Scholarlink

Automatic Speech Recognition (ASR) algorithms are crucial for advancing audio-speech applications. The widespread distribution of audio data on edge devices motivates the use of Federated Learning (FL), where devices send gradient updates to a central server for privacy-preserving model training. We show that we can use gradient descent to reconstruct speech features following gradient privacy leakage attacks on other domains. By using first order methods, the optimization speed is significantly fast and the optimization procedure is simple.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Reconstructing Speech Features of Automatic Speech Recognition Systems in Federated Learning by Gradient Descent

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Minh N. Bui,
  • Trung Dang,
  • Paul Cherian,
  • Trac D. Tran,
  • Peter Chin

摘要

Automatic Speech Recognition (ASR) algorithms are crucial for advancing audio-speech applications. The widespread distribution of audio data on edge devices motivates the use of Federated Learning (FL), where devices send gradient updates to a central server for privacy-preserving model training. We show that we can use gradient descent to reconstruct speech features following gradient privacy leakage attacks on other domains. By using first order methods, the optimization speed is significantly fast and the optimization procedure is simple.