Claim Missing Document
Check
Articles

Found 1 Documents
Search

Leakage-Aware Machine Learning for Imbalanced Network Intrusion Detection: A Comparative Study of Classical and Ensemble Classifiers Pandhu Pramarta; Pujiastuti Pujiastuti; Sri Mardiyati
Journal Digital Technology Trend Vol. 5 No. 2 (2026): December
Publisher : Yayasan Kawanad

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.56347/jdtt.v5i2.451

Abstract

Network intrusion detection requires reliable machine learning models capable of identifying malicious traffic while maintaining valid generalization performance. However, class imbalance and cross-split feature overlap may affect the reliability of model evaluation. This study proposes a leakage-aware evaluation framework for binary network intrusion detection using the UNSW-NB15 dataset. Logistic Regression, Linear Support Vector Machine, Random Forest, and class-weighted variants of the two linear classifiers were evaluated using official and leakage-controlled test sets. A cross-split audit identified 1,302 exact overlapping feature patterns affecting 8,541 test records (10.37%), which were excluded to construct a leakage-controlled test set of 73,791 records. All classifiers exhibited performance reductions after overlap control, indicating differences between conventional and stricter generalization estimates. Random Forest achieved the best leakage-controlled performance, with 86.23% accuracy, 88.11% F1-score, 85.88% Macro F1-score, 97.68% ROC-AUC, and 97.87% average precision. It also achieved a 98.37% attack detection rate with a 1.63% false negative rate. Class weighting improved the balanced performance of both linear classifiers. These findings demonstrate the importance of jointly considering class imbalance and train-test independence when evaluating machine-learning-based intrusion detection systems.