Unmasking High Performance Phishing Detection with Feature Fusion of a Tree-based Transfer Learning Boosted Approach using SMOTE-Tomek Balancing: A Pilot Study
Abstract
Daily exchange of data has eased its proliferation, accessibility, availability and
interoperability of accompanying devices. This is further eased with widespread
proliferation of smartphones, which has also in turn – birthed attacks targeted at
network infrastructures to exploit interconnected, shared resources. Its rising
success rate and increased susceptibility to these attacks can be traced to user's
personality traits and behaviours, which renders them repeatedly vulnerable to
such exploits especially those rippled across spoofed websites. Our study posits a
stacked approach to classify malicious contents over a spoofed website. It explores
3-base classifiers namely Cultural Genetic Algorithm, Random Forest, and
Korhonen Modular Neural Network – whose output is utilized as input for
XGBoost meta-learner. Our study resolved dataset imbalance challenge using the
SMOTE-Tomek mode; while, the selected predictors was resolved using the relief
rank feature selection. Results shows ensemble yields F1 0.95, Accuracy 1.00,
Recall 0.98, Precision 1.00, and Specificity 1.00 – to accurately classify all 2,764
cases of its test dataset. Ensemble outperformed benchmarks and was effectively
classified phishing (cues and lures) contents on websites.