DeepICU: imbalanced classification by using deep neural networks for network intrusion detection

Cyber intrusions are becoming more commonplace, more dangerous, and more sophisticated. Therefore, there is a desperate need for a robust intrusion detection system. In a healthy network environment, a majority of the connections are initiated by benign behaviours. Despite a wide variety of attacks, they only occupy a limited fraction of the observed network traffic. The imbalanced class distribution implicitly forces conventional classifiers to be biased toward the majority/benign class, thus leave many attack incidents undetected. In this paper, we design a new intrusion detection system named DeepICU based on deep neural networks. To address the class imbalance issue, we design two novel loss functions, i.e., attack-sharing loss and attack-discrete loss, that can effectively move the decision boundary towards the attack classes. Extensive experimental results on three benchmark datasets demonstrate the high detection accuracy of DeepICU. In particular, compared with eight state-of-the-art approaches, DeepICU always provides the best class-balanced accuracy.