Machine Learning-Based Intrusion Detection System for Secure Communication Networks

Abstract

Secure communication networks must distinguish legitimate traffic from rapidly changing malicious activity without introducing excessive false alarms or processing delay. This methodology paper presents a machine-learning-based network intrusion detection system (NIDS) designed around flow-level traffic analysis, reproducible preprocessing, feature selection, supervised classification, and security-oriented performance evaluation. The framework uses modern benchmark traffic, with UNSW-NB15 as the primary dataset and CICIDS2017 as an external robustness reference, because both contain diverse benign and attack behaviors that are more representative than legacy KDD-only evaluations [1], [2], [7]. The proposed study compares a linear baseline, support vector machine, random forest, and multilayer perceptron under identical train/validation/test rules. Model selection is based on macro-F1, recall, false-alarm rate, ROC-AUC, and inference latency rather than accuracy alone. A literature-based validation of the methodology shows that learning-based IDSs can achieve strong detection performance, but results are highly sensitive to dataset age, class imbalance, feature engineering, and evaluation protocol [5], [6], [9]–[12]. The paper therefore emphasizes leakage-free preprocessing, held-out testing, class-aware metrics, and deployability for secure communication networks.