Malicious Traffic Detection with Label Noise Based on Semi-supervised Learning

The application of Deep neural networks (DNNs) in malicious network traffic detection has been intensively studied over the past decades. Despite the great success of DNNs, many real-world traffic datasets may have noisily labeled data, which could significantly degrade the performance of DNNs. Therefore, it is important to develop a label noise-resistant framework for malicious traffic detection. To this end, we propose a semi-supervised learning-based framework for label noise-resistant malicious traffic detection. Specifically, we divide the samples in the dataset into noisy data (unlabeled set) and clean data (labeled set) based on prediction loss, thus transforming the problem into a common semi-supervised learning problem. In addition, we replace the cross-entropy loss with the generalized cross-entropy loss during model training, which effectively suppresses the effect of noisy data. Finally, we compare our approach with common baseline methods and state-of-the-art noise-labeled classifiers on two popular intrusion detection datasets. The experimental results show the effectiveness of our framework.