Machine Learning Based Framework for Efficiently Predicting Security Attacks in WSN

Wireless Sensor Networks (WSNs) are very common in critical applications yet they are very susceptible to different security attacks. The current intrusion detection methods primarily focus on accuracy improvement, whereas issues like computational complexity, prediction time, model size, and price are frequently neglected, and therefore a wide range of solutions is inapplicable in real-time, resource-limited WSN conditions. A number of more recent papers that use a combination of ensemble and deep learning techniques claim good detection rates, though their high complexity and energy requirements make them impractical. To overcome these shortcomings, the current paper introduces a lightweight Decision Tree based Decision Forest model to predict and classify attacks in WSNs with the benchmark WSN-BFSF dataset. The framework is also evaluated with several machine learning models such as the Random Forest, SVM, Naive Bayes and Logistic Regression models which were evaluated with evaluation metrics such as accuracy, precision, recall, F1-score, prediction cost, training time, model size and prediction speed. It has been experimentally proven that the proposed model has the accuracy of 97.5% and faster prediction with less training time, less model size, and cost of misclassifications, which indicates an effective trade-off between detection performance and efficiency of resources.