Unsupervised representation learning for BGP anomaly detection using graph auto-encoders

The Border Gateway Protocol (BGP) is crucial for the communication routes of the Internet. Anomalies in BGP can pose a threat to the stability of the Internet. These anomalies, caused by a variety of factors, can be challenging to detect due to the massive and complex nature of BGP data traces. Various machine learning techniques have been employed to overcome this issue. The traditional approach involves the extraction of ad hoc features, which, although effective, results in a significant loss of information and may be biased towards a certain type of anomaly. A recent supervised machine learning pipeline learns representations from BGP graphs derived from BGP data traces. Although this solution achieves good anomaly detection results, the representations learned are specific to the types of anomalies within the training data. To overcome this limitation, in this paper, we propose to learn the representations of normal BGP behaviour in an unsupervised manner using a Graph Auto-Encoder (GAE). This approach ensures that the representations are not limited to the specific set of anomalies included in the training set. These representations associated with a Multi-Layer Perceptron (MLP)-based detector allowed to achieve an accuracy rate of 99% in detecting large-scale events, outperforming previous literature results.