🛡️ Digital Public Safety AI Platform

AI-powered platform to detect and disrupt digital arrest scams, counterfeit currency, fraud networks, and crime hotspots - built for the "AI for Digital Public Safety" hackathon challenge.

---

📖 Overview

India registered 1.14 million cybercrime complaints in 2023, and "digital arrest" scams alone defrauded citizens of over ₹1,776 crore in the first nine months of 2024. This platform tackles the problem with five integrated AI modules - shifting law enforcement and citizens from reactive investigation to proactive threat detection.

✨ Features

Module	What it does
🚨 Digital Arrest Scam Detection	Analyzes call transcripts, WhatsApp/SMS/email text → scam probability, risk level, explanation, recommended action
💵 Counterfeit Currency Detection	Computer vision analysis of currency note images → genuine/counterfeit verdict, confidence score, highlighted suspicious regions
🕸️ Fraud Network Graph Intelligence	Maps phone/device/account linkages from transaction data → fraud cluster detection, interactive graph visualization
🗺️ Geospatial Crime Intelligence	Plots fraud incidents on a map → hotspot detection, district-wise risk scoring
🤖 Citizen Fraud Assistant	Multilingual chatbot - explains scams, analyzes suspicious messages, guides reporting to cybercrime.gov.in / 1930

🔐 Access Control (Roles)

Access is gated by a JWT login and enforced server-side - the role is not a client-side toggle. Sign in decides which modules you can reach:

Role	Login	Can access
Police	police / police123	All modules
Bank	bank / bank123	Dashboard, Scam Checker, Currency Checker
Citizen	citizen / citizen123	Dashboard, Scam Checker, Citizen Assistant

Currency (/currency/*) requires police or bank; Fraud Graph (/graph/*) and Geo Intel (/geo/*) are police-only - other roles get an HTTP 403. Auth behaviour is configurable via env vars (JWT_SECRET_KEY, JWT_EXPIRE_HOURS, REQUIRE_AUTH).

🏗️ Architecture

┌─────────────────────────────────────────────────────────┐
│                     React Frontend (Vite)                │
│  Dashboard | Scam Checker | Currency Scan | Graph | Map   │
└───────────────────────────┬────────────────────────────┘
                             │ REST (axios)
┌───────────────────────────▼────────────────────────────┐
│                    FastAPI Backend (Uvicorn)              │
│  /scam  /currency  /graph  /geo  /chat   routers          │
├─────────────┬───────────┬───────────┬───────────┬────────┤
│ Scam ML     │ CV Model  │ NetworkX  │ Geo utils │ LLM API │
│ (sklearn)   │ (OpenCV)  │ (fraud    │ (heatmap/ │ (Groq)  │
│             │           │  clusters)│ hotspot)  │         │
└─────────────┴───────────┴───────────┴───────────┴────────┘
                             │
                    ┌────────▼────────┐
                    │  SQLite (app.db)  │
                    │ scams, currency,  │
                    │ transactions,     │
                    │ incidents         │
                    └───────────────────┘

🧰 Tech Stack

Layer	Technology	Why
Frontend	React + Vite + Tailwind CSS	Fast dev, minimal config
Backend	FastAPI + Uvicorn	Async, auto Swagger docs, Python-native
Database	SQLite (SQLAlchemy ORM)	Zero setup, easy Postgres upgrade path
Scam Detection	TF-IDF + Scikit-learn Logistic Regression + keyword rules	Lightweight, explainable, fast
Currency Detection	OpenCV heuristics (edge/sharpness/color/texture analysis)	Explainable, no training data required
Graph Intelligence	NetworkX + Pyvis	Fraud ring clustering + interactive visualization
Geospatial	React-Leaflet + OpenStreetMap	Free, no API key required
Chatbot	Groq API (Llama 3.1 8B Instant)	Fast, free-tier, multilingual

📁 Project Structure

digital-safety-ai/
├── backend/
│   ├── app/
│   │   ├── main.py
│   │   ├── database.py
│   │   ├── models.py
│   │   ├── routers/
│   │   │   ├── scam.py
│   │   │   ├── currency.py
│   │   │   ├── graph.py
│   │   │   ├── geo.py
│   │   │   └── chat.py
│   │   ├── ml/
│   │   │   ├── scam_classifier.py
│   │   │   ├── train_scam_model.py
│   │   │   ├── currency_model.py
│   │   │   ├── graph_engine.py
│   │   │   ├── chat_service.py
│   │   │   └── artifacts/          # trained model files
│   │   ├── utils/
│   │   │   └── geo_utils.py
│   │   └── static/                 # generated graph HTML
│   ├── requirements.txt
│   └── .env
└── frontend/
    ├── src/
    │   ├── App.jsx
    │   ├── api.js
    │   └── pages/
    │       ├── Dashboard.jsx
    │       ├── ScamChecker.jsx
    │       ├── CurrencyChecker.jsx
    │       ├── FraudGraph.jsx
    │       ├── GeoIntel.jsx
    │       └── CitizenAssistant.jsx
    ├── package.json
    └── tailwind.config.js

🚀 Getting Started

Prerequisites

• Python 3.10+
• Node.js 18+
• A free Groq API key (for the chatbot module)

1. Clone the repo

git clone https://github.com/<your-username>/digital-safety-ai.git
cd digital-safety-ai

2. Backend setup

cd backend
python3 -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate

pip install -r requirements.txt

Create a .env file in backend/ (all values are optional - the app boots without them; only the chatbot needs the Groq key):

GROQ_API_KEY=your_free_groq_key_here     # optional - chatbot degrades gracefully if unset
DATABASE_URL=sqlite:///./app.db          # swap for postgresql://… in production
JWT_SECRET_KEY=change-me-in-production    # optional - a default is used otherwise
JWT_EXPIRE_HOURS=8                        # optional
# ALLOWED_ORIGINS=http://localhost:5173   # optional CORS allow-list
# REDIS_URL=redis://localhost:6379/0      # optional - falls back to in-memory cache

Note: the counterfeit-currency CNN is optional and not installed by default (it stays disabled and falls back to OpenCV heuristics). To enable it, install torch/torchvision for your platform and drop trained weights into backend/app/ml/weights/.

Train the scam classifier (one-time):

python -m app.ml.train_scam_model

Run the backend:

uvicorn app.main:app --reload

API docs available at → http://127.0.0.1:8000/docs

3. Frontend setup

cd frontend
npm install
npm run dev

App available at → http://localhost:5173

🔌 API Reference

Method	Endpoint	Role	Description
POST	/auth/login	public	Exchange username/password for a JWT
GET	/auth/me	any	Current token payload
GET	/dashboard/stats	any	Aggregated dashboard statistics (cached)
POST	/scam/analyze	any	Analyze text for scam indicators
GET	/scam/history	any	Recent scam analysis history
POST	/scam/report/{id}	any	Generate NCRB-style PDF for a scam record
POST	/currency/analyze	police, bank	Upload currency note image for verification
GET	/currency/history	police, bank	Recent currency check history
POST	/graph/ingest	police	Ingest transaction records
GET	/graph/analyze	police	Get fraud clusters
GET	/graph/view	public	Interactive fraud network graph (HTML, iframe-embeddable)
POST	/graph/query	police	Query the graph - COUNT, SUSPICIOUS, HUBS, or a node id
POST	/geo/ingest	police	Ingest fraud incident coordinates
GET	/geo/heatmap	police	Raw incident points
GET	/geo/hotspots	police	Detected crime hotspots
GET	/geo/district-risk	police	District-wise risk scoring
POST	/audio/analyze	any	⚠️ Demo stub - simulated deepfake result
POST	/chat/message	any	Chat with the citizen fraud assistant (rate-limited)

🧪 Testing

Use the Swagger UI at /docs to test each endpoint interactively, or use the frontend pages directly:

Page	Route
Dashboard	/
Scam Checker	/scam
Currency Checker	/currency
Fraud Graph	/graph
Geo Intelligence	/geo
Citizen Assistant	/assistant

⚠️ Known Limitations

• Scam classifier trained on a small synthetic dataset - production use would need a larger labeled corpus of real scam transcripts.
• Currency detection uses classical CV heuristics (no labeled FICN dataset available) rather than a trained CNN - a documented, explainable trade-off for hackathon scope. Swapping in a fine-tuned CNN is a planned improvement.
• Audio deepfake detection is a demo stub that returns a simulated result - it is clearly labeled as such in the UI and API and is not real spoofing detection.
• SQLite is used for simplicity - swap DATABASE_URL for PostgreSQL in production.
• Auth ships with in-source demo users and a default JWT secret - set JWT_SECRET_KEY and back it with a real user store for production.

🔮 Future Improvements

• Fine-tuned CNN for counterfeit currency detection on labeled FICN datasets
• Real-time call/voice deepfake detection (speech AI)
• Direct NCRB / cybercrime.gov.in reporting integration
• Neo4j/graph database for larger-scale fraud network analysis
• Mobile app + WhatsApp/IVR integration for the citizen assistant
• Court-admissible intelligence report export (PDF)

👥 Team

Built for the AI for Digital Public Safety hackathon challenge - Defeating Counterfeiting, Fraud & Digital Arrest Scams.

📄 License

MIT License - see LICENSE for details.