TL;DR
This paper explores the privacy risks associated with personalized acoustic models in federated learning for speech recognition. The authors propose novel attack methods that can infer a speaker's identity from these personalized models without direct access to their speech data, demonstrating significant vulnerabilities in current federated learning setups.
What Problem Does This Paper Address?
Federated learning (FL) has emerged as a promising paradigm for training machine learning models, particularly in speech recognition, by allowing models to be trained on decentralized user data without the data ever leaving the user's device. This approach is lauded for its privacy-preserving capabilities, as raw user data is not shared with a central server. However, the paper highlights a critical vulnerability: while raw data is protected, the updates to the models themselves, especially personalized acoustic models (PAMs) tailored to individual speakers, can inadvertently leak sensitive information. Specifically, the authors investigate whether access to these PAMs, which are essentially refined versions of a global model for a specific user, can be exploited to identify that user. This is a crucial concern for privacy in FL, as PAMs are central to personalization in many speech applications.
Methods
The paper introduces two distinct attack models to infer speaker identity from personalized acoustic models.
- Attack Model A1: This method is based on analyzing the "footprint" of acoustic models on a fixed, independent dataset called an "indicator dataset." It involves comparing the activation patterns of a specific hidden layer (denoted by 'h') in both the global model and personalized models when processing utterances from this indicator dataset. For each personalized model, statistical measures (mean and standard deviation) of these activations are computed. A similarity score is then calculated between pairs of personalized models based on the Euclidean distance between their statistical representations. This score is used to perform speaker verification.
- Attack Model A2: This approach employs a neural network (NN) trained specifically to predict speaker identity. This NN takes as input the activation patterns (similar to those computed in Attack Model A1) derived from the indicator dataset using both the global and personalized models. The NN architecture includes Time Delay Neural Network (TDNN) layers and a statistical pooling layer, similar to x-vector extractors used in speaker recognition. The trained NN then outputs embeddings that can be used for speaker verification.
Both attack models operate without direct access to the target speaker's speech data, relying solely on the parameters of the updated personalized models and the indicator dataset.
Datasets
The experiments were conducted using the TED-LIUM 3 corpus, a publicly available dataset containing approximately 452 hours of English speech from around 2,000 speakers.
- Train-G: Used to train the initial global acoustic model.
- Part-1 & Part-2: Used to obtain personalized acoustic models for different speakers. These datasets have disjoint speaker sets from Train-G.
- Indicator Dataset: A newly curated dataset for the attack models, consisting of 320 utterances from 32 speakers. Importantly, these speakers are disjoint from those in Train-G, Part-1, and Part-2. This dataset is crucial for analyzing the "footprint" of the models without contamination from training or testing data.
Key Insights and Ideas
- PAMs are a significant privacy risk: The study convincingly demonstrates that personalized acoustic models, even when trained in a federated manner, are not inherently private and can be exploited to reveal speaker identity.
- The "footprint" reveals identity: The core idea that the way a model processes unseen data (its "footprint" on an indicator dataset) encodes speaker-specific information is a powerful insight. This suggests that model behavior, not just raw data, is a source of privacy leakage.
- Early layers are information-rich: The first hidden layer of the acoustic models was found to contain a substantial amount of speaker-specific information, particularly in its standard deviation values. This is a surprising finding, as one might expect later layers to capture more abstract, identity-related features.
- Attack models are highly effective: Both proposed attack models achieve remarkably low Equal Error Rates (EERs), indicating a high success rate in identifying speakers.
Key Results
The paper reports strong performance for both attack models:
- Attack Model A1: Achieved an EER as low as 0.86% when using the first hidden layer (h=1) and a combination of mean and standard deviation statistics.
- Attack Model A2: Showed significant improvement over A1 when using the 5th hidden layer (h=5), reducing the EER to 1.94%. However, for h=1, A2 did not outperform A1.
- The study highlights that for A1, using standard deviation alone yielded the lowest EER on the first layer, while using means alone was less informative. The combination of both statistics generally provided superior results.
Limitations and Future Work
- Indicator Dataset Dependency: The effectiveness of the attacks relies on the availability of an indicator dataset. The paper does not discuss how an attacker would acquire or construct such a dataset in a real-world scenario.
- Attack Model A2 Performance: While A2 shows promise, its performance was not consistently better than A1, particularly for earlier layers. The paper notes that for h=1, A2 did not offer an improvement.
- Future Work: The authors propose to leverage the finding that early layers of PAMs contain significant speaker information to develop more effective speaker verification systems. They also plan to explore further attack strategies and privacy-preserving techniques for federated learning in speech recognition.
Summary updated