Adversarial attack on Speech-to-Text Recognition Models

Xiaolei Liu, Kun Wan, Yufei Ding

TL;DR

This paper introduces a novel method for generating audio adversarial examples that attack Automatic Speech Recognition (ASR) systems. By strategically adjusting the number and intensity of perturbations (Weighted Perturbation Technology and Sampling Perturbation Technology), they achieve significantly faster attack generation and improved robustness against noise compared to prior work, all within minutes.

What Problem Does This Paper Address?

State-of-the-art ASR systems, while powerful, are vulnerable to adversarial attacks. These attacks involve subtly modifying audio inputs so that an ASR system misinterprets them, transcribing them into unintended phrases. While adversarial attacks on images are well-studied, audio adversarial examples present unique challenges:

  • Efficiency: Generating effective audio adversarial examples is computationally expensive and time-consuming, often taking hours. This impracticality limits their real-world threat assessment.
  • Robustness: Existing audio adversarial examples are often fragile. They can lose their adversarial property when subjected to even minor noise, which is common in real-world audio recording and playback.
  • Metric Selection: Unlike image adversarial attacks that rely on well-established Lp-norm metrics, there's no clear consensus on which metrics are most suitable for audio, impacting the quality and imperceptibility of the generated examples.

This paper aims to address these limitations by developing a more efficient, robust, and imperceptible method for generating audio adversarial examples.

Methods

The authors propose two core techniques:

  • Weighted Perturbation Technology (WPT): This method focuses on optimizing the generation process by assigning different weights to distortions at various points in the audio. It leverages an Audio Sequence Location (ASL) model to identify "key points" in the audio that are most critical for the ASR's transcription. By applying larger weights to these key points, WPT accelerates the attack generation. Additionally, WPT reduces the learning rate in later stages to refine the attack without overshooting.
  • Sampling Perturbation Technology (SPT): This technique enhances the robustness of adversarial examples by reducing the number of perturbed audio points. It's based on the observation that ASR models (specifically those using Connectionist Temporal Classification - CTC) process audio sequentially. By strategically perturbing fewer points, SPT makes the adversarial examples less susceptible to noise and more resilient to transformations.

The paper also investigates various loss functions for measuring the difference between adversarial and original audio, including Total Variation Denoising (TVD), to improve the imperceptibility of the generated attacks.

Datasets

  • Mozilla Common Voice dataset (MCVD): This is an open, publicly available dataset used for training speech-enabled applications. The authors used the first 100 test instances from this dataset for their experiments.

Key Insights and Ideas

  • Targeted Perturbation is Key: Instead of uniformly perturbing the entire audio signal, focusing perturbations on "key points" identified by an ASL model significantly speeds up the attack generation.
  • Less is More for Robustness: Reducing the number of perturbed points (SPT) makes adversarial examples more robust to noise, a critical factor for real-world applicability.
  • WPT and SPT are Complementary: These two techniques work together to achieve both speed and robustness, addressing the core challenges of audio adversarial attack generation.
  • Audio Metrics Matter: The choice of loss function for measuring audio similarity is crucial. Metrics that better reflect human auditory perception (like the proposed TVD-based metric) lead to more imperceptible adversarial examples, even if raw SNR values aren't dramatically improved.

Key Results

  • Speed: The proposed method can generate adversarial audio examples in 4 to 5 minutes, a dramatic improvement over the hours required by previous state-of-the-art methods.
  • Robustness: The SPT-based and SPT-EOT-based approaches demonstrate significantly better robustness against noise compared to baseline methods and even EOT-based methods when the noise distribution differs from the assumed one.
  • Efficiency: The paper reports high success rates and good SNR/dB(δ) values, indicating effective attacks that are also relatively clean. For instance, Table 1 shows a success rate of 1 with an average SNR of 31.9 dB and dB(δ) of 46.92 for their attack.

Limitations and Future Work

  • Limitations:
    • While the paper claims general robustness, it acknowledges that EOT-based methods perform better when the noise distribution is identical or very similar to the one they were trained on.
  • The paper focuses on white-box attacks, assuming full knowledge of the ASR model's parameters.
  • Future Work: The authors suggest that their future work will focus on the defense of audio adversarial examples, which is a natural next step after developing effective attack methods. They also highlight the need for further research into appropriate loss functions for audio adversarial attacks.

Summary updated

Abstract

Recent studies have highlighted audio adversarial examples as a ubiquitous threat to state-of-the-art automatic speech recognition systems. Thorough studies on how to effectively generate adversarial examples are essential to prevent potential attacks. Despite many research on this, the efficiency and the robustness of existing works are not yet satisfactory. In this paper, we propose~\textit{weighted-sampling audio adversarial examples}, focusing on the numbers and the weights of distortion to reinforce the attack. Further, we apply a denoising method in the loss function to make the adversarial attack more imperceptible. Experiments show that our method is the first in the field to generate audio adversarial examples with low noise and high audio robustness at the minute time-consuming level.
BibTeX
@inproceedings{liu2020adversarial,
  author = {Xiaolei Liu and Kun Wan and Yufei Ding},
  title = {{Adversarial attack on Speech-to-Text Recognition Models}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  url = {https://openreview.net/forum?id=wOnXPNu0Ge},
}
Figure