Topology‐Aware Spectral Graph Learning for Unsupervised IoT Device Clustering and Anomaly Detection

The heterogeneity and high level of dynamism in the topology and specificity of nodes in IoT networks pose challenges for clustering devices and detecting anomalies. Machine learning approaches in their current form also tend to overlook the graph principles of IoT systems. This chapter addresses these limitations by developing SGCN, a spectral graph convolutional approach specifically designed for IoT networks. The contributions include a spectral formulation for heterogeneous IoT topologies, an unsupervised clustering method using learned graph embeddings, an anomaly detection approach based on spectral analysis and comprehensive experimental results demonstrating superior performance compared to existing methods. Future research directions may include extending spectral methods to dynamic and temporal networks, improving weight optimization and refining filtering mechanisms. The proposed framework provides a solid foundation for IoT security monitoring and the management of complex network environments.