Scholarlink

Federated Learning (FL) has become a key paradigm for privacy-preserving training of Deep Learning (DL) models in medical image classification, allowing hospitals and clinical institutions to collaboratively build models without directly sharing sensitive patient data. While FL ensures data locality, recent findings reveal that gradient-sharing still poses a privacy risk, as malicious adversaries can potentially reconstruct original medical images from shared gradients. This presents a critical concern in clinical applications where patient confidentiality is paramount. In this work, we propose a robust defense mechanism based on a private key-lock module. Our approach secures arbitrary model architectures by transforming gradients through a key-dependent layer before sharing. Only these locked gradients are sent to the server for global aggregation, ensuring that image reconstruction is infeasible without access to the private key. We provide theoretical justification for our method’s security and demonstrate its effectiveness through extensive experiments across multiple medical image classification benchmarks. The results show that our key-lock mechanism significantly reduces privacy leakage while maintaining strong diagnostic performance, offering a practical and efficient defense for secure FL deployment in healthcare settings.

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

Safeguarding Privacy for Medical Data with a Novel Key-Lock Module in Federated Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Hanchi Ren,
  • Jingjing Deng,
  • Xianghua Xie,
  • Xiaoke Ma,
  • Jianfeng Ma

摘要

Federated Learning (FL) has become a key paradigm for privacy-preserving training of Deep Learning (DL) models in medical image classification, allowing hospitals and clinical institutions to collaboratively build models without directly sharing sensitive patient data. While FL ensures data locality, recent findings reveal that gradient-sharing still poses a privacy risk, as malicious adversaries can potentially reconstruct original medical images from shared gradients. This presents a critical concern in clinical applications where patient confidentiality is paramount. In this work, we propose a robust defense mechanism based on a private key-lock module. Our approach secures arbitrary model architectures by transforming gradients through a key-dependent layer before sharing. Only these locked gradients are sent to the server for global aggregation, ensuring that image reconstruction is infeasible without access to the private key. We provide theoretical justification for our method’s security and demonstrate its effectiveness through extensive experiments across multiple medical image classification benchmarks. The results show that our key-lock mechanism significantly reduces privacy leakage while maintaining strong diagnostic performance, offering a practical and efficient defense for secure FL deployment in healthcare settings.