Graph Convolutional Networks for DDoS Attack Detection in a Lossy Network

This study introduces a resilient solution for the detection of DDoS attacks in IoT systems, leveraging the capabilities of Graph Convolutional Networks (GCN). By conceptualizing IoT devices as nodes within a graph structure, we present a refined detection mechanism capable of operating efficiently even in lossy network environments. Two graph construction methodologies, namely distance-based and correlation-based approaches, are introduced and evaluated against a spectrum of performance metrics including binary accuracy, F1 score, and Area Under the Receiver Operating Characteristic Curve (AUC). By studying different levels of network connection loss and various attack situations, we demonstrate that the distance-based graph structure is effective and tough in spotting DDoS attacks, substantiating its good performance even in lossy network scenarios. The results indicate a remarkable resilience of the GCN model, showcasing not only the heightened detection rates in aggressive attack settings with an F1 score up to 85% but also its sustained performance in environments with up to 50% connection loss with at most 6% drop in F1 score. The findings from this study highlight the advantages of utilizing GCN for the security of IoT systems which benefit from high detection accuracy while being resilient to connection disruption.