Dataset for Intrusion Detection Systems (IDS) are often 'imbalanced', where normal data conditions are too abundant while attack data is very scarce. This condition makes machine learning performance in detecting attacks poor, especially for attacks that are indeed rare. Previous research addressed this issue by applying Stacked Denoising Autoencoder (SDAE), which can reduce the dimensions of features in large datasets. In this study, researchers explored the potential use of SDAE but combined it with Generative Adversarial Networks (GAN). SDAE was adopted not for dimension reduction, but with a different approach to augmentation or multiplying minority attack class data. With the SDAE-GAN hybrid model, the researchers proposed an architecture using an SDAE-based generator trained in a GAN framework to generate realistic synthetic data for each minority attack class in the UNSW-NB15 dataset. This process balanced the data distribution before training the classification model. The researchers then tested the model using Support Vector Machine (SVM) on the balanced data. The results were very significant, with the SVM model achieving 97% accuracy, an F1-Score of 98.44%, and, most importantly, a near-perfect recall value of 99.67% in detecting attack classes. This shows that the SDAE-GAN hybrid augmentation method can overcome the imbalance problem and drastically improve the IDS's ability to detect attacks from minority classes.
Copyrights © 2026