phish-transformer

Phish-Transformer

A lightweight, pure-CPU transformer that detects phishing URLs and warns users inside Chrome.

Version 1.0 – 2025
Custom Commercial Licence – see LICENSE for terms.

Tested on Python 3.10–3.13

---

Contents

• Installation
• Quick Start
• Pipeline Flow
• API Usage
• Model Specs
• Evaluation
• Chrome Extension
• Packaging
• Possible Extensions
• Contributing
• License

What It Does

1. Learns URL patterns from the UCI PhisUSIIL Phishing URL Dataset.
2. Tokenizes each URL into 75 ASCII characters.
3. Feeds the sequence into a ~ 45k parameter transformer encoder
4. Returns a probability via a Flask REST endpoint /predict .
5. Displays a red banner in Chrome when the probability > 0.5.

---

Limitations

The model judges only the URL string (characters, sub-domains, TLD, etc.) and does not fetch or inspect page content.
It is designed as a zero-latency first filter against look-alike domains; content-based attacks require additional page-analysis layers.

Future Improvements

• Integrate content-based page analysis to catch phishing beyond URL strings.
• Support real-time updates from threat feeds to improve detection of new domains.
• Extend multilingual support and handle punycode / internationalized domains.
• Explore edge deployment optimizations (INT8 quantization, 1-D CNN variant).
• Implement user feedback loop for continual model retraining.

---

Folder Map

phish_transformer/
├── datasets/                      # UCI CSV + splits
├── models/                        # *.pt checkpoints (git-ignored)
├── src/                           # modular code
│   ├── data/                      # tokenizer + preprocessing
│   ├── model/                     # transformer definition
│   ├── training/                  # training loop
│   └── inference/                 # Flask API + evaluation
├── extension/                     # Chrome MV3 extension
├── tests/                         # pytest suite
├── main.py                        # end-to-end pipeline
├── requirements.txt               # dependencies
└── README.md                      # this file

---

Installation

git clone https://github.com/YOU/phish-transformer.git
cd phish-transformer
python -m venv venv
source venv/bin/activate        # Windows : venv\Scripts\activate
pip install -r requirements.txt # includes torch, flask, pandas, etc.

---

Quick Start

# 1. Train & export
python main.py

# 2. Start API
flask --app src/inference/app run --port 8000

# 3. Load extension
Chrome → Extensions → Developer mode → Load unpacked → select extension/

# 4. Test
pytest tests/

---

Pipeline Flow

---

API Usage

POST /predict Request :

{"url" : "https://example.com"}

Response :

{"phishing" : 0.97}

CLI test

curl -X POST http://127.0.0.1:8000/predict \
-H "Content-Type: application/json" \
-d '{"url" : "https://example.com"}'

GUI testers

• VS Code Thunder Client: install extension → POST → URL above → Body → raw JSON → same payload.
• Postman: new request → POST → URL → Body → raw → JSON → send.

---

Model Specs

---

Evaluation

python src/inference/evaluate.py   # produces roc.png

---

Chrome Extension

1. Start Flask Server.
2. Chrome → Extensions → Developer mode → Load unpacked → select extension/.
3. Browse any site; red banner appears when score > 0.5.

---

Packaging for Chrome Web Store

1. Zip the extension/ folder (must contain manifest.json at top level).
2. Upload zip in Chrome Developer Dashboard.
3. Host the ≥ 1 MB model externally (Google Drive, CDN) and update fetch() URL in content.js.

---

Possible Extensions

• Distil to 1-D CNN for < 200 kb edge deployment
• Quantization (INT8) for 2x speed-up
• On device Chrome ML (TensorFlow-Lite)
• Multilingual URL support (Unicode normalizer)
• Continuous retraining with user feedback

---

Contributing

1. Fork
2. Feature branch (git checkout -b feat/awesome)
3. Commit with clear messages
4. Push & open pull requests

---

Acknowledgements

• UCI PhisUSIIL Phishing URL Dataset
• PyTorch, Flask, Chrome Extension APIs

---

Licence

This project is licensed under the GNU General Public Licence v3.0 – see LICENSE for details.
Commercial use is allowed only if you open-source your entire derivative under the same licence.