Abstract

The modern community increasingly depends more than ever on a variety of platforms for data exchange and communication due to the ongoing rapid evolution and development of the internet and computing technologies. These frameworks offer autonomy and accuracy but tend to lack advanced security and safety measures, consisting of valuable data, which cybercriminals have taken new ways to take advantage of by developing potential techniques for compromising security and breaching data. By using malicious software, viruses, jamming attacks, and injection attacks, etc., in several IoT and WSN frameworks in use by both residential and commercial sectors, cybercriminals are making it a major concern for stakeholders and potential users. Based on these issues, the proposed methodology is designed using advanced Machine Learning techniques for detecting jamming attacks within a WSN. The methodology uses 2 distinct datasets taken from Kaggle, first the “CICIot2023 dataset” that consists of 6 classes associated with DDoS attacks, and the “DOS Mosaic dataset” which consists of 3 classes associated with DoS attacks. An extensive exploratory data analysis (EDA) technique is applied to both datasets to extract data limitations, which were later improved by applying data preprocessing and feature extraction techniques for data enhancement and feature selection. For model training, the proposed model RFMetaClassifier is presented, which is an ensemble stacked model developed using Multilayer perceptron, Random Forest, and Logistic Regression within the base layer and Random Forest model as the Meta Layer. Furthermore, 6 ML classifiers, including Decision Tree, Random Forest, XGboost, Logistic Regression, Naive Bayes, and Catboost, along with 3 Hybrid models CNN+GRU, CNN+Random Forest, and CNN+Decision Tree, are applied on both of the datasets. The aim was to implement a series of different models alongside the proposed RFMetaClassifier for a more detailed performance comparison. Out of these implemented algorithms, RFMetaClassifier produced the best results on both datasets based on selected performance evaluation metrics. The proposed model produced 100% of training and 100% of validation accuracy on the ‘DoS Mosaic Dataset’ as well as 100% and 99.98% of training and validation accuracy, respectively, on the “CICIoT2023 dataset”. The proposed methodology surpasses the ML and hybrid techniques utilized within the existing related work by producing the highest accuracy among all. This work can substantially prove to be very helpful in the detection and classification of jamming attacks in various industrial sectors, as well as for domestic use.