Safeguarding federated learning models in intelligent internet of medical things against data poisoning attacks

Abstract

Federated learning (FL) has gained wide attention as a way to train machine learning models without sharing sensitive raw data, which is essential in the Internet of Medical Things (IoMT). By enabling collaboration across distributed healthcare devices, FL supports the development of accurate predictive models while preserving patient privacy. However, its adoption in IoMT also raises new risks, particularly data poisoning attacks that can degrade model performance and undermine trust in medical applications. This paper provides a comprehensive analysis of FL in the IoMT domain, covering its key applications, challenges, and emerging research directions. In addition, we introduce a data poisoning mitigation algorithm that combines outlier detection with robust aggregation. Outlier detection is used to identify suspicious updates based on statistical deviation thresholds. At the same time, robust aggregation techniques such as Trimmed Mean and Krum ensure that only reliable contributions influence the global model. This two-level defense strengthens the resilience of FL against both extreme and subtle poisoning strategies. Experimental results show that the proposed approach achieves a post-attack accuracy of 87% and lowers the attack success rate to 17%. By addressing both opportunities and security challenges, this work advances the role of FL in secure and privacy-preserving IoMT systems.