The dynamic and constantly evolving landscape of cyber threats demands innovative methods capable of adapting to the complex relationships and structures inherent in network data. Traditional methods often struggle to adequately capture the intricacies of dynamic networks, especially in terms of evolving temporal dynamics and multiscale dependencies. The proposed solution, Enhanced V-GCN, combines the structural insights of Graph Convolutional Networks (GCNs) with the temporal modeling capabilities of Variational Autoencoders (VAEs), further augmented by multiscale convolutions and attention mechanisms. Multiscale convolutions enable the model to aggregate information across broader neighborhood ranges, while attention mechanisms prioritize the most critical nodes and edges, dynamically adapting to changes within the network. This enhanced approach allows V-GCN to effectively capture both nodal and structural patterns, significantly improving performance in node classification tasks. The Enhanced V-GCN model has demonstrated superior performance in node classification, outperforming baseline models with an accuracy of 98.00%, precision of 97.93%, recall of 98%, and an F1-score of 97.92%, indicating robust classification capabilities and exceptional generalization across diverse network structures.
Read full abstract