Towards Practical Phishing Detection: Addressing Challenges with Hybrid Machine Learning Architectures
C. Sivasankar, A. Tamilarasan, S. Christy, S. Parthiban · 2025
This study presents a new Hybrid Detection Model combining XGBoost with Deep Neural Networks (DNN) to take advantage of feature engineering power and pattern recognition strengths. The recommended solution is then thoroughly tested against three well-developed machine learning methods - Random Forest (RF), Support Vector Machine (SVM), and Logistic Regression (LR) - on six performance measures: Accuracy, Precision, Recall, F1-Score, ROC-AUC, and PR-AUC. Although traditional models (RF, SVM, LR) show evident perfect classification, overfitting issues are implied by these unattainable results. Conversely, the Hybrid Model is more accurate in its performance on noisy, real-world data (Accuracy: 0.866 ± 0.012, ROC-AUC: 0.934 ± 0.008). The highest discriminative features in URL length, domain age, and special character frequency are extracted by SHAP interpretability analysis. These results indicate that although simpler models might have artificially high metrics on clean datasets, the suggested hybrid architecture provides better robustness for practical deployment environments with concept drift and adversarial noise.