CertiScan Backend

Deepfake Document Detection API — AI-powered detection of manipulated document images using U-Net segmentation.

Overview

CertiScan Backend is a FastAPI-based REST API for detecting document forgery and deepfake manipulation. Upload an image and receive a confidence score and visual mask showing potentially manipulated regions.

• Model: U-Net with EfficientNet-B0 encoder
• Framework: FastAPI + Uvicorn
• Deployment: Ready for Vercel serverless

Quick start (local development)

# Clone and navigate to project
cd backend

# Create and activate virtual environment
python -m venv .venv
. .venv/Scripts/Activate.ps1    # Windows PowerShell
source .venv/bin/activate       # macOS/Linux

# Install dependencies
pip install -r requirements.txt

# Run server
uvicorn main:app --reload

Server will be at http://127.0.0.1:8000

View interactive API docs: http://127.0.0.1:8000/docs

Deploying to Vercel

Deploy this backend as serverless functions to Vercel:

1. Push to GitHub
2. Connect repo to Vercel
3. Set environment variables (see .env.example):
   • UNET_MODEL_PATH — path to model checkpoint
   • CORS_ORIGINS — comma-separated frontend domains
4. Deploy — Vercel auto-detects vercel.json config

VERCEL_DEPLOYMENT.md — Complete setup guide with troubleshooting

Important Notes

• ⏱️ Vercel functions have a 60-second timeout (model loading takes 15-30s)
• 💾 No persistent storage — output masks are returned as encoded data
• 🌡️ Cold starts cause slowdowns — pre-warm with periodic health checks

See VERCEL_DEPLOYMENT.md for workarounds and best practices.

API usage

Check health

curl http://localhost:8000/health

Predict document authenticity

curl -F "file=@document.jpg" http://localhost:8000/predict

Response:

{
  "prediction": "Real",
  "confidence": 0.95,
  "threshold_used": 0.1,
  "model_version": "unet_finetuned_v2.pth",
  "mask_path": "outputs/abc123.png",
  "mask_url": "http://localhost:8000/outputs/abc123.png"
}

Environment variables

See .env.example for all available options:

• UNET_MODEL_PATH — path to model checkpoint
• CORS_ORIGINS — comma-separated list of allowed CORS origins
• ALLOW_UNTRAINED_MODEL — allow running without trained weights (dev only)

Project structure

.
├── backend/
│   ├── main.py                 # FastAPI app
│   ├── routes/predict.py       # Prediction endpoint
│   ├── model/loader.py         # Model loading
│   ├── utils/
│   │   ├── preprocess.py       # Image preprocessing
│   │   └── inference.py        # Model inference
│   ├── unet_finetuned_v2.pth   # Model weights
│   └── README.md               # Detailed backend docs
├── Procfile                    # Railway config
├── runtime.txt                 # Python version
├── Dockerfile                  # Container config
├── requirements.txt            # Python dependencies
└── .env.example               # Environment template

Testing

pip install pytest
pytest -q

Troubleshooting

See backend/README.md for detailed troubleshooting guide.

Common issues:

• Model not found: Set UNET_MODEL_PATH or use ALLOW_UNTRAINED_MODEL=1 (dev)
• CORS errors: Update CORS_ORIGINS environment variable
• Port in use: Change port in development or check for existing process

Contributing

• Create feature branches from main
• Keep commits focused and well-documented
• Add unit tests for new features
• Update requirements.txt when adding dependencies

License

[Add license info here]

---

For detailed backend documentation, see backend/README.md