Poisoning resilient federated learning for secure internet of medical things a systematic review

Abstract

The current trend of the Internet of Medical Things (IoMT) has facilitated intelligent healthcare, including real-time monitoring, remote diagnosis, and data-driven clinical decision support. This substantially increases the attack surfaces of medical cyber–physical systems. As a privacy-conserving paradigm, Federated Learning (FL) has been proposed to enable distributed IoMT devices to collaboratively train models without exchanging raw medical data. However, FL-based IoMT systems are highly susceptible to poisoning attacks, intrusion-based attacks, gradient leakage, and communication-level adversaries, especially in heterogeneous and resource-constrained setups. This paper presents a systematic and comprehensive review of poisoning-resilient FL for securing IoMT systems, which summarizes over 134 peer-reviewed articles published since 2020. Using PRISMA 2020 principles, we construct a unified, multidimensional taxonomy of data poisoning, model poisoning, backdoor attacks, Byzantine attacks, Sybil attacks, intrusion-based attacks, and communication-level attacks in medical FL deployments. In contrast to existing fragmented surveys, this work directly correlates attack types with cross-layer defense mechanisms, including robust aggregation, anomaly-based intrusion detection, differential privacy, cryptography, blockchain-enhanced FL, and lightweight hybrid strategies tailored for low-power IoMT devices. A comparative evaluation is conducted using real-world healthcare datasets, analyzing the trade-offs among detection accuracy, latency, communication overhead, computational cost, and energy efficiency. Finally, the research challenges and future research directions are identified to enable regulatory compliance, real-time clinical applications, and scalable edge-fog-cloud orchestration for next-generation secure IoMT systems.