IbrahimDammak/tunisian_id_fraud

# Tunisian CIN Fraud Detection A modular pipeline that decides whether a photograph of the **front side** of a Tunisian national ID card (CIN — *Carte d'Identité Nationale*) is genuine, a **presentation attack** (printed copy, screen replay, re-laminated card), or a **digital injection** (face swap, text tampering, splicing). Ships as a FastAPI service with a React capture UI and an admin dashboard. > **Status: validated prototype, not production.** It works end to end and its > core module is measured against ISO/IEC 30107-3, but it is trained on a very > small dataset and several thresholds are uncalibrated. Read > Honest status before building on it. --- ## Table of contents - Quickstart - Architecture - How a card is scored - API - Project layout - Results - Honest status - Where to pick up - Data and privacy - References --- ## Quickstart **Requirements:** Python 3.10 (developed on 3.10.6), Node 18+, ~4 GB disk for model weights and dependencies. CPU-only is fine; everything runs without a GPU. ```bash python -m venv venv venv\Scripts\activate pip install -r requirements.txt ``` ```bash cd frontend && npm install && npm run build && cd .. ``` ```bash python -m uvicorn api.main:app --host 0.0.0.0 --port 8000 ``` Then open: | URL | What it is | |---|---| | `localhost` | Capture UI — guided camera, auto-capture, verdict | | `localhost` | Admin dashboard — past runs, scores, evidence images | | `localhost` | Interactive OpenAPI docs | **First run is slow** — EasyOCR downloads its Arabic model and `transformers` downloads `facebook/dinov2-large` (~1.2 GB). Both are cached afterwards. > ⚠️ **Use the project venv.** The shipped `.pkl` models are pickled with > **scikit-learn 1.7.2**. Under 1.9.x they still load but emit *"may lead to > invalid results"*. `requirements.txt` pins this — don't loosen it without > retraining. Verify with: ```bash python -c "import sys, sklearn; print(sys.executable, sklearn. …