Focal Loss Based Two-Stage Training for Class Imbalance Network Intrusion Detection

Sijie Luo,Zhiheng Zhao,Qiyuan Hu
DOI: https://doi.org/10.1109/icftic54370.2021.9647139
2021-01-01
Abstract:With the rapid development of the industrial Internet, the security of industrial control network has become a vital topic. The network intrusion detection system (NIDS), as an important security device in contemporary networks, is of vital importance in detecting network traffic and ensuring network security. In this paper, a two-stage training Deep Neural Network (DNN) based on cross-entropy and focal loss, called CE-FL-NIDS, is proposed to handle with the problems caused to NIDS due to the uneven data distribution in the traffic dataset. The effectiveness of the proposed model is verified by performing 10-class classification on the UNSW-NB15 dataset. The experimental results show that: i) compared with the two-stage training DNN based on cross-entropy (CE-CE-NIDS), the proposed CE-FL-NIDS has increased the detection rate and F1-score by 8.76% and 6.37% respectively; ii) compared with the two-stage training DNN based on focal loss (FL-FL-NIDS), CE-FL-NIDS has increased the accuracy, DR and F1-score by 3.10%, 2.77%, 2.62% respectively. As compared to its counterparts, CE-FL-NIDS maintains the representation learning ability of DNN and improves the performance of the class imbalance NIDS.
What problem does this paper attempt to address?