Bi-GAN-LDA for cybersecurity: a hybrid deep learning framework for advanced network anomaly detection

Abstract Intrusion Detection Systems (IDS) play a crucial role in modern cybersecurity by identifying and mitigating malicious activities in network traffic. However, existing IDS models suffer from high false positive rates, class imbalance issues, and inefficient feature selection, which hinder their ability to detect sophisticated cyber threats. In this study, study proposes Bi-GAN-LDA IDS, a novel hybrid deep learning framework that integrates Bidirectional Generative Adversarial Networks (Bi-GANs) for synthetic attack sample generation and Linear Discriminant Analysis (LDA) for optimized feature selection. Additionally, a custom focal loss function is introduced to enhance the classification of minority attack classes. The efficacy of the proposed Bi-GAN-LDA intrusion detection framework was rigorously validated using a diverse set of benchmark datasets, namely NSL-KDD, UNSW-NB-15, CICIDS-2017, ADFA-LD, and UNR-IDD. Notably, on the ADFA-LD dataset, the model achieved an F1-score of 99.5%, marking a 2.8% performance gain over existing GAN-based IDS frameworks. Furthermore, a substantial 22% reduction in false positive rates was observed when compared to conventional deep learning-based detectors. These improvements underscore the robustness of the proposed method, particularly in addressing the challenge of class imbalance, minimizing false alarms, and enhancing the reliability of real-time anomaly detection in contemporary IDS environments.