HYBRIDNET-IDS: AN INNOVATIVE CNN, BILSTM, AND TRANSFORMER FRAMEWORK UTILISING ATTENTION FOR IMPROVED NETWORK INTRUSION DETECTION

Abstract

The rapid spread of network-connected technologies and the growing complexity of attacks have shown that traditional signature-based intrusion detection systems are not good enough, especially when it comes to finding zero-day attacks and advanced persistent threats. This paper introduces HybridNet-IDS, an augmented deep learning framework that integrates Convolutional Neural Networks, Bidirectional Long Short-Term Memory, and Transformer architectures to collectively capture spatial, temporal, and long-range dependencies in network traffic. A hybrid feature selection method that combines Mutual Information, Extra Trees significance, and Recursive Feature Elimination reduces the feature space from 42 to 20 discriminative characteristics. We use the Optuna framework and a Tree-structured Parzen Estimator (TPE) sampler to find the best model settings through Bayesian hyperparameter optimisation. Tests on the UNSW-NB15 benchmark dataset show that HybridNet-IDS outperforms six optimised classical machine learning baselines. It has a Matthews Correlation Coefficient of 0.8085, a ROC-AUC of 97.69%, and an accuracy rate of 90.36%. Five-fold cross-validation and bootstrap resampling confirm statistical robustness and generalisation stability. Also, explainable AI-based analysis shows clearly how each feature affects decisions, which makes people more trusting and open. The results show that HybridNet-IDS is a good, optimised, and easy-to-understand choice for current intrusion detection systems.