Website Phishing Attack Detection Using Innovative Meta Learning-Based Ensemble Approach

Phishing remains the most pervasive cybersecurity threat, exploiting human vulnerabilities through sophisticated social engineering. While traditional heuristic and blacklist-based solutions struggle against evolving attacks, this paper proposes a novel stacked ensemble model combining Artificial Neural Networks (ANN) and Bagging K-Nearest Neighbors (KNN) with Logistic Regression (LR) as a meta-learner classifier. Our hybrid framework leverages the strength of ANN in identifying global non-linear patterns and the proficiency of Bagging KNN in local pattern detection, achieving superior performance through optimal decision-level fusion. Experimental results on 2019-2024 datasets demonstrate that the proposed ensemble model outperformed state-of-the-art techniques with a 97.20% accuracy score, a precision of 98%, and a recall of 98%. The proposed model employs hyperparameter-optimised architectures including a 512-256-128-128 layer ANN with Relu activation and 30% dropout, along with a Bagging KNN configuration using 10 estimators with K=7 and Euclidean distance. Key advantages of our stacking approach include enhanced generalization with 2.16% accuracy improvement over the best standalone model, adaptability to emerging attack patterns through weighted averaging, and scalability through the modular design that allows incremental model additions. This study introduces an innovative hybrid ensemble method that stacks ANN and Bagging KNN, incorporating a meta-learning strategy optimized for real-time phishing detection. The model also utilizes SMOTE oversampling to overcome the problem of class disproportion and uses k-fold cross-validation to establish stable performance testing. This study highlights the significance of feature engineering, ensemble methods, and hybrid approaches in enhancing phishing detection.