Publication

Abstract : Anomaly Detection Systems (ADS) are essential for protecting vital digital infrastructures from cyberattacks. However, because of the dynamic nature of incursion techniques and the lack of updated and varied training samples, the efficacy of machine learning-based ADS models frequently encounters constraints. In this work, we propose a unique method of incorporating generative adversarial networks (GANs) to enhance ADS performance. Data scarcity is a major issue with ADS training datasets, which we solve by using GANs to produce synthetic network traffic data that closely resembles real-world behavior. To generate real-world network traffic patterns that are specifically designed to depict abnormal activity, we employ three different GANs models: Wasserstein GAN, Conditional Tabular GAN, and Vanilla GAN. We show a notable enhancement in ADS performance for detecting such activity when employing this artificial data resampling method. We give empirical proof of the success of our technique through extensive trials conducted on the CIC-IDS2017 benchmark dataset supplemented with data generated by a GAN. In today's networked and susceptible digital world, our research indicates that adding GANs to ADS can improve anomaly detection performance, especially for attacks with little training data. This presents a viable option for bolstering an organization's cybersecurity posture.