Improved lightweight cloud-based multiclass cyber intrusion detection: optimising FastViT with Boruta-SHAP and class balancing on CSE-CIC-IDS2018

Cloud environments have recently become increasingly
exposed to the growing complexity and volume of network traffic,
thereby elevating the demand for robust, scalable, and intelligent
intrusion detection systems (IDS). This study introduces an
effective deep learning-based IDS that integrates advanced feature
selection with a lightweight, high-performing deep learning model,
FastViT, to address critical limitations observed in prior work,
particularly in handling multiclass classification and severe class
imbalance. Many previous studies evaluate model performance on
a limited subset of attack types or treat each attack in isolation,
often neglecting dataset balancing techniques. This leads to poor
generalisation and reduced effectiveness in detecting complex and
infrequent attacks under realistic conditions. To address these
limitations, this study proposes a FastViT-based approach, a
hybrid convolutional-transformer architecture, trained and
evaluated on the CSE-CIC-IDS2018 dataset, which includes all
attack categories. A hybrid feature selection strategy combining
Boruta and SHAP is employed to enhance interpretability and
improve learning efficiency by identifying the most relevant
features. Dataset balancing is achieved through a combination of
the Synthetic Minority Over-sampling Technique (SMOTE) and
undersampling, ensuring fair representation of minority classes.
Comparative experiments on both balanced and imbalanced
datasets demonstrate significant improvements, particularly in
recall, for underrepresented and challenging attack types such as
Web and Infiltration.