<![CDATA[Phishing website attacks are a severe cyber security threat that manipulates users into surrendering sensitive credentials. Machine learning detection based on URL lexical features offers a proactive, real-time alternative to reactive blacklists, but large-scale datasets such as URL-Phish 2025 (89.56% benign vs. 10.44% phishing across 111,660 records) exhibit extreme class imbalance that biases classifiers toward the majority class. This study analyzes the performance of combining the Synthetic Minority Over-sampling Technique (SMOTE) with the Light Gradient Boosting Machine (LightGBM) classifier, following the experimental baseline of Dam and Tran (2025). The dataset was divided using a stratified 75:10:15 split into training (83,745), validation (11,166), and testing (16,749) sets. SMOTE was applied only to the training set, balancing it to 75,000:75,000 samples in 0.5339 seconds. On 16,749 test samples, the optimal LightGBM + SMOTE model (Scenario II) achieved Accuracy of 97.94% (+4.29% over baseline), Recall of 98.23% (+56.83%), F1-Score of 90.88% (+33.25%), and ROC-AUC of 99.32%, while reducing False Negative threat leakage by 96.98% (from 1,025 to 31 cases), at an inference latency of only 0.0039 ms per URL. These results confirm that SMOTE and LightGBM together form an effective, real-time-ready cybersecurity detection solution.]]>
Copyrights © 2026
