AI-Driven Phishing Detection
Adversarially-tested ML over 235k URLs for SOC automation
Context
An AI-for-Cybersecurity project building and critically evaluating supervised models for phishing-URL classification on the PhiUSIIL Phishing URL dataset — with a deliberate focus on adversarial robustness rather than headline accuracy.
The problem
Phishing remains a dominant initial-access vector (MITRE T1566.002). Signature lists can’t keep pace, and models that score 99%+ on clean test data can collapse under obfuscation. The question: which model families generalise, and how brittle are they when attackers fight back?
My approach
I engineered lexical and host-based URL features, then trained and compared bagging (Random Forest), boosting (XGBoost, LightGBM, CatBoost) and a neural network (MLP). Crucially, I designed experiments to measure how feature obfuscation and poisoned training samples shift the decision boundary — moving the analysis beyond simple classification accuracy.
Feature engineering
Strong discriminators emerged from URL structure: legitimate business URLs trend short (SEO-driven), while phishing URLs show high length variance, random tokens and disposable-infrastructure signals such as IsHTTPS=0.
Model comparison
All five families exceeded a macro-F1 of 0.99 on clean data, with the MLP reaching 99.97% accuracy. Tree ensembles offered the best stability / efficiency trade-off for real-time deployment.
Adversarial reality check
Obfuscation techniques pushed macro-F1 below the headline figures — demonstrating that clean-data accuracy is deceptive for deployed defences, and that a production threshold must account for class imbalance (phishing is <1% of real traffic) and false-positive cost.
Outcome
A rigorous, deployment-aware evaluation that treats ML as a security control — not a leaderboard — and ties detection back to the MITRE ATT&CK phishing technique and a CIA-triad impact analysis.