AI-Driven Real-Time Severity Prediction for Cyber Attacks using Machine Learning
Epifelward Niño O. Amora, Jose C. Agoylo, Jimson A. Olaybar, Josephine C. Muñasque, Patrick D. Cerna · 2025
Cybersecurity operations are increasingly challenged by the high volume and complexity of cyberattacks, necessitating automated solutions for real-time threat prioritization. This study proposes a machine learning-based framework for real-time prediction of cyberattack severity levels—classified as Low, Medium, or High—using engineered and preprocessed network traffic features. A novel temporal feature, Time Since Last Attack, was incorporated to capture contextual attack frequency. The proposed system was evaluated using Random Forest, Extreme Gradient Boosting (XGBoost), and a Neural Network model on a synthetic cybersecurity dataset. Among these, XGBoost achieved the highest classification accuracy of 94.5%, with corresponding precision, recall, and F1-scores exceeding 94%. The Random Forest and Neural Network models followed with 91.7% and 89.4% accuracy, respectively. All models were assessed using macro-averaged metrics to ensure class-balanced evaluation. The results confirm the efficacy of ensemble learning techniques in prioritizing cybersecurity alerts based on severity, thus enhancing decision-making in security operations. Key contributions include a comparative model analysis, implementation of an interpretable feature engineering strategy, and validation of predictive accuracy in a simulated real-time environment.