Using NIDSGAN to generate practical adversarial network traffic

The researchers propose a GAN-based attack algorithm, NIDSGAN, which can generate realistic adversarial network traffic to bypass machine learning-based network intrusion detection systems (NIDS). Under white-box, black-box, and restricted black-box settings, the algorithm's average success rates in attacking two state-of-the-art DNN-based NIDS are 99%, 85%, and 70%, respectively, and about 70% on average against logistic regression, SVM, decision tree, and KNN-type NIDS. The authors state that ML-based NIDS that have not deployed defenses against adversarial traffic may be compromised in the future.