SC-Net: Network Intrusion Detection with Deep Supervised Contrastive Learning and Normalized Classifier

Network intrusion detection (NID) has attracted much attention as it is essential in preventing security threats and protecting networks from attacks. However, existing methods face the following challenges: (1) poor feature extraction capability; (2) not well-designed to address the class imbalance problem; (3) failure to take full use of label information and learn classification-oriented features, degrading the NID performance. To this end, we proposed SC-Net, a two-stage training model with deep supervised learning and a normalized classifier, to overcome the abovementioned challenges. During the pretraining stage, the learned embedding will be optimized by both a supervised contrastive loss and a classification loss, so that the embedding with the same label will be more compact in the feature space. After that, in the finetuning stage, the weight of the classifier will be normalized for catering to classification tasks in scenarios of a class imbalance dataset. The experiment shows that SC-Net outperforms all comparative models in four metrics.