Scholarlink

This paper introduces PriMo, a novel stochastic optimization algorithm designed for decentralized machine learning with robust data privacy protections. PriMo achieves differential privacy for each training device’s local dataset, safeguarding sensitive information while maintaining high training performance. Our algorithm is crafted to match the training loss and convergence of a centralized, non-private baseline. Through theoretical analysis and numerical experiments, we show that PriMo provides strong privacy guarantees without compromising model performance, making it a promising privacy-preserving solution for decentralized optimization.

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

PRIMO: A Privacy-Preserving Decentralized Machine Learning Algorithm

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Le Trieu Phong,
  • Tran Thi Phuong

摘要

This paper introduces PriMo, a novel stochastic optimization algorithm designed for decentralized machine learning with robust data privacy protections. PriMo achieves differential privacy for each training device’s local dataset, safeguarding sensitive information while maintaining high training performance. Our algorithm is crafted to match the training loss and convergence of a centralized, non-private baseline. Through theoretical analysis and numerical experiments, we show that PriMo provides strong privacy guarantees without compromising model performance, making it a promising privacy-preserving solution for decentralized optimization.