Scholarlink

We propose a robust federated learning system for classifying electrocardiograms (ECGs) that protects data privacy and is resistant to label-flipping and gradient-scaling attacks. The system uses logistic regression models trained on distributed 187-dimensional ECG time-series data. We employ a lightweight, confidence-based defense mechanism that evaluates client reliability based on prediction agreement with peers, enabling detection and exclusion of malicious clients without compromising model accuracy. Our dynamic weighting strategy uses prediction consensus with a threshold of 0.15 to gauge client trustworthiness. We used PTB Diagnostic ECG Database for the framework. Empirical testing on the PTB dataset demonstrates robustness, maintaining 92% accuracy under attack compared to 97% in benign scenarios. Our framework also features federated feature aggregation, which enhances classification stability by combining representative features from diverse clients without exposing raw data, thereby supporting model generalizability across heterogeneous medical environments. Our defense system mitigates adversarial attacks by 60% compared to an undefended FL system, with computational overhead under 15%. These findings demonstrate the framework’s applicability in privacy-preserving medical scenarios where security and data integrity are paramount, providing a foundation 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

A Secure Federated Learning Framework for Decentralized ECG Classification on Heterogeneous Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Prakhar Shukla,
  • Vaibhav Shukla,
  • Om Dabral,
  • Mounil Hiren Kankhara,
  • Bharat Singh,
  • Bagesh Kumar

摘要

We propose a robust federated learning system for classifying electrocardiograms (ECGs) that protects data privacy and is resistant to label-flipping and gradient-scaling attacks. The system uses logistic regression models trained on distributed 187-dimensional ECG time-series data. We employ a lightweight, confidence-based defense mechanism that evaluates client reliability based on prediction agreement with peers, enabling detection and exclusion of malicious clients without compromising model accuracy. Our dynamic weighting strategy uses prediction consensus with a threshold of 0.15 to gauge client trustworthiness. We used PTB Diagnostic ECG Database for the framework. Empirical testing on the PTB dataset demonstrates robustness, maintaining 92% accuracy under attack compared to 97% in benign scenarios. Our framework also features federated feature aggregation, which enhances classification stability by combining representative features from diverse clients without exposing raw data, thereby supporting model generalizability across heterogeneous medical environments. Our defense system mitigates adversarial attacks by 60% compared to an undefended FL system, with computational overhead under 15%. These findings demonstrate the framework’s applicability in privacy-preserving medical scenarios where security and data integrity are paramount, providing a foundation for secure FL deployment in healthcare settings.