BAYESIAN-OPTIMIZED ENSEMBLE SUPPORT VECTOR MACHINE MODEL FOR PHISHING EMAIL DETECTION

Abstract

With the rapid growth of email use, phishing and malware attacks have become more frequent and sophisticated, often slipping past traditional defenses such as blacklists and rule-based filters. Existing detection models, including SVM, XGBoost, and CNN, have improved accuracy but still depend heavily on manually crafted features and struggle to adapt to new or evolving attack patterns. This challenge creates the need for a more flexible and intelligent detection approach capable of learning and adapting to emerging email threats. This study aims to develop an ensemble phishing email detection model combining SVM and XGBoost, optimize it using Bayesian tuning, and evaluate its performance through accuracy, precision, recall, F1-score, and ROC-AUC metrics. This study used an ensemble approach that combines SVM and XGBoost to detect phishing emails. Various SVM models, including Baseline, Grid Search, SGD, and Bayesian-optimized versions, were developed and tested. An optimized Bayesian model was developed to improve accuracy, with performance evaluated using accuracy, precision, recall, F1-score, and ROC-AUC. A well-known Kaggle phishing dataset was used for fair comparison. After cleaning and reducing 10,000 emails with 1,250 features to 9,872 emails and 500 cleaned features, the Baseline SVM reached 0.9287 accuracy, Grid Search SVM improved to 0.96, and SGD SVM slightly dropped to 0.92. The Bayesian SVM performed best at 0.9667, showing greater stability and generalization. The Bayesian-optimized Hybrid Ensemble SVM–XGBoost achieved 0.992 accuracy and 0.9992 ROC-AUC, confirming its strong reliability and effectiveness in phishing detection. Stacking substantially enhanced model stability, generalization, and real-time reliability