Core Concepts
Proposing a novel privacy-preserving model for spoken language understanding using hidden layer separation and adversarial training.
Abstract
The content discusses the importance of privacy in spoken language understanding systems, introducing a novel model to prevent attacks on speech recognition and identity recognition. It outlines experiments, datasets used, and results to show the effectiveness of the proposed model.
Abstract:
SLU is crucial for human-computer interaction in IoT devices.
Privacy breaches due to user-sensitive information in speech.
Proposed privacy-preserving model using hidden layer separation and adversarial training.
Introduction:
Voice-controlled IoT devices gaining popularity.
Privacy risks associated with limited storage space models.
End-to-end SLU systems vulnerable to privacy breaches.
Data Extraction:
Experiments over two SLU datasets show proposed method reduces accuracy of attacks close to random guess.
Related Work:
Early SLU tasks transitioned from ASR-NLU models to end-to-end systems.
Voice privacy protection methods evolving with deep learning advancements.
Privacy-preserving SLU:
Model separates hidden layer for SLU, ASR, and IR tasks.
Adversarial training enhances privacy preservation ability.
Experiments:
Datasets used include LibriSpeech, VoxCeleb1, FSC, SLURP, TED-Lium.
Setup includes feature extraction, encoder-decoder configurations.
Results show proposed model maintains SLU accuracy while reducing attacker success rate.
Stats
Experiments over two SLU datasets show that the proposed method can reduce the accuracy of both the ASR and IR attacks close to that of a random guess.
Quotes
"Users do not want to expose their personal sensitive information to malicious attacks by untrusted third parties."
"The proposed method maintains the performance of SLU well while reducing the success rate of attackers close to that of a random guess."