Summary of Noise Masking Attacks and Defenses For Pretrained Speech Models, by Matthew Jagielski et al.
Noise Masking Attacks and Defenses for Pretrained Speech Models
by Matthew Jagielski, Om Thakkar, Lun Wang
First submitted to arxiv on: 2 Apr 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper investigates noise masking attacks on speech models, which can lead to privacy leakage by revealing sensitive information from training data. The authors build upon previous work that demonstrated such attacks on automatic speech recognition (ASR) models. Here, they extend these attacks to pretrained speech encoders and show how fine-tuning the encoder for ASR and then applying noise masking can recover private information from pretraining data. The paper also discusses methods to improve the precision of these attacks and evaluates various countermeasures. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Noise models are often trained on secret data to make them better, but this can reveal personal details. Researchers have shown that attackers can trick these models into saying sensitive things by adding noise to part of an audio recording. The new study takes this idea and applies it to other types of speech models. They show that if you fine-tune a model to recognize words and then add noise, it will still reveal secrets from when the model was trained. The paper explores ways to make these attacks more accurate and looks at methods to prevent them. |
Keywords
* Artificial intelligence * Encoder * Fine tuning * Precision * Pretraining
