities of audio-based AI models. Future research will focus on refining these spectral transformations and developing more robust defensive mechanisms against such sophisticated spectral attacks.

Summary updated

Abstract

The success of adversarial attacks to speaker recognition is mainly in white-box scenarios. When applying the adversarial voices that are generated by attacking white-box surrogate models to black-box victim models, i.e. \textit{transfer-based} black-box attacks, the transferability of the adversarial voices is not only far from satisfactory, but also lacks interpretable basis. To address these issues, in this paper, we propose a general framework, named spectral transformation attack based on modified discrete cosine transform (STA-MDCT), to improve the transferability of the adversarial voices to a black-box victim model. Specifically, we first apply MDCT to the input voice. Then, we slightly modify the energy of different frequency bands for capturing the salient regions of the adversarial noise in the time-frequency domain that are critical to a successful attack. Unlike existing approaches that operate voices in the time domain, the proposed framework operates voices in the time-frequency domain, which improves the interpretability, transferability, and imperceptibility of the attack. Moreover, it can be implemented with any gradient-based attackers. To utilize the advantage of model ensembling, we not only implement STA-MDCT with a single white-box surrogate model, but also with an ensemble of surrogate models. Finally, we visualize the saliency maps of adversarial voices by the class activation maps (CAM), which offers an interpretable basis to transfer-based attacks in speaker recognition for the first time. Extensive comparison results with five representative attackers show that the CAM visualization clearly explains the effectiveness of STA-MDCT, and the weaknesses of the comparison methods; the proposed method outperforms the comparison methods by a large margin.
BibTeX
@misc{yao2023interpretable,
  author = {Jiadi Yao and Hong Luo and Xiao-Lei Zhang},
  title = {{Interpretable Spectrum Transformation Attacks to Speaker Recognition}},
  year = {2023},
  eprint = {2302.10686},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2302.10686},
}
Figure