Deep hybrid architecture for multi-class detection of network layer attacks in WSN

The growth of Wireless Sensor Networks (WSNs) in essential fields such as health care, defense, and environmental monitoring poses severe cybersecurity threats, specifically to the network layer. Network layer attacks such as Blackhole, Flooding, and Selective Forwarding are common and tend to have a serious impact on the data integrity, the reliability, and the lifetime of the network. To counter these vulnerabilities, this study proposes the development of a hybrid deep learning (DL) model combining the Gated Recurrent Units (GRU) and Long Short-Term Memory (LSTM) structures for the detection of multi-class WSNs attacks. The proposed model combines the temporal learning capability of GRU with the long-term dependency learning strength of LSTM to accurately detect malicious attacks. Using WSNBFSF dataset, different types of experiments have been conducted to evaluate the performance of the developed model in accuracy, precision, recall, and F1-score metric measures. Experimental results indicate that the GRU+LSTM hybrid model offers improved classification performance when compared to traditional DL models, i.e., Deep Neural Network (DNN), Recurrent Neural Networks (RNN), Artificial Neural Network (ANN), Attention, LSTM, and GRU models for multi-class classification tasks. The model achieved an overall accuracy of 97.41% (without SMOTE) and 98.74% using a Synthetic Minority Oversampling Technique (SMOTE) technique and k-fold validation along with macro F1 and weighted F1 of 0.9874 across multiple attack classes. These results indicate that the proposed architecture effectively captures complex sequential attack patterns and provides a reliable solution for intelligent intrusion detection in WSN environments.