Speech Privacy Leakage from Shared Gradients in Distributed Learning

Zhuohang Li, Jiaxin Zhang, Jian Liu

TL;DR

Researchers have developed a two-stage "gradient inversion" attack that can reconstruct private audio recordings from the mathematical updates (gradients) shared during federated learning. The study reveals that common speech features like Mel-spectrograms are highly vulnerable, allowing attackers to recover not only what was said but also the specific identity of the speaker with over 90% accuracy.

What Problem Does This Paper Address?

Distributed learning, particularly Federated Learning (FL), is often touted as a privacy-saving technology. Instead of sending raw data (like your voice recordings) to a central server, your device processes the data locally and only shares "gradients"—mathematical descriptions of how the model should update its weights.

While "gradient leakage" has been extensively studied in the image domain (where attackers can reconstruct pictures of faces or objects), the speech domain has remained relatively unexplored. Speech data is unique because models rarely "see" the raw waveform; instead, they process condensed acoustic features like Mel-spectrograms or MFCCs (Mel-frequency cepstral coefficients). This paper asks: If an attacker only sees the gradients from a speech model, can they work backward to hear your private conversations?

Methods

The authors propose a two-stage inversion method to bridge the gap between abstract model updates and raw audio:

  1. Feature Reconstruction: The attacker treats the shared gradients as a puzzle. By setting up an optimization problem, they search for a set of acoustic features (like a spectrogram) that, when fed into the model, would produce the exact same gradients they intercepted. They use a technique called anisotropic total variation as a regularizer to ensure the reconstructed features look like natural speech patterns.
  2. Waveform Reconstruction: Once the features are recovered, the attacker needs to turn them back into sound. For Mel-spectrograms, they use the Griffin-Lim algorithm to estimate the missing phase information and synthesize audio. For MFCCs, they add an extra step using an Inverse Discrete Cosine Transform (iDCT) to get back to a spectrogram representation first.

The attack assumes an "honest-but-curious" server—one that follows the FL protocol but secretly tries to reconstruct user data from the updates it receives.

Datasets

The study evaluated the attack using the Speech Commands dataset, a standard benchmark for keyword spotting.

  • Scope: They focused on 10 common words (e.g., "yes", "no", "up", "down", "stop").
  • Audio Quality: 1-second recordings sampled at 16kHz.
  • Model: A convolutional neural network (CNN) designed for keyword recognition, similar to those used in production environments for "wake word" detection.

Key Insights and Ideas

  • The Feature Gap: Unlike images, where the "input" is the final product, speech models have a "front-end" feature extractor. The authors found that this layer acts as a partial barrier, but not an impenetrable one.
  • Mel-spectrograms vs. MFCCs: There is a massive difference in vulnerability between feature types. Mel-spectrograms are nearly as easy to invert as raw data. MFCCs, however, are much more resilient. Because MFCCs use a decibel scale and have higher variance, the optimization process often gets "stuck," leading to distorted, less intelligible reconstructions.
  • Side-Channel Leakage: This is the "aha" moment. Even if a model is only trained to recognize a word (keyword spotting), the gradients still contain enough information to identify the speaker. This means a user might think they are only helping a model learn the word "Stop," while accidentally handing over their unique "voiceprint."

Key Results

The researchers measured success using both signal quality metrics (MSE, PESQ) and privacy-centric metrics:

  • Intelligibility: Reconstructions from Mel-spectrogram gradients achieved a PESQ score > 2.0, indicating the audio was clear and intelligible.
  • Speaker Re-identification: Using a pre-trained speaker verification model (ECAPA-TDNN), the researchers could match the reconstructed audio to the original speaker with a 90.5% success rate for Mel-spectrograms.
  • MFCC Resilience: MFCC-based models were significantly safer, with a speaker re-identification success rate of only 2.5% from gradients, though some keyword information was still potentially at risk.

Limitations and Future Work

The authors acknowledge a few key areas for further exploration:

  • Model Complexity: This study focused on keyword spotting (short utterances). It remains to be seen how well these attacks scale to complex Automatic Speech Recognition (ASR) models trained on long-form sentences.
  • Vocoder Quality: The use of the Griffin-Lim algorithm for audio synthesis is a classic but older approach. The authors suggest that using modern neural vocoders (like WaveNet or HiFi-GAN) could potentially make the leaked audio even more realistic and dangerous.
  • Defense Evaluation: While the paper identifies the vulnerability, it does not deeply explore defenses like Differential Privacy (DP) or Secure Multi-Party Computation (SMPC), which are the next logical steps for securing speech-based FL systems.

Summary updated

Abstract

Distributed machine learning paradigms, such as federated learning, have been recently adopted in many privacy-critical applications for speech analysis. However, such frameworks are vulnerable to privacy leakage attacks from shared gradients. Despite extensive efforts in the image domain, the exploration of speech privacy leakage from gradients is quite limited. In this paper, we explore methods for recovering private speech/speaker information from the shared gradients in distributed learning settings. We conduct experiments on a keyword spotting model with two different types of speech features to quantify the amount of leaked information by measuring the similarity between the original and recovered speech signals. We further demonstrate the feasibility of inferring various levels of side-channel information, including speech content and speaker identity, under the distributed learning framework without accessing the user’s data.
BibTeX
@inproceedings{li2023speech,
  author = {Zhuohang Li and Jiaxin Zhang and Jian Liu},
  title = {{Speech Privacy Leakage from Shared Gradients in Distributed Learning}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2023},
  url = {https://openreview.net/forum?id=6r0wJIT51a},
}

Loading related content...

Figure