Evaluating the Generalization Gaps of Intrusion Detection Systems Across DoS Attack Variants

Intrusion Detection Systems (IDS) play a vital role in safeguarding networks, yet their effectiveness is often challenged, as cyberattacks evolve in new and unexpected ways. Machine learning models, although very powerful, usually perform well only on data that closely resembles what they were trained on. When faced with unfamiliar traffic, they often misclassify. In this work, we examine this generalization gap by training IDS models on one Denial-of-Service (DoS) variant, DoS Hulk, and testing them against other variants such as Goldeneye, Slowloris, and Slowhttptest. Our approach combines careful preprocessing, dimensionality reduction with Principal Component Analysis (PCA), and model training using Random Forests and Deep Neural Networks. To better understand model behavior, we tuned decision thresholds beyond the default 0.5 and found that small adjustments can significantly affect results. We also applied Shapley Additive Explanations (SHAP) to shed light on which features the models rely on, revealing a tendency to focus on fixed components that do not generalize well. Finally, using Uniform Manifold Approximation and Projection (UMAP), we visualized feature distributions and observed overlaps between training and testing datasets, but these did not translate into improved detection performance. Our findings highlight an important lesson: visual or apparent similarity between datasets does not guarantee generalization, and building robust IDS requires exposure to diverse attack patterns during training.