Impact Analysis of Classifier Training on the Performance of Machine Learning based Intrusion Detection Systems

With a rapid rise in emerging technologies and connected networks, ensuring security has become a top requirement. Cyber-attacks in networks are constantly evolving and becoming more challenging to secure. Intrusion Detection System (IDS) is essential in maintaining the security and integrity of such network traffic. However, with rapidly growing threats in the network security landscape, handling IDS databases is becoming difficult. Machine Learning (ML) model-based approaches effectively address these issues in the context of IDS. The performance of an IDS in real-time can significantly vary depending on the model type, classifiers trained, attack type, the amount of training and testing data, and feature selection, among many others. This paper tests the impact of classifier (Decision Tree, Random Forest, Multi-layer Perceptron) training on IDS and provides an analysis of how each classifier training model performs when tested with newer IoT Datasets that consider a wider range of network attacks. This paper also discusses a step-by-step model for training the individual classifiers mentioned above. It discusses the impact of training on the performance when considering different attack types, label (normal or attack) conditions, and data balance of normal-attack traffic.