VoiceBlock: Privacy through Real-Time Adversarial Attacks with Audio-to-Audio Models

Patrick O'Reilly, Andreas Bugler, Keshav Bhandari, Max Morrison, Bryan PardoNeurIPS 2022 · OpenReview

The Privacy Challenge in Voice Communications

Mass surveillance of voice communications is a growing concern, fueled by the rapid adoption of deep-learning-based automatic speaker recognition (ASR). These systems can automatically identify individuals or link utterances across different recordings, often without warrants or consent.

While "adversarial attacks"—small, intentional audio perturbations designed to fool AI—offer a potential defense, they face two major hurdles in the context of voice:

  1. Latency: Most effective attacks require iterative, offline optimization, making them too slow for live streaming or phone calls.
  2. Perceptibility: Many attacks rely on adding "noise" to the waveform. To human ears, this sounds like a conspicuous hiss or distortion, severely degrading the user experience.

VoiceBlock bridges this gap by creating a system that is fast enough for live use and subtle enough that humans barely notice it.

How VoiceBlock Works

VoiceBlock moves away from "additive" noise and instead uses multiplicative filtering. The system is an audio-to-audio model consisting of three main components:

VoiceBlock Architecture and Embedding Space
  • The Encoder: Extracts four types of features from the audio: pitch, loudness, phonetic posteriorgrams (which capture linguistic content), and spectrogram features. It also uses a Feature-wise Linear Modulation (FiLM) layer, which takes a short "enrollment" embedding of the user's voice to tailor the attack to that specific speaker.
  • The Bottleneck: A recurrent module (using LSTMs) that incorporates context from past frames. Crucially, it uses a tiny "lookahead" of 5 frames (about 48ms) to see slightly into the future. This significantly strengthens the attack while keeping total latency around 56ms—well within the acceptable range for VoIP calls.
  • The Decoder: Instead of generating new audio from scratch, the decoder predicts the parameters for a time-varying Finite Impulse Response (FIR) filter. This filter is applied to the original audio, essentially "shaping" the frequencies to mislead the AI without introducing the artifacts common in synthetic speech.

The model is trained using a dual objective: an adversarial loss to push the speaker's identity away from their known profile in the AI's embedding space, and an auxiliary loss to ensure the resulting audio sounds as close to the original as possible.

Key Insights

  • Filtering vs. Adding: The core innovation is that filtering is inherently less intrusive than adding noise. Because filtering only attenuates or amplifies existing frequencies, it avoids creating the "unnatural" sounds that human ears easily detect.
  • Preserving Content with PPGs: By using Phonetic Posteriorgrams (PPGs), the model ensures that while the identity of the speaker is masked, the content of what they are saying remains perfectly clear.
  • Real-Time Feasibility: The authors proved that a complex adversarial attack doesn't require a massive GPU cluster. VoiceBlock runs on a single CPU thread, making it practical for integration into mobile apps or browser extensions.

Experimental Results

The researchers trained VoiceBlock using LibriSpeech and evaluated it on VoxCeleb1, a large-scale dataset containing over 150,000 utterances from 1,251 speakers. They tested against two different speaker recognition models: ResNetSE34v2 (a "seen" model it was trained against) and Y-Vector (an "unseen" model to test transferability).

Speech Quality and Attack Effectiveness Metrics
  • Effectiveness: On the ResNet model, Top-1 identification accuracy plummeted from 99% to just 3%. Even on the unseen Y-Vector model, accuracy dropped to 35%, significantly outperforming "universal" noise-based attacks.
  • Audio Quality: Using the PESQ metric (an industry standard for voice quality where higher is better), VoiceBlock scored an impressive 3.74, compared to just 1.03 for white noise and 1.35 for other universal attacks.
  • Human Perception: In a MUSHRA-style listening study, humans rated VoiceBlock audio as virtually identical in quality to the original "clean" recordings, whereas baseline anonymization methods were rated significantly lower.

Limitations and Future Work

While highly effective, VoiceBlock has a few constraints. First, it requires an enrollment period; the model currently needs about one minute of the user's speech beforehand to create a "source embedding." Second, there is still a transferability gap when attacking models it wasn't specifically trained on (the "black-box" scenario).

Future work will focus on improving the generalization of the attack—making it equally effective against any recognition system without prior knowledge. The authors also plan to explore how this approach might protect against other types of invasive audio profiling, such as emotion or health state detection.

Summary updated

Abstract

As governments and corporations adopt deep learning systems to collect and analyze user-generated audio data, concerns about security and privacy naturally emerge in areas such as automatic speaker recognition. While audio adversarial examples offer one route to mislead or evade these invasive systems, they are typically crafted through time-intensive offline optimization, limiting their usefulness in streaming contexts. Inspired by architectures for audio-to-audio tasks such as denoising and speech enhancement, we propose a neural network model capable of adversarially modifying a user's audio stream in real-time. Our model learns to apply a time-varying finite impulse response (FIR) filter to outgoing audio, allowing for effective and inconspicuous perturbations on a small fixed delay suitable for streaming tasks. We demonstrate our model is highly effective at de-identifying user speech from speaker recognition and able to transfer to an unseen recognition system. We conduct a perceptual study and find that our method produces perturbations significantly less perceptible than baseline anonymization methods, when controlling for effectiveness. Finally, we provide an implementation of our model capable of running in real-time on a single CPU thread. Audio examples and code can be found at https://interactiveaudiolab.github.io/project/voiceblock.html.
BibTeX
@inproceedings{oreilly2022voiceblock,
  author = {Patrick O'Reilly and Andreas Bugler and Keshav Bhandari and Max Morrison and Bryan Pardo},
  title = {{VoiceBlock: Privacy through Real-Time Adversarial Attacks with Audio-to-Audio Models}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2022},
  url = {https://openreview.net/forum?id=8gQEmEgWAkc},
}
Figure