Explainable Analysis of ANN-Based Intrusion Detection Classifiers: A Comparative Evaluation on ToN-IoT and RT-IoT 2022 Datasets

Abstract

Internet of Things (IoT) has opened up many innovative applications but the unique characteristics of IoT devices have made it vulnerable to cyber-attacks. To deal with these threats, robust mechanisms are to be built to ensure safe and reliable IoT ecosystems. In this work, an Artificial Neural Network (ANN) based approach has been investigated to build intrusion detection models that are capable of analyzing anomalous behavior of network users and flag deviations from normal network activities. Four ANN based classification techniques, viz., Radial Basis Function Network (RBFN), Self-Organizing Map (SOM), Deep Neural Network (DNN), and Multi-pass Learning Vector Quantization (MLVQ) have been used to classify communication data in an IoT network. In order to enhance the model performance, feature engineering has been applied using five feature selection techniques namely, information gain, gain ratio, symmetrical uncertainty, Relief-F, and chi-square. Results show that the DNN classifier consistently achieved the highest detection accuracy across all feature-selection subsets (up to 99.78%), while Relief-F produced the strongest overall feature subset across classifiers in ToN_IoT dataset. DNN classifier consistently achieved the highest detection accuracy across all feature-selection subsets (up to 99.52%), while Symmetric Uncertainty produced the strongest overall feature subset across classifiers of RT_IoT2022 dataset.