Scholarlink

<p>Federated Learning (FL) is a widely adopted approach that enables multiple clients to collaboratively train a model while keeping their data local. This decentralized setup reduces the need for data transfer, thus enhancing both privacy and security. However, FL remains vulnerable to Membership Inference Attacks (MIAs), which aim to determine whether specific data records were included in a model’s training set. In this study, we systematically analyze FL vulnerabilities using a real-world healthcare dataset. Based on the two-level structure of the data set, we define MIAs at record-level and client-level and evaluate the influence of various FL configurations, such as sampling strategies and client selection methods, on MIAs’ susceptibility. Furthermore, we propose a threshold-tuning mechanism to enhance client-level MIAs and examine the impact of synthetic data on the model’s resistance to MIAs. Our findings demonstrate the dual benefits of client selection, which simultaneously improves model performance and reduces MIAs’ vulnerabilities. By emphasizing these benefits, this study provides actionable guidance for developing privacy preserving FL frameworks that protect sensitive information without compromising model effectiveness.</p>

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

A study of membership inference attacks on a federated health care application

  • 期刊论文

  • 发表日期：

    2025年6月16日

查看全文
  • Christos Chrysanthos Nikolaidis,
  • Pavlos S. Efraimidis

摘要

Federated Learning (FL) is a widely adopted approach that enables multiple clients to collaboratively train a model while keeping their data local. This decentralized setup reduces the need for data transfer, thus enhancing both privacy and security. However, FL remains vulnerable to Membership Inference Attacks (MIAs), which aim to determine whether specific data records were included in a model’s training set. In this study, we systematically analyze FL vulnerabilities using a real-world healthcare dataset. Based on the two-level structure of the data set, we define MIAs at record-level and client-level and evaluate the influence of various FL configurations, such as sampling strategies and client selection methods, on MIAs’ susceptibility. Furthermore, we propose a threshold-tuning mechanism to enhance client-level MIAs and examine the impact of synthetic data on the model’s resistance to MIAs. Our findings demonstrate the dual benefits of client selection, which simultaneously improves model performance and reduces MIAs’ vulnerabilities. By emphasizing these benefits, this study provides actionable guidance for developing privacy preserving FL frameworks that protect sensitive information without compromising model effectiveness.