Towards Robust Learning using Diametrical Risk Minimization for Network Intrusion Detection

Currently, deep neural networks show great promise in the detection of malicious network traffic at machine speed. However, these networks are typically trained using Empirical Risk Minimization (ERM), which is not robust to misclassified or altered training data. We propose applying Diametrical Risk Minimization (DRM), which is shown to lead to more robust optimization solutions, to train deep neural networks to classify malicious network traffic. Using two different network traffic datasets, we find that when state-of-the-art deep neural networks are trained on partially mislabeled data, utilizing DRM results in higher accuracy compared to equivalent models trained with ERM. More importantly, when models are tested against previously unseen cyber-attack types, models trained with DRM correctly identify the previously unseen cyber-attacks more often. We then show that these deep neural networks are computationally tractable to deploy in real-time on edge computing systems utilizing commercial-off-the-shelf hardware.