Alert Prediction in Computer Networks Using Deep Graph Learning

Cyber-attacks have surged in intensity and complexity, often outpacing existing detection systems. Advanced attack prediction algorithms can help mitigate these threats, reducing their impact and strengthening network security. While prior research has advanced attack predictability and source estimation, deep graph modeling has yet to integrate such advancements fully. This article introduces a novel method for binary link prediction in cyber-attack networks, leveraging dynamic graph deep learning and temporal graph neural networks (TGN). The approach begins by constructing a time-evolving attack graph based on historical alert data within a specific time window. This dynamic graph is then fed into a TGN, which automatically extracts structural and temporal features. The model predicts the likelihood of future attack links between entities, enabling more effective identification of potential cyber threats before they occur. In this article, three different deep networks were evaluated for predicting attacks. The dynamic graph method demonstrated 10–18% better performance compared to the other methods, improving accuracy in link prediction.