A Stable Generative Adversarial Network Architecture for Network Intrusion Detection

Many approaches have been proposed for detecting and categorizing malicious activities over the years. The adversarial training process has recently been applied to solve this task, yielding remarkable results. Generative Adversarial Networks (GANs) can model complex distributions of high-dimensional data, which is useful for anomaly detection. Few studies have examined the use of GANs to detect network intrusions. This paper aimed to develop a new architecture for generative and discriminative training to improve the detection of multi-attack types with a stable training process using ensemble convolutional neural networks (CNNs). By applying the stacking ensemble learning method to the public datasets, NSL-KDD and UNSW-NB15, efficient intrusion detection is achieved compared to the state-of-the-art performance. Additionally, the training process is more stable with this novel architecture, and the model converges faster. The proposed method's accuracy, precision, sensitivity, and F1 score are obtained at 86.36%, 86.29%, 86.36%, and 85.81% for the NSL_KDD dataset, respectively. For the UNSW-NB15 dataset, the accuracy, precision, sensitivity and F1 score are 89.43%, 91.51%, 89.44%, and 89.72%, respectively.