AutoKAN: A Federated Lightweight Anomaly Detection Framework for Securing Constrained IoT Healthcare Diabetes Monitoring Systems
Nguyen Khanh Son, Arun Kumar Sangaiah, Chuang Chun-Chi, Honda Hsu, Chung-Chian Hsu, Chuan‐Yu Chang · IEEE Transactions on Consumer Electronics · 2025
The adoption of Internet of Things (IoT) technology in healthcare has significantly enhanced patient care by improving both the efficiency and cost-effectiveness of healthcare delivery systems. Specifically, for diabetic patients, IoT facilitates continuous health monitoring, enabling healthcare professionals to track patients’ conditions in real time and detect potential complications at an early stage. However, the growing number of IoT devices has also introduced security concerns, particularly in terms of cyber threats that exploit resource limitations such as power and memory. Especially in the healthcare sector, where data is highly sensitive and distributed in nature, privacy is one of the most critical aspects. A promising privacy-preserving technology, known as federated learning, addresses these challenges effectively. This research presents a lightweight federated anomaly detection framework tailored for constrained environments, including small-scale IoT devices. We conceptualize IoT devices as essential components within the edge continuum and leverage anomaly detection models to enhance security. Our anomaly detection model is built on an autoencoder architecture; however, rather than relying on conventional multilayer perceptron (MLP) networks, we utilize Kolmogorov–Arnold Networks (KAN) with an adaptive threshold to minimize parameter complexity and enable real-time deployment. Additionally, the federated learning mechanism is applied to ensure the privacy of patient data is safeguarded. Experimental results indicate that our model achieves an accuracy exceeding 99.5% and a precision of 100%, outperforming traditional autoencoders while utilizing 50% fewer parameters and achieving twice the speed in training and inference time. This demonstrates the superiority of our approach in anomaly detection, particularly in fields where precision is the most important metric, such as medical applications.