Detecting Phishing Websites Using Machine Learning

Abstract—Phishing websites remain a dominant entry-point for cyber-attacks. This study proposes a light-weight machine learning pipeline that spots malicious URLs in real time. We collected 734 benign links from Alexa and 412 confirmed phishing links from PhishTank, extracted 30 lexical and host-based features, and trained five classifiers. The Random Forest model achieved the top performance with 93% accuracy, 0.98 precision and 0.97 recall on a hold-out test set. To improve interpretability, SHAP values were employed, revealing that URL length, sub-domain count and HTTPS flag contribute most to the decision process. We then deployed the model as a Flask micro-service, processing a median of 700 URLs per second on commodity hardware. Comparative tests against Google Safe Browsing and VirusTotal show our approach detects newly registered phishing domains up to 48 hours earlier. The code and labelled dataset are released open-source to foster reproducibility and future work in proactive web-security.