Scholarlink

We introduce the problem of private participation in federated learning (FL) systems. In this problem, different data owners can participate in different FL training tasks without revealing exactly which task they are involved in. It is extremely important in some metadata-sensitive scenarios (e.g., a patient does not want to disclose the fact that he/she is diseased but wants to contribute to the disease study). Despite the inherent privacy assurance of conventional FL techniques and recent advances in secure aggregations, such private participation remains an open issue. This work introduces VizardFL, an FL framework that efficiently enables private participation. At a high level, VizardFL is built out of distributed trust across two servers that keep client participation private as long as there is no collusion.

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

VizardFL: Enabling Private Participation in Federated Learning Systems

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yichen Zang,
  • Chengjun Cai,
  • Wentao Dong,
  • Cong Wang

摘要

We introduce the problem of private participation in federated learning (FL) systems. In this problem, different data owners can participate in different FL training tasks without revealing exactly which task they are involved in. It is extremely important in some metadata-sensitive scenarios (e.g., a patient does not want to disclose the fact that he/she is diseased but wants to contribute to the disease study). Despite the inherent privacy assurance of conventional FL techniques and recent advances in secure aggregations, such private participation remains an open issue. This work introduces VizardFL, an FL framework that efficiently enables private participation. At a high level, VizardFL is built out of distributed trust across two servers that keep client participation private as long as there is no collusion.