BGNN: Detection of BGP Anomalies Using Graph Neural Networks
Kévin Hoarau, P Tournoux, Tahiry Razafindralambo · 2022 IEEE Symposium on Computers and Communications (ISCC) · 2022
The Border Gateway Protocol (BGP) builds the communication routes at the Internet scale. Anomalies in BGP have several causes and can impact the Internet stability. BGP data traces are complex and require specific methods such as machine learning to be processed for anomaly detection. Two types of features are used to study large scale events with machine learning models: graph features or statistical features. Despite the recent interest for the concept of Graph Neural Network (GNN), there is no proposal that adapts GNN for BGP anomaly detection directly from the BGP graph. In this paper, we propose BGNN, a GNN model which detects if a node is involved in a large scale BGP anomaly. Our results show a maximum accuracy of 96% and the model can detect an anomaly after 6 minutes with 90% accuracy. These results are promising and suggest GNN for BGP anomaly detection are worth investigating,