DA-GAN: Domain Adaptation for Generative Adversarial Networks-assisted Cyber Threat Detection

The rising development of machine learning (ML) techniques has become the motivation for research in applying their outstanding features to facilitate intelligent intrusion detection systems (IDSs). However, ML-based solutions also have drawbacks of high false positive rates and vulnerability to sophisticated attacks such as adversarial ones. Therefore, continuous evaluation and improving those systems are necessary tasks, which can achieve by simulating mutated real-world attack scenarios. Taking advantage of the Generative Adversarial Network (GAN) and Domain Adaptation technique, our approach proposes DA-GAN, a framework that can generate mutated network attack flows. Those crafted flows then work as supplemental training data for ML-based IDS to improve its robustness in dealing with new and complicated attacks. Our framework is implemented and evaluated via experiments on the public CIC-IDS2017 and CIC-IDS2018 datasets. The results prove the effectiveness of the proposed framework in continuously strengthening ML-based IDS in the fight against network attack actors.