Predicting DoS-Probe-R2L-U2R Intrusions in Wireless Sensor Networks Using an Ensemble Deep Learning Model

Wireless Sensor Networks (WSNs) have helped to make significant advances in various sectors, but their increasing adoption has generated security issues. Network Intrusion Detection Systems (NIDS) are essential for safeguarding Wireless Sensor Networks. Previous studies looked at threat detection using machine learning and deep learning approaches, but there was no precise feature selection method for threat classification. As a result, this study developed a Network Intrusion Detection (NID) Model for WSNs by selecting essential features using Recursive Feature Elimination (RFE) and Random Forest (RF) as the estimator and predicting threats using a Deep Learning Ensemble (DLE) method. The NSL-KDD dataset was utilised to establish the experimental setup. 123 features were reduced to 16 through Recursive Feature Elimination and Random Forest estimator. Common threats to WSNs include Denial of Service (DoS), Probe, Remote-to-Local (R2L), and User-to-Root (U2R). Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Deep Belief Networks (DBNs) from deep learning were trained as the base models on the WSN dataset. A Deep Learning Ensemble (DLE) was developed by integrating the base models’ predictions using a voting classifier, and its performance was evaluated and compared to similar models using Accuracy, Precision, recall, and F1 metrics. The performance evaluation results on RFE selected features showed that the CNN model, with an average accuracy score of 96.1%, outperforms RNN and DBN with average accuracy scores of 95.6% and 88.2%, respectively. The developed DLE model had performance score values of 99.6%, 98.9%, 98.4%, and 98.7% for accuracy, precision, recall, and F1-score, respectively. The result of the evaluation of the developed DLE model was benchmarked with the Hybrid Model (CNN & LSTM), RNN (LSTM), and DNN with accuracy values of 97.7%, 92.8%, and 95.6%, respectively. The developed DLE model outperformed the benchmarked models. The developed DLE model, suggested by this research, can be employed with various optimisation methods for an extensive exploration of significant datasets and dimensionality reduction to boost classification accuracy.