Audio Adversarial Examples: Targeted Attacks on Speech-to-Text

Nicholas Carlini, David A. Wagner

Abstract

We construct targeted audio adversarial examples on automatic speech recognition. Given any audio waveform, we can produce another that is over 99.9% similar, but transcribes as any phrase we choose (recognizing up to 50 characters per second of audio). We apply our white-box iterative optimization-based attack to Mozilla's implementation DeepSpeech end-to-end, and show it has a 100% success rate. The feasibility of this attack introduces a new domain to study adversarial examples.

Introduction

Adversarial examples—inputs specifically crafted to deceive machine learning models—have been a major focus of research in computer vision. However, the audio domain has proven more resilient. Previous audio attacks were largely untargeted (causing general failure) or relied on "hidden" sounds outside human hearing.

This research addresses the challenge of targeted attacks on Automatic Speech Recognition (ASR) systems. Unlike image classifiers, ASR pipelines like Mozilla’s DeepSpeech are complex, involving Mel-Frequency Cepstrum (MFC) preprocessing and recurrent neural networks (RNNs) that process sequential data. This paper demonstrates that it is possible to force these systems to transcribe any audio clip as a specific, chosen phrase with a 100% success rate.

Methodology

The attack is framed as an optimization problem. For a given source audio $x$ and a target transcription $t$, the goal is to find a perturbation $\delta$ such that the model transcribes $x + \delta$ as $t$, while keeping $\delta$ small enough to be nearly inaudible.

The targeted attack process: adding a tiny perturbation to an original audio clip results in a new transcription while remaining nearly identical to the original.

Key technical innovations include:

  • End-to-End Optimization: The attack differentiates through the entire DeepSpeech pipeline, including the non-differentiable-looking MFC preprocessing and the deep LSTM network.
  • Refined Loss Function: While Connectionist Temporal Classification (CTC) loss is standard for training, the authors found that a two-step approach—first identifying a temporal alignment and then optimizing a logit-based loss—significantly reduced distortion.
  • Iterative Refinement: Using the Adam optimizer, the noise is iteratively tuned over thousands of steps to achieve the desired transcription with minimal signal change.

Key Insights

  • Information Density: The researchers found they could embed up to 50 characters of text per second of audio, allowing short clips to produce long, complex transcriptions.
  • Non-linearity of Audio: Unlike image models, which often succumb to simple "single-step" attacks (like FGSM) due to local linearity, ASR systems require iterative optimization because of the complex preprocessing and deep recurrent layers.
  • Versatility: The attack is powerful enough to turn music into speech or even make spoken words transcribe as silence (empty strings) by adding very low-level noise (-45dB).

Results and Evaluation

Using the Mozilla Common Voice dataset, the researchers tested the attack on 100 test instances, each paired with 10 random target transcriptions.

  • 100% Success Rate: Every single source-target pair was successfully generated.
  • Inaudible Distortion: The mean distortion was measured at -38 dB using the refined loss function. This level of noise is comparable to the ambient background noise in a quiet room, making the perturbation virtually undetectable to human listeners.
  • Computational Efficiency: Although generating a single example takes roughly an hour on a GPU, the process is parallelizable, allowing for the creation of multiple adversarial examples simultaneously.

Limitations and Future Directions

Currently, these attacks are "white-box," requiring full access to the target model's internal parameters. Furthermore, they are "digital-only"; the precise perturbations are often lost when played over speakers and re-recorded due to room acoustics and hardware limitations.

Future work will focus on:

  1. Over-the-air Robustness: Developing perturbations that survive physical playback and recording.
  2. Black-box Attacks: Creating attacks that do not require internal model knowledge.
  3. Transferability: Testing if adversarial audio designed for one system (like DeepSpeech) can successfully fool others, such as commercial voice assistants.

Summary updated

Abstract

We construct targeted audio adversarial examples on automatic speech recognition. Given any audio waveform, we can produce another that is over 99.9% similar, but transcribes as any phrase we choose (recognizing up to 50 characters per second of audio). We apply our white-box iterative optimization-based attack to Mozilla's implementation DeepSpeech end-to-end, and show it has a 100% success rate. The feasibility of this attack introduce a new domain to study adversarial examples.
BibTeX
@inproceedings{carlini2018audio,
  author = {Nicholas Carlini and David A. Wagner},
  title = {{Audio Adversarial Examples: Targeted Attacks on Speech-to-Text}},
  booktitle = {IEEE Symposium on Security and Privacy Workshops},
  year = {2018},
  url = {https://openreview.net/forum?id=lH9FAMDjcCV},
}
Figure