Speech Privacy Leakage from Shared Gradients in Distributed Learning

Zhuohang Li, Jiaxin Zhang, Jian Liu

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.

Introduction

Distributed learning, particularly Federated Learning (FL), is frequently championed as a privacy-preserving architecture. By design, FL allows devices to process data locally, sharing only mathematical updates—known as gradients—with a central server rather than the raw data itself. While the vulnerability of these gradients has been documented extensively in the computer vision domain (where images can often be reconstructed from model updates), the speech domain has remained relatively under-examined.

Speech data presents a unique challenge for attackers because models rarely process raw waveforms directly. Instead, they utilize condensed acoustic features such as Mel-spectrograms or Mel-frequency cepstral coefficients (MFCCs). This research investigates whether an "honest-but-curious" server can exploit these shared gradients to reconstruct private audio recordings and extract sensitive side-channel information, such as the speaker's identity.

Methodology: The Two-Stage Inversion Attack

To bridge the gap between abstract gradient updates and audible speech, the researchers developed a two-stage inversion method designed to recover the original input signal.

1. Feature Reconstruction

In the first stage, the attacker treats the intercepted gradients as an optimization puzzle. The goal is to find a set of acoustic features that, when processed by the local model, would generate the exact gradients observed. To ensure the reconstructed features resemble natural speech rather than random noise, the authors employ anisotropic total variation as a regularizer. This helps maintain the structural integrity of the spectrogram patterns during the optimization process.

2. Waveform Reconstruction

Once the acoustic features are recovered, they must be converted back into a playable audio format.

  • Mel-spectrograms: The researchers utilize the Griffin-Lim algorithm to estimate missing phase information and synthesize the audio waveform.
  • MFCCs: Because MFCCs are more compressed, an additional step involving an Inverse Discrete Cosine Transform (iDCT) is required to return the data to a spectrogram representation before waveform synthesis.

Experimental Setup

The attack was evaluated using the Speech Commands dataset, a standard benchmark for keyword spotting. The study focused on 10 common keywords (e.g., "yes", "no", "stop") recorded as 1-second clips at 16kHz. The target model was a convolutional neural network (CNN) optimized for wake-word detection, representing a typical production-grade speech application.

Key Findings

The Vulnerability of Feature Types

The research highlights a significant disparity in privacy resilience between different acoustic features. Mel-spectrograms proved highly vulnerable; the gradients contained enough information to reconstruct audio with a PESQ score greater than 2.0, indicating high intelligibility.

In contrast, MFCCs offered substantially better protection. Due to their use of a decibel scale and higher variance, the optimization process frequently failed to converge on a clear reconstruction, resulting in distorted and less intelligible audio.

Side-Channel Information Leakage

Perhaps the most concerning finding is the leakage of information unrelated to the primary task. Even though the model was trained solely for keyword spotting (recognizing what was said), the gradients leaked significant data regarding who was saying it.

Using a pre-trained speaker verification model (ECAPA-TDNN), the researchers attempted to match the reconstructed audio to the original speaker. For Mel-spectrogram-based models, the speaker re-identification success rate reached 90.5%. For MFCC-based models, this risk was mitigated significantly, dropping to a 2.5% success rate.

Discussion and Limitations

This study demonstrates that the "front-end" feature extractor of a speech model is not an absolute barrier to privacy. While MFCCs provide a degree of inherent protection, Mel-spectrograms—which are increasingly popular in modern deep learning—leave users highly exposed.

However, several factors may influence the severity of this leakage in real-world scenarios:

  • Model Complexity: The current study focused on short-form keyword spotting. The scalability of these attacks to complex Automatic Speech Recognition (ASR) systems involving long-form sentences remains an open question.
  • Synthesis Quality: The use of the Griffin-Lim algorithm represents a baseline. The authors suggest that the application of modern neural vocoders (such as HiFi-GAN) could potentially produce even more realistic and identifiable audio from leaked gradients.
  • Defensive Strategies: Future work must evaluate the efficacy of standard privacy defenses, such as Differential Privacy (DP) or Secure Multi-Party Computation (SMPC), in neutralizing these inversion attacks without compromising model accuracy.

Conclusion

As distributed learning becomes a standard for speech-enabled devices, the assumption that "sharing gradients equals sharing privacy" must be re-evaluated. The ability to reconstruct intelligible audio and identify speakers from model updates suggests that additional layers of protection are necessary to truly secure user voice data in federated environments.

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},
}
Figure