| Sumario: | Wireless Sensor Networks (WSNs) have become a core technology for modern applications, including surveillance, smart agriculture, and Internet of Things (IoTs) ecosystems, but they remain highly vulnerable to malicious nodes that compromise link reliability and communication performance. Conventional anomaly detection and cryptographic approaches are often inadequate due to the resource limitations of sensor nodes and the dynamic nature of network topologies. This study aims to design a secure and robust detection framework that accurately identifies malicious nodes while maintaining link stability. The paper proposes a hybrid model integrating federated Graph Convolutional Networks (GCN) for spatial learning and Long ShortTerm Memory (LSTM) networks for temporal sequence analysis, combined with Q learning for self-healing routing. The methodology uses the WSN BFSF dataset with multi-domain feature extraction across traffic, energy, topology, and storage, class balancing with SMOTE, and feature selection through mutual information. Experimental validation using NS3 simulations demonstrates a classification accuracy of 99.70 percent with precision, recall, and F1 scores exceeding 0.996 for all classes. Link stability also improved significantly, with packet delivery ratio rising from 0.57 to 1.0 and average delay reduced from 3.32 to 3.09 units. The findings confirm that the proposed framework enhances both the detection of malicious activity and the resilience of wireless sensor networks in real-world scenarios.
|
|---|