Cybersecurity Risk Prediction Using Graph Neural Networks

Abstract

Cybersecurity threats have grown in complexity and scale, demanding advanced prediction models that account for intricate interconnections within cyber infrastructures. This paper proposes a novel approach for cybersecurity risk prediction using Graph Neural Networks (GNNs). By modeling computer systems, network logs, and threat vectors as dynamic graphs, we capture relational and temporal dependencies critical for forecasting cyber risks. We utilize a curated dataset of historical attack traces, system logs, and vulnerability disclosures, transforming them into attributed graphs. Our proposed GNN architecture integrates node-level anomaly detection, edge-based threat propagation modeling, and temporal graph attention mechanisms. Experimental results show that our GNN-based model significantly outperforms traditional machine learning baselines in predicting emerging cyber threats, offering superior precision and early warning capabilities. The findings highlight GNNs as a promising direction for proactive cyber risk assessment in real-time environments.