Constrained optimization based adversarial example generation for transfer attacks in network intrusion detection systems

Deep learning has enabled network intrusion detection rates as high as 99.9% for malicious network packets without requiring feature engineering. Adversarial machine learning methods have been used to evade classifiers in the computer vision domain; however, existing methods do not translate well into the constrained cyber domain as they tend to produce non-functional network packets. This research views the payload of network packets as code with many functional units. A meta-heuristic based generative model is developed to maximize classification loss of packet payloads with respect to a surrogate model by repeatedly substituting units of code with functionally equivalent counterparts. The perturbed packets are then transferred and tested against three test network intrusion detection system classifiers with various evasion rates that depend on the classifier and malicious packet type. If the test classifier is of the same architecture as the surrogate model, near-optimal adversarial examples penetrate the test model for 69% of packets whereas the raw examples succeeds for only 5% of packets. This confirms hypotheses that NIDS classifiers are vulnerable to adversarial attacks, motivating research in robust learning for cyber.