<![CDATA[Cyberattack classification supports timely threat identification and network security monitoring. However, model effectiveness depends not only on the selected algorithm but also on whether the available features contain sufficient information to distinguish attack classes. This study evaluates the effect of feature selection on multiclass cyberattack classification and assesses the discriminative capability of a synthetic cybersecurity dataset. The dataset contains 40,000 records from three nearly balanced classes: distributed denial-of-service, malware, and intrusion. Preprocessing included removing high-cardinality and potentially post-detection attributes, handling missing security indicators, one-hot encoding categorical variables, and applying a stratified 80:20 training/testing split. We evaluated Random Forest, Logistic Regression, and XGBoost using all usable features and the ten highest-ranked features identified through Random Forest importance. We measured performance using accuracy, macro precision, macro recall, and macro F1-score, and included a stratified dummy classifier as a baseline. Logistic Regression using all features achieved the highest accuracy of 33.95%, while XGBoost using the Top-10 subset achieved the highest macro F1-score of 33.69%. The dummy baseline obtained 33.90% accuracy, showing that none of the models produced a practically meaningful improvement over chance-level classification. Feature reduction slightly improved Random Forest and XGBoost but reduced Logistic Regression performance. These findings indicate that the dataset features have limited discriminative capability for separating the three attack categories. Further research should validate the analysis using real-world traffic, alternative feature-selection methods, and cross-dataset experiments before interpreting model complexity and reported accuracy.]]>
