Skip to content
Projects

chamanthmvs/Phishing-Website-Detection

View on GitHub

chamanthmvs/Phishing-Website-Detection has 211 GitHub stars and 75 forks. Star changes are not measured yet. Last checked 2026-11-16 00:12 UTC. Last push 2026-11-08.

How these counts are measured

Measured star growth

in 24h
Update delayed
in 7 days
Update delayed
in 15 days
Update delayed
in 1 month
Update delayed

Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.

About

It is a project of detecting phishing websites which are main cause of cyber security attacks. It is done using Machine learning with Python

Machine Learning

README.md

Original

Phishing Website Detection

Phishing Website Detection classifies URLs as legitimate or potentially phishing using machine learning. It is built as two decoupled services:

  • a FastAPI backend that owns feature extraction and the trained models and exposes a JSON API;
  • a Next.js frontend that talks to the backend only through that API.

For the visual project presentation, visit the project page.

At a glance: 20 lexical URL features · 3 compared models (Random Forest, Gradient Boosting, Logistic Regression) · ~90% accuracy · 0.96 ROC AUC · clean API/UI separation · fully offline, reproducible training. See docs/IMPROVEMENTS.md for the full history.

Architecture

┌──────────────────────┐      HTTP / JSON       ┌───────────────────────────┐
│  Next.js frontend     │  ───────────────────►  │  FastAPI backend           │
│  (frontend/)          │   GET  /api/models     │  (backend/)                │
│  no model, no data,   │   POST /api/scan       │  - featureExtraction.py    │
│  no Python            │  ◄───────────────────  │  - models/*.sav + metadata │
└──────────────────────┘                        └───────────────────────────┘

The frontend contains no model, no dataset, and no scikit-learn — every prediction is requested from the backend API. This keeps the two layers independently developed, tested, and deployed.

How it works

The model uses lexical features — 20 signals computed purely from the URL string, with no request to the site being scanned. This makes every scan fast, deterministic, privacy-preserving, and reproducible, and it behaves identically on a laptop and on a locked-down cloud host.

Family Example signals
Length URL / hostname / path length
Composition counts of ., -, @, ?, &, =, %, digits; digit ratio
Structure subdomain count, raw-IP host, // redirection, repeated http token
Reputation heuristics HTTPS usage, prefix/suffix -, URL shorteners, phishing keywords

Model comparison

Three models are trained and compared on a held-out 25% test set (2,553 URLs; 1,560 legitimate / 993 phishing). The best is served as the API default.

Model Accuracy ROC AUC CV accuracy
Random Forest (default) 89.5% 0.959 90.2%
Histogram Gradient Boosting 90.0% 0.957 90.0%
Logistic Regression 80.4% 0.874 81.0%

Exact numbers and feature importances live in backend/models/models_metadata.json, regenerated on every training run.

Quick start

1. Backend

cd backend
python3 -m venv ../.venv && source ../.venv/bin/activate
pip install -r requirements.txt
python train.py                       # train the models (offline, reproducible)
uvicorn app.main:app --reload --port 8000

2. Frontend (in a second terminal; requires Node.js 20.9+)

cd frontend
npm install
cp .env.local.example .env.local      # defaults to http://localhost:8000
npm run dev                           # http://localhost:3000

See backend/README.md and frontend/README.md for details.

Tests

cd backend
pip install -r requirements-dev.txt
python -m pytest                      # feature extraction + registry + API

Repository layout

backend/     FastAPI API, feature extraction, training, models, tests
frontend/    Next.js app (App Router, TypeScript)
docs/        GitHub Pages presentation + improvement log
notebooks/   original exploratory notebooks (historical; see note below)

A note on the notebooks

notebooks/ holds the original exploratory work, which relied on live WHOIS/DNS lookups and an Alexa traffic-rank API that was shut down in 2022. The current pipeline replaces that with a purely lexical, offline, reproducible approach. The notebooks are retained for historical context only.

Disclaimer

A heuristic tool for learning, experimentation, and responsible security research. A prediction is not a guarantee that a website is safe or malicious — always use good judgement.

License

Released under the MIT License.

Built by

Built by chamanthmvs.