Prediction and graph visualization of cyber attacks using graph attention networks

Abstract

This study proposes a hybrid approach for visualizing cyberattacks by combining the deep learning-based GAT model with JavaScript-based graph visualization tools. The model processes large, heterogeneous data from the UNSW-NB15 dataset to generate dynamic and meaningful graphs. In the data cleaning phase, missing and erroneous data were removed, unnecessary columns were discarded, and the data was transformed into a format suitable for modeling. Then, the data was converted into homogeneous graphs, and heterogeneous structures were created for analysis using the GAT model. GAT prioritizes relationships between nodes in the graph with an attention mechanism, effectively detecting attack patterns. The analyzed data was then converted into interactive graphs using tools like SigmaJS, with attacks between the same nodes grouped to reduce graph complexity. Users can explore these dynamic graphs in detail, examine attack types, and track events over time. This approach significantly benefits cybersecurity professionals, allowing them to better understand, track, and develop defense strategies against cyberattacks.