BGP Anomaly Detection with Balanced Datasets

We use machine learning techniques to build predictive models for anomaly detection in the Border Gateway Protocol (BGP). Imbalanced datasets of network anomalies pose limitations to building predictive models for anomaly detection. In order to achieve better classification performance measures, we use resampling methods to balance classes in the datasets. We use undersampling, oversampling and combination techniques to change class distributions of the datasets. In this paper we build predictive models based on preprocessed network anomaly datasets of known Internet network anomalies and observe improvement in classifier performance measures compared to those reported in our previous work. We propose to use resampling combination techniques on datasets along with Decision Tree and Naïve Bayes classifiers in order to achieve the best trade-off between (1) the F-measure and the length of model training time, and (2) avoiding overfitting and loss of information