Performance evaluation of deep learning models for intrusion detection using network traffic
TL;DR
This study examines a one-dimensional Convolutional Neural Network and a hybrid model, investigating how both architectures can detect network attacks in binary and multiclass classification settings, and provides actionable insights for practitioners choosing between deep learning and classical approaches under real-world NIDS deployment constraints.
Abstract
Digital infrastructures are growing rapidly and cyberattacks are becoming increasingly sophisticated. As a result, Network-based Intrusion Detection Systems have become a critical component of modern cybersecurity. Signature-based methods sometimes fail to detect new threats, where machine learning and deep learning can provide a valuable alternative. This study examines a one-dimensional Convolutional Neural Network and a hybrid model called CNN-BiLSTM, investigating how both architectures can detect network attacks in binary and multiclass classification settings. Both architectures were evaluated on the CIC-IDS2017 and UNSW-NB15 benchmark datasets through a rigorous experimental protocol incorporating stratified 5-fold cross-validation, Focal Loss, QuantileTransformer scaling, and SMOTE oversampling. Random Forest and XGBoost were additionally tested on UNSW-NB15 as classical baselines. The results show that CNN achieves near-perfect binary classification performance on CIC-IDS2017, with 99.75% accuracy and a ROC-AUC of 0.9999. CNN-BiLSTM reduces false negatives by approximately 36% in the binary UNSW-NB15 setting, albeit at the cost of 4–6 times higher inference latency. In multiclass classification, CNN outperforms CNN-BiLSTM on CIC-IDS2017 with a higher macro F1-Score (0.701 vs. 0.671). XGBoost achieves comparable F1 performance to deep learning models in significantly less training time, once again demonstrating the enduring value of ensemble methods. Minority attack class detection remains a fundamental unresolved challenge across all configurations, laying the groundwork for future research on class-wise threshold calibration and Transformer-based architectures. The findings provide actionable insights for practitioners choosing between deep learning and classical approaches under real-world NIDS deployment constraints.