Latency-Aware Comparative Evaluation of Machine Learning Classifiers for Network Intrusion Detection Using the LAAI Metric on KDD99

The increasing sophistication of cyber threats to networked infrastructure has increased the need for accurate and effective network intrusion detection systems (NIDS) (NIDS). Although machine learning (ML) methods have performed exceptionally well in classification with canonical intrusion detection datasets, existing comparative literature does not consider inference latency as an important evaluation criterion, providing model recommendations that are correct but computationally infeasible to apply in practice. This paper presents a top-down, reproducible analysis of twelve machine learning classifiers (linear, probabilistic, tree-based, ensemble, and neural paradigms) on the KDD Cup 1999 (KDD99) benchmark using the Latency-Adjusted Accuracy Index (LAAI) as the main ranking tool to combine predictive accuracy and computational efficiency. All models use an identical preprocessing pipeline and identical hardware timing protocol to ensure consistency in the methodology. Results show that CatBoost achieves the highest LAAI score of 0.9946 (accuracy 99.94), then Decision Tree (LAAI 0.9991), and XGBoost (LAAI 0.9852). More importantly, K-Nearest Neighbours (KNN) with a test accuracy of 99.87% comes with the lowest LAAI of 0.3864 with extremely high inference latency of 1.585 ms - an accuracy-latency paradox which is not captured by traditional evaluation metrics. In the Tier I cluster, the mean accuracy is 99.71% (SD = 0.61%) while the mean LAAI is 0.9857 (SD = 0.0094). In contrast, the mean accuracy across the whole dataset is 92.22% (SD = 14.81%) while the mean LAAI is 0.8497 (SD = 0.1997), which highlights the significant performance disparity between top and bottom tier classifiers. This suggests a four-level deployment classification to guide NIDS model selection for practitioners. The LAAI-based rankings are demonstrated to be generally applicable with the help of cross-validation against six independent benchmark studies.