Scholarlink

<p>Federated learning (FL) has emerged as a promising paradigm for privacy-preserving analytics in distributed healthcare systems, yet its deployment for real-time cybersecurity in mobile health (mHealth) environments remains insufficiently understood. This study presents a domain-specific federated threat detection framework that integrates differential privacy (DP) and homomorphic encryption (HE) within a unified FL pipeline for Internet-of-Medical-Things (IoMT) ecosystems. This study presents a domain-specific federated learning (FL) framework for privacy-preserving cyber threat detection in distributed mobile healthcare environments. The proposed framework integrates differential privacy (DP) and homomorphic encryption (HE) into the federated aggregation process, maintaining secure model training without exposing sensitive patient data or electronic health records (EHRs). Four neural network architectural models, such as Dense, a convolutional neural network (CNN), long short-term memory (LSTM), and Gated Recurrent Unit (GRU), were systematically evaluated under centralized and federated configurations, including hybrid methods, such as federated averaging (FedAvg), FedAvg + DP, FedAvg + HE, and FedAvg + DP+HE, which were deployed for an experiment. Additionally, experimental findings show that centralized models achieve near-perfect detection accuracy of &gt; 0.98 (98%); however, they lack data protection, although federated models exhibit marginal performance trade-offs of approx. 0.94 (95%) and 0.96 (96%) with significantly improved privacy guarantees. The incorporation of DP explores measurable reductions in precision and recall as HE introduces computational overhead with minimal accuracy loss. A combination of DP and HE configurations shows strong security guarantees, though at the cost of higher latency and complexity. From the evaluated architectures, LSTM and GRU under FedAvg + HE performed the most balanced trade-off between detection performance, privacy protection, and system scalability. Contrasting with the prior studies that investigate privacy-preserving FL mechanisms in isolation, this work provides a comprehensive system-level evaluation of differential privacy and homomorphic encryption within a unified mHealth cybersecurity context. The study emphasizes practical deployment considerations, including communication overhead, latency, model robustness, and heterogeneous edge environments, thereby offering actionable understandings for real-world healthcare systems. The findings further reveal the transformative potential of federated privacy-preserving learning in protecting the desired advanced healthcare ecosystems and enable a robust foundation for future research in reliable distributed cyber intelligence and defense systems.</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

Advancing Federated Learning Frameworks for Privacy-preserving Cyber Threat Detection in Healthcare Systems

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月30日

查看全文
  • Anayo Chukwu Ikegwu,
  • Uzoma Rita Alo,
  • Henry Friday Nweke,
  • Deborah Uzoamaka Ebem

摘要

Federated learning (FL) has emerged as a promising paradigm for privacy-preserving analytics in distributed healthcare systems, yet its deployment for real-time cybersecurity in mobile health (mHealth) environments remains insufficiently understood. This study presents a domain-specific federated threat detection framework that integrates differential privacy (DP) and homomorphic encryption (HE) within a unified FL pipeline for Internet-of-Medical-Things (IoMT) ecosystems. This study presents a domain-specific federated learning (FL) framework for privacy-preserving cyber threat detection in distributed mobile healthcare environments. The proposed framework integrates differential privacy (DP) and homomorphic encryption (HE) into the federated aggregation process, maintaining secure model training without exposing sensitive patient data or electronic health records (EHRs). Four neural network architectural models, such as Dense, a convolutional neural network (CNN), long short-term memory (LSTM), and Gated Recurrent Unit (GRU), were systematically evaluated under centralized and federated configurations, including hybrid methods, such as federated averaging (FedAvg), FedAvg + DP, FedAvg + HE, and FedAvg + DP+HE, which were deployed for an experiment. Additionally, experimental findings show that centralized models achieve near-perfect detection accuracy of > 0.98 (98%); however, they lack data protection, although federated models exhibit marginal performance trade-offs of approx. 0.94 (95%) and 0.96 (96%) with significantly improved privacy guarantees. The incorporation of DP explores measurable reductions in precision and recall as HE introduces computational overhead with minimal accuracy loss. A combination of DP and HE configurations shows strong security guarantees, though at the cost of higher latency and complexity. From the evaluated architectures, LSTM and GRU under FedAvg + HE performed the most balanced trade-off between detection performance, privacy protection, and system scalability. Contrasting with the prior studies that investigate privacy-preserving FL mechanisms in isolation, this work provides a comprehensive system-level evaluation of differential privacy and homomorphic encryption within a unified mHealth cybersecurity context. The study emphasizes practical deployment considerations, including communication overhead, latency, model robustness, and heterogeneous edge environments, thereby offering actionable understandings for real-world healthcare systems. The findings further reveal the transformative potential of federated privacy-preserving learning in protecting the desired advanced healthcare ecosystems and enable a robust foundation for future research in reliable distributed cyber intelligence and defense systems.