Skip to content
← Projects

BERT-CNN Phishing URL Detection

BeTNet, a hybrid BERT-CNN classifier that detects phishing URLs at 98.33% accuracy and 99.31% sensitivity.

Aug 2024 – Jun 2025 · URECA
  • research
  • AI / ML

Problem

Phishing is usually the first move in an attack, and reputation blacklists can’t keep pace with domains registered minutes before a campaign. This URECA research project classifies a URL from its own structure — no blacklist lookup, no page load — so detection happens before a victim ever clicks through.

Approach

The model, BeTNet, pairs BERT’s contextual embeddings of the URL with a CNN over those embeddings: BERT captures the structure and semantics of the URL string, and the CNN sharpens the local patterns that separate phishing from legitimate traffic. Implemented and trained in Python.

Results

BeTNet performance
98.33%
Accuracy
99.31%
Sensitivity
phishing caught
97.40%
Specificity
legitimate passed

In the confusion matrix, true positives and true negatives dominate: false negatives (phishing scored as legitimate) are very rare — the 99.31% sensitivity — while false positives stay low, reflected in the 97.40% specificity. BeTNet catches phishing aggressively without over-flagging legitimate sites.