TL;DR
This paper introduces a novel method for creating adversarial audio attacks that are robust in real-world, over-the-air scenarios. By leveraging psychoacoustic principles and generating realistic room impulse responses, they demonstrate attacks that are both imperceptible to humans and effective at fooling Automatic Speech Recognition (ASR) systems, even when played through speakers.
What Problem Does This Paper Address?
Automatic Speech Recognition (ASR) systems, ubiquitous in voice assistants, are vulnerable to adversarial attacks. Previous attacks often operated solely in the digital domain or in controlled environments, limiting their real-world applicability. Furthermore, many attacks produced audible distortions, making them easily detectable. This paper addresses the challenge of creating adversarial audio attacks that are:
- Physically realizable: They can be executed in the real world, not just in simulations.
- Robust: They maintain their effectiveness despite the transformations audio undergoes when played through speakers and captured by microphones (e.g., room acoustics, noise).
- Imperceptible: They are difficult for humans to detect, preserving the naturalness of the audio.
Methods
The core of this work lies in a novel formulation of a psychoacoustic-based loss function for adversarial audio generation. Key techniques include:
- Psychoacoustic Masking: The authors exploit the human auditory system's properties, specifically how certain sounds can mask others. This allows them to embed adversarial perturbations within frequencies that are less perceptible to human ears.
- Time-Domain Attack Generation: Instead of complex frequency-domain manipulations, the attack is formulated and optimized primarily in the time domain. This simplifies the process and reduces memory requirements.
- Automated Room Impulse Response (RIR) Generation: To simulate real-world acoustic environments, they use a room simulator to generate RIRs that capture how sound propagates and reflects in different spaces. This enhances the robustness of the attacks against over-the-air transmission.
- Perceptual Evaluation of Speech Quality (PESQ): To objectively measure the imperceptibility of the attacks, they use the PESQ score, an established metric that correlates well with human perception of speech quality, eliminating the need for subjective listening tests.
- Projected Gradient Descent (PGD): This optimization method is used to find the adversarial perturbation, constrained by an L2-norm to keep the attack localized.
Datasets
- LibriSpeech: Used for training the DeepSpeech ASR model. The dataset consists of 960 hours of clean audio with corresponding transcriptions. Data augmentation (noise, pitch, and time shifting) was applied during training.
- No new benchmarks were introduced. The evaluation was performed on the Librispeech test set.
Key Insights and Ideas
- Exploiting Human Hearing for Stealth: The most significant insight is that by understanding and leveraging psychoacoustic masking, adversarial attacks can be made virtually undetectable to humans while still being effective against ASR systems.
- Real-World Robustness is Achievable: The integration of realistic acoustic modeling (via RIRs) demonstrates that adversarial attacks can be made robust to the complexities of over-the-air transmission.
- Time-Domain Simplicity: Formulating the attack in the time domain, after extracting perceptual properties, offers a more efficient and stable approach compared to purely frequency-domain methods.
- PESQ as an Objective Metric: Using PESQ for evaluation provides a reliable, automated way to assess the perceptual quality of adversarial audio, bridging the gap between objective signal metrics and human perception.
Key Results
- High Success Rate: The generated adversarial audio achieved a 100% success rate (WER=CER=0) on sampled audio files when no perceptual weighting was used (α=1). Even with perceptual weighting (α=0.8), the success rate remained 100%.
- Improved Perceptual Quality: Perceptually weighted attacks (lower α) resulted in higher PESQ scores (closer to 4.5, indicating excellent quality) compared to non-perceptually weighted attacks, despite a slight decrease in SNR.
- Over-the-Air Effectiveness: The attacks demonstrated robustness in over-the-air tests, with significant WER and CER reductions observed as the SNR increased. A language model decoder further improved attack performance in these scenarios.
- Imperceptibility: The PESQ scores, particularly with perceptual weighting, indicated that the adversarial audio was perceived as very close to the original signal.
Limitations and Future Work
- White-Box Attacks: The current attacks are "white-box," meaning they require full knowledge of the ASR model. Future work aims to explore "black-box" attacks, which are more realistic.
- Model Generalizability: The transferability of these attacks to different ASR architectures needs further investigation.
- Pre-processing Steps: The attacks do not currently account for common ASR pre-processing steps like beamforming or noise reduction, which could significantly impact their effectiveness.
- Computational Cost: Constructing these attacks currently takes several minutes, which is not feasible for real-time applications. Future research will focus on making the attack generation computationally more efficient.
- Two-Speaker Scenario: While tested, the two-speaker over-the-air scenario showed varying results, with the language model decoder performing particularly well. Further exploration of multi-speaker environments is warranted.
Summary updated