Killian Cressant, Ph.D. student in the ROC Team, is planned to defend his Ph.D. thesis on December 18, 2026, at 2:00 p.m., in the Amphitheatre Gaston-Planté, Accès 35 – Level 1, 2 rue Conté, 75003 Paris, France.
Thesis Title: Graph Neural Networks for anomaly detection in network systems
Thesis Title(French): Réseaux de neurones en graphes pour la détection d’anomalies dans les réseaux de télécommunications
Abstract:
In this doctoral thesis, we explored the use of Graph Neural Networks (GNNs) for anomaly detection in networks. To this end, we generated graphs that allow an effective application of these models to network data, while enabling the generalization of this method to other types of applications through the combined use of correlation in the raw data and the semantics of metadata, such as features names. We then conducted experiments to demonstrate the effectiveness of this type of graph construction on various datasets. Subsequently, we investigated ways to improve this approach through optimization methods and the use of virtual nodes. The use of virtual nodes is still relatively recent in the field of GNNs. We therefore first tried to determine whether they could improve model performance. To this end, we designed a generic metric, called LEED, which partially addresses the two main problems of GNNs, namely over-smoothing and over-squashing. By measuring locally over-smoothing phenomena and then using this local information to generate virtual nodes, we were able, for the first time, to establish a connection between these two major problems. After evaluating the LEED method on standard datasets, we developed a framework called EAGLE specifically designed for anomaly detection and tested the use of this method directly on network systems datasets. This framework was designed to enable anomaly detection that accounts for both temporal and spatial characteristics. We therefore designed this new model in two parts: a Graph Recurrent Unit to process temporal data and a GNN-based autoencoder to process spatial information.
Killian’s publications:
Powered by HAL 
