<![CDATA[Distributed Denial of Service (DDoS) attacks generate high-dimensional network traffic that poses significant challenges for machine learning-based detection systems in terms of predictive accuracy and computational efficiency. This study presents a systematic evaluation of Mutual Information (MI) based feature selection applied to three modern boosting algorithms, namely XGBoost, LightGBM, and CatBoost, using the CIC-DDoS2019 dataset. A controlled experimental design was employed, where data partitioning was performed prior to resampling, and SMOTE was applied exclusively to the training set to prevent data leakage. Feature selection was conducted by identifying the top 25 features based on MI score saturation analysis. The results demonstrate that MI-based feature selection consistently improves classification performance while substantially reducing training time across all models. Among the evaluated methods, LightGBM achieves the best trade-off between accuracy and computational efficiency, reaching an accuracy of 99.88% with significantly reduced training cost. These findings indicate that feature quality plays a critical role in shaping the learning behaviour of boosting algorithms and that MI-based feature selection functions as a structural mechanism for enhancing model stability and scalability in high-dimensional DDoS detection scenarios.]]>
Copyrights © 2026
