Adversarial Network Traffic: Towards Evaluating the Robustness of Deep-Learning-Based Network Traffic Classification

Network traffic classification is used in various applications such as network management, policy enforcement, and intrusion detection systems. Although most encrypt their some of them dynamically change port numbers, Machine Learning (ML) especially Deep (DL)-based classifiers have shown impressive performance classification. In this article, we evaluate the robustness DL-based against Adversarial Traffic (ANT). ANT causes to predict incorrectly using Universal Perturbation (UAP) generating methods. Since there no need buffer before sending ANT, it generated live. We partition input space into three categories: packet classification, flow content time series To generate propose new attacks injecting UAP traffic. AdvPad attack injects a packets classifiers. AdvPay payload dummy AdvBurst specific number with crafted statistical features based on selected burst The results indicate little traffic, highly decreases all categories.