ZERO-DAY ATTACK PREDICTION USING ENSEMBLE MACHINE LEARNING WITH THREAT INTELLIGENCE DATA

Abstract

Modern cybersecurity faces ongoing difficulties because zero-day attacks exploit unknown system weaknesses which make traditional signature-based protection systems useless. The research develops an innovative ensemble machine learning system which unites threat intelligence data from multiple sources with Random Forest (RF) and Extreme Gradient Boosting (XGBoost) and Long Short-Term Memory (LSTM) models for detecting zero-day attacks. The proposed framework unites three different data sources which include CVE/NVD vulnerability databases and honeypot network telemetry and dark web intelligence feeds through a single analytical process. The model achieves better predictive results through its soft-voting mechanism and advanced feature engineering which combines behavioral and contextual and temporal indicators. The ensemble model outperformed all baseline learners in experimental testing with real-world data by achieving 94% accuracy and 0.95 ROC-AUC and 0.83 Z-DR while reducing false positives by 20% compared to the best single model (LSTM). The Wilcoxon signed-rank test (p < 0.01) shows that the improvements achieved hold statistical significance.

The framework demonstrated operational effectiveness through case studies of CVE-2024-3094 and CVE-2025-21788 by producing predictive alerts before official disclosure dates as shown in recent predictive cyber threat intelligence and early-warning system research. The research shows that uniting diverse threat intelligence with ensemble learning techniques creates an intelligence-based cyber defense system which represents a major step toward developing future zero-day protection systems. The research establishes a flexible framework which enables organizations to implement adaptive intelligence-based cybersecurity solutions for their enterprise and national defense systems.