Machine Learning and Data Balancing Methods for Intrusion Detection Enhancement

Abstract

The continuous advancements of the technologies has lead also to the use of Wireless Sensor Networks (WSNs). WSNs have been extensively utilized in various surveillance applications, including industrial process monitoring, consumer applications, military operations, machine health monitoring, emergency relief efforts, and agricultural monitoring. However, the deployment of WSN faces several challenges such as the uneven sensor dispersion from haphazard deployment may have detrimental effects. Researchers have proposed a number of approaches, including intrusion detection systems, routing protocols, cryptography, and key management. Therefore, finding affordable, reliable, and effective solutions becomes essential. This work aids real-time intrusion detection in resource-constrained WSNs. In this article, we have evaluated three Machine Learning (ML) Models including SVM, Random Forest and Extreme Gradient Boosting (XGBoost) using WSN-DS dataset. The evaluation metrics used to evaluate the performance of the hybrid models (Combination of balancing techniques with ML methods) are the training time (in Seconds), the accuracy, and the Area Under the Curve (AUC) to validate the performance of the ML algorithms. This experiment examines the effect of data balancing techniques to the accuracy rate of classification using a comparative study. We have compared the accuracy rate of the chosen models according to Unbalanced (Original Distribution), SMOTE, Random Oversampling, Random Undersampling, ADASYN, and SMOTE + ENN methods. As a results, the hybrid model, Smote + ENN and XGBoost model, performed the highest scores among the other algorithms by an AUC of 99.99%, accuracy of 99.80%, and 6452 seconds as training time.