Wasserstein Conditional Generative Adversarial Networks for Class Balancing in Intrusion Detection Datasets

Abstract

Conventional network intrusion detection systems have numerous obstacles in handling data. These issues may significantly impact its efficacy and efficiency. In real-world scenarios, attacks are rare compared to the high volume of normal network activities, creating a significant imbalance in the dataset. This imbalance causes the model to concentrate more on normal traffic data, thus decreasing its sensitivity in identifying attacks and impacting the overall efficacy of the intrusion detection system. Training data plays a critical role when training an intrusion detection system. Generating enough training data is a challenging task. One approach to handle this challenge is to utilize Generative Adversarial Networks, a machine learning technique that generates synthetic data by placing a generator and a discriminator—two neural networks contradict each other. The generator creates realistic data examples, while the discriminator assesses them, improving the data’s validity through iterative training. This paper suggests employing Wasserstein Conditional Generative Adversarial Networks (WCGANs) to tackle the imbalanced class issue and improve the effectiveness of intrusion detection systems. Providing realistic adversarial examples, the model enhances deep neural network training, hence complementing deep learning techniques. This research focuses on handling the difficulty of class imbalance in network intrusion detection models using WCGAN. By generating synthetic data for both normal and attack categories, the system improves the detection of underrepresented labels. WCGAN can also be leveraged to generate realistic network traffic samples, thus enhancing the robustness of the classifiers in both binary and multi-class scenarios.