Knowledge Improved Hybrid DNN–KAN Framework for Intrusion Detection in Wireless Sensor Networks

Wireless Sensor Networks (WSNs) are increasingly vulnerable to sophisticated cyber threats, necessitating advanced intrusion detection systems (IDS) that balance high accuracy with interpretability. This paper presents a Knowledge-Improved Hybrid Deep Neural Network-Kolmogorov Arnold Network (DNN-KAN) Framework for intrusion detection in WSNs, integrating data-driven learning with domain-specific knowledge to enhance detection performance. The proposed framework preprocesses and merges multiple datasets (WSN, NSL-KDD, and CICIDS2017), extracts features using Principal Component Analysis (PCA), and constructs a knowledge graph to embed expert-defined rules via Graph Convolutional Networks (GCNs). The hybrid architecture employs a DNN for hierarchical feature extraction and replaces the fully connected layer with a KAN layer, leveraging the Kolmogorov-Arnold Theorem to decompose complex relationships into interpretable univariate functions. Evaluated on key metrics, the model achieves 99.87% accuracy, 99.85% precision, and a 0.9985 ROC-AUC score, significantly outperforming traditional DNNs, standalone KANs, and state-of-the-art models like CNN-LSTM and Transformer-based IDS. By incorporating domain knowledge into the loss function, the framework reduces false positives and improves generalization, even with limited training data. This work demonstrates that combining knowledge-driven constraints with hybrid neural architectures can yield highly accurate, interpretable, and robust intrusion detection for WSN security.