Optimizing Intrusion Detection with Hybrid Deep Learning Models and Data Balancing Techniques
Md. Nasif Sarwar, Md. Shohel Arman, Touhid Bhuiyan, Fatama Binta Rafiq · 2025
The effectiveness of hybrid deep learning models in detecting network intrusions on imbalanced datasets was tested. Conventional IDS often misses rare attacks due to class imbalance. Three models were evaluated: CNN+DNN, Autoencoder+DNN, and DNN+XGBoost, using the CSC IDS 2018 dataset balanced with SMOTE, FSMOTE, and CTGAN techniques. Feature selection was performed using SHAP and correlation studies. The models were evaluated on accuracy, precision, recall, F1-score, and ROC AUC. The Autoencoder+DNN and CNN+DNN models, especially with FSMOTE and SMOTE, showed the best performance. This research highlights the potential of combining advanced feature selection, data balancing, and deep learning techniques to enhance network intrusion detection, offering a robust framework for improved cybersecurity strategies.