The increasing complexity of enterprise networks has significantly intensified the risk of sophisticated cyberattacks, particularly zero-day attacks that exploit previously unknown software vulnerabilities. Conventional signature-based intrusion detection systems and standalone machine learning models often fail to identify these attacks due to the absence of prior attack signatures and their limited capability to learn evolving behavioural patterns. This paper proposes a Hybrid Stacking-Based Ensemble Learning Framework that integrates Random Forest, Support Vector Machine, Extreme Gradient Boosting, and Long Short-Term Memory models through a Logistic Regression meta-classifier to enhance zero-day attack prediction. The framework combines statistical, behavioural, and temporal feature engineering to effectively characterize malicious network activities while reducing false-positive alarms. Public benchmark datasets, including NSL-KDD, UNSW-NB15, and CIC-IDS2017, together with enterprise telemetry data, are employed to evaluate the proposed framework. Comprehensive preprocessing, feature normalization, and Bayesian hyperparameter optimization are incorporated to improve model stability and predictive capability. Experimental evaluation demonstrates superior performance compared with conventional machine learning models in terms of Accuracy, Precision, Recall, F1-score, and ROC-AUC. The proposed framework achieves improved generalization or previously unseen attack patterns and provides a scalable solution for enterprise cybersecurity. The findings indicate that integrating heterogeneous learning models with temporal feature modelling significantly strengthens proactive intrusion detection and supports the development of intelligent security systems capable of defending against emerging cyber threats.

 

Keywords— Zero-Day Attack, Ensemble Learning, Stacking, Enterprise Security, Machine Learning, LSTM, Intrusion Detection.