Robust Detection of Malicious Traffic in Label Noise and Imbalanced Datasets

To address both label noise and data imbalance in malicious traffic detection, we propose a novel network traffic intrusion detection system, RDMT. The core idea is that, ideally, traffic samples of the same class and similar behavioral patterns should exhibit similarities in the feature space. Additionally, samples with the same true labels should be located close to each other in the decision space. Given the presence of label noise, it is crucial to select correctly labeled samples and accurately infer the labels of other samples within the same class to mitigate the impact of the noise. During training, we focus on learning from minority class samples to effectively handle the issue of class imbalance.