Reproduce “A Conditional Tabular GAN-Enhanced Intrusion Detection System for Rare Attacks in IoT Networks” on OpenResearch

About this paper

Safaa Menssouri, El Mehdi Amhoud

Researchers at Mohammed VI Polytechnic University introduce CTGSM-DNN, a two-stage intrusion detection system that leverages Conditional Tabular GAN (CTGAN) for generating synthetic data for rare attack classes and SMOTEENN for subsequent dataset balancing and refinement. The system achieved 99.90% overall accuracy in multiclass classification and up to 80% detection accuracy for rare attacks like SQL Injection and Brute Force-XSS on the CSE-CIC-IDS2018 dataset.

Key insights

• CTGAN effectively handles the mixed data types and complex dependencies of tabular network data, enabling the generation of high-quality synthetic samples for rare attack classes.
• The hybrid approach of combining CTGAN for initial data augmentation with SMOTEENN for subsequent cleaning and balancing creates a robust and representative dataset for training.
• Employing Focal Loss with the DNN classifier significantly enhances the model's ability to focus on and accurately classify minority attack classes, overcoming the limitations of standard loss functions.

Summary by alphaXiv.