Year: 2025
Type: preprint
Abstract: Federated learning (FL) enables collaborative machine learning across distributed healthcare institutions while preserving patient data privacy, making it particularly valuable for healthcare IoT systems that underpin critical social infrastructure such as digital health services. However, its distributed nature introduces vulnerabilities to adversarial attacks that can compromise model accuracy and pose critical risks to patient safety and public trust in AI-enabled healthcare systems. These threats are further intensified by server-side uncertainty, where aggregation servers cannot reliably assess client trustworthiness or detect malicious activity in real time. This paper evaluates the adversarial robustness of FL in healthcare IoT systems under such realistic uncertainty conditions. We examine three key attack vectors-backdoor, label flipping, and gradient manipulation attacks-using data derived from the DiaTrend dataset, which comprises continuous glucose monitor (CGM) and insulin pump data from 54 patients with diabetes. For our experimental evaluation, we utilized data from 12 subjects: 10 subjects for federated model training and 2 held-out subjects for independent performance validation to ensure robust and generalizable results. Experimental findings reveal severe vulnerabilities in standard federated averaging: gradient manipulation attacks caused up to a 3,145% increase in mean squared error (MSE) and an average prediction deviation of 79.1 mg/dL-levels that represent potentially life-threatening errors in clinical contexts. We further assess Byzantine-robust aggregation methods (Krum, Trimmed Mean, and Median) as proactive, attack-agnostic defenses. Defense performance varies substantially: Median aggregation achieves 94.4% recovery against gradient manipulation attacks, Trimmed Mean nearly restores performance against label flipping, while Krum achieves only 4.2% mitigation against backdoor attacks. Our findings highlight that proactive Byzantine-robust aggregation is critical for the safe deployment of FL in healthcare IoT systems. The exceptional resilience of Median aggregation demonstrates its viability for secure federated CGM analytics in safety-critical applications.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7108324296