A Deep Learning Based Semi-Supervised Network Intrusion Detection System Robust to Adversarial Attacks

Abstract

Network intrusion detection systems (NIDS) are used to detect abnormal behavior in network traffic, which is vital for secure communication. Recently, deep learning based solutions have been adopted for NIDS which suffer from two main problems. Most of them are based on supervised learning and cannot utilize the information that can be obtained from unlabeled data. Also, deep learning based methods are shown to be vulnerable to adversarial attacks. In this paper, we propose a novel semi-supervised and adversarially robust deep learning based approach which can utilize both labeled and unlabeled training samples. Our IDS first performs K-Means clustering to soft label part of the unlabeled data and then obtain a decision tree based on labeled and soft labeled samples. It then pretrains an autoencoder based multi-layer perceptron and later learns separate multi-layer perceptrons on each individual leaf of the decision tree. Our results show that the performance of our system is comparable to state-of-the art supervised learning approaches and outperforms existing state-of-the-art semi-supervised NIDS. Furthermore, we have extensively tested the adversarial robustness of our method using the popular blackbox Fast Gradient Sign Method (FGSM) and Generative Adversarial Network based IDSGAN approaches. Comparisons with other state-of-the-art NIDS baselines show that our proposed mechanism provides significantly higher adversarial detection rates, proving the robustness of our system to adversarial attacks.