💳 Real-Time Fraud Detection System

🚀 Live Demo: https://fraud-detection-api-xage.onrender.com/docs

A production-ready Fraud Detection REST API built using XGBoost and FastAPI, designed to minimize expected financial loss instead of maximizing raw accuracy.

This project demonstrates financial risk modeling, cost-sensitive machine learning, and cloud deployment.

---

🚀 Project Overview

This system is built using the Kaggle Credit Card Fraud dataset:

• 284,807 transactions
• 492 fraud cases
• 0.17% fraud rate (extreme class imbalance)

The model detects fraudulent transactions and recommends actions based on expected financial impact, not just probability.

---

🧠 Key Features

• Extreme class imbalance handling
• XGBoost with scale_pos_weight
• ROC-AUC and PR-AUC evaluation
• Cost-sensitive decision logic
• Expected loss modeling
• REST API with FastAPI
• Interactive Swagger documentation
• Cloud deployment ready

---

💰 Business Logic (Core Strength)

Instead of using a fixed probability threshold, the system compares:

Expected Loss if Allowed
= P(Fraud) × Fraud_Loss

Expected Loss if Blocked
= (1 − P(Fraud)) × False_Block_Cost

Assumptions:

• Fraud allowed → $1000 loss
• Legit transaction blocked → $10 inconvenience cost

The system selects the action that minimizes expected financial loss.

---

📊 Model Performance

• ROC-AUC: ~0.97+
• PR-AUC: ~0.85+
• Handles extreme 0.17% fraud imbalance
• Optimized for financial risk reduction

---

📂 Project Structure

fraud-detection-system/ │ ├── data/ # Dataset (ignored in GitHub) ├── src/ # Preprocessing & training scripts ├── models/ # Saved model artifacts ├── app/ # FastAPI application ├── requirements.txt └── README.md

---

🔌 API Usage

Run Locally

pip install -r requirements.txt
python -m uvicorn app.main:app --reload

Open:

http://127.0.0.1:8000/docs
POST /predict

Example Response:

{
  "fraud_probability": 0.87,
  "expected_loss_if_allowed": 870.0,
  "expected_loss_if_blocked": 1.3,
  "recommended_action": "Block Transaction 🚨"
}
🌐 Deployment

The API is cloud-deployable using:

Render

Railway

AWS

Azure

Docker containers

Swagger documentation available at:

https://fraud-detection-api-xage.onrender.com/docs
🛠 Tech Stack

Python

XGBoost

Scikit-learn

FastAPI

Uvicorn

Pandas

🏆 Why This Project Is Strong

This project demonstrates:

Financial risk modeling

Cost-sensitive ML optimization

Handling extreme imbalance datasets

Production-grade API design

Cloud deployment capability

Business-driven decision systems

👩‍💻 Author

Soumyasree Mitra

Aspiring Machine Learning Engineer