A Bagged Multi-Layer Perceptron Framework for Robust Phishing Website Detection

Abstract

Phishing attacks remain a persistent cybersecurity threat, causing substantial economic and operational losses worldwide. Although ensemble learning and deep neural networks have been widely applied to phishing detection, many existing approaches suffer from high computational cost, limited interpretability, or insufficient statistical validation of performance gains over strong baselines. This study proposes a Bagged Multi-Layer Perceptron (BMLP) framework designed to achieve robust generalization with controlled variance while maintaining practical deployment efficiency. The proposed approach integrates Principal Component Analysis (PCA) for dimensionality reduction with bootstrap aggregation of neural networks to reduce model correlation and overfitting. Experiments were conducted using the Web Page Phishing Detection dataset from Kaggle, consisting of 11,430 labelled URLs. PCA was fitted exclusively on the training data to prevent information leakage, reducing the original 88 features to 52 components while preserving 90.3% of the variance. Performance was evaluated using 5-fold cross-validation, complemented by ablation studies and paired statistical tests. Results show that BMLP achieves the highest mean accuracy among evaluated models and demonstrates statistically significant improvements over Random Forest and competitive performance relative to XGBoost and Single MLP, with effect sizes indicating meaningful practical gains. Computational analysis further shows that BMLP satisfies real-time processing requirements (≈1,000 URLs/sec) with moderate training overhead and acceptable memory consumption on consumer-grade hardware. This work provides a statistically grounded and reproducible evaluation of a BMLP-based phishing detector, highlighting a balanced trade-off between predictive performance, robustness, and computational efficiency suitable for real-world cybersecurity applications.