Fraud-Detector: A Real-Time Fraud Detection System Based on XGBoost

Basic Project Information

• Original Author/Maintainer: Bumerdene073
• Source Platform: GitHub
• Original Link: https://github.com/Bumerdene073/fraud-detector
• Release Date: 2026-06-14

Core Highlights Fraud-Detector is an open-source real-time fraud detection tool based on the XGBoost model. It supports recognition of 6 common fraud patterns, has a response time under 100ms, provides explainable AI analysis and local Docker deployment solutions, balances accuracy and real-time performance, and is suitable for scenarios like finance and e-commerce.

Fraud-Detector focuses on real-time fraud detection, aiming to address the need for instant identification of fraudulent activities in scenarios like financial transactions and online payments. Its core advantages lie in combining high machine learning accuracy (XGBoost model F1 Score >91%) with real-time low latency (response time <100ms), while providing explainable AI analysis to help users understand the basis of fraud risk scores. The system uses FastAPI to build API interfaces and Docker containerization to ensure environment consistency and ease of maintenance.

Core Features

1. Real-Time Fraud Detection: Supports 6 patterns (abnormal amount, abnormal frequency, abnormal geographic location, abnormal time pattern, abnormal device fingerprint, deviation from behavior pattern).
2. Explainable AI Analysis: Provides fraud type classification, confidence score, and detailed description of suspicious activities.

Technology Stack and Advantages

Component	Technology Selection	Performance Indicator
Machine Learning Model	XGBoost	F1 Score >91%
API Framework	FastAPI	Response Time <100ms
Containerization	Docker	Environment Isolation, Fast Deployment

XGBoost Advantages: High accuracy, handling imbalanced data, feature importance analysis, regularization to prevent overfitting. FastAPI Design: Asynchronous processing supports high concurrency, automatic document generation, type safety, and data validation.

System Requirements

• Windows10+ (64-bit), 4GB+ RAM, dual-core 2GHz+ processor, 500MB available space, Docker Desktop installed.

Installation Steps

1. Download the latest version from GitHub Releases;
2. Install Docker Desktop;
3. Extract files to a local directory;
4. Start the service: cd fraud-detector && docker-compose up;
5. Visit http://localhost:8000.

Usage Methods

• File Upload: Supports CSV, Excel batch data;
• Manual Input: Enter transaction data one by one via the web interface;
• Detection results include fraud type summary, confidence score, and detailed explanation.

Fraud-Detector adopts a local deployment architecture to ensure data security:

1. Data Does Not Leave the Local Environment: Sensitive transaction data is processed locally and not uploaded to external servers;
2. Session Isolation: Each detection data is processed only in the current session;
3. Optional Log Upload: Logs are uploaded only when users actively choose to do so for troubleshooting. This design is suitable for industries with strict data privacy requirements like finance and healthcare.

1. Financial Transaction Monitoring: Banks and payment institutions integrate the system to assess transaction risks in real time, block high-risk transactions or trigger manual review;
2. E-commerce Platform Risk Control: Detect abnormal orders (brushing, malicious refunds, account theft, etc.) to protect the rights of merchants and consumers;
3. Enterprise Internal Audit: Scan financial data to identify internal fraudulent behaviors like abnormal expenditures, duplicate reimbursements, and fake invoices.

Project Value

• Out-of-the-box: Provides complete deployment documentation and pre-trained models to lower the threshold for use;
• Balances performance and accuracy: 91%+ F1 Score with <100ms response time, meeting production environment needs;
• Explainability: Helps business personnel understand model decisions and improves trust;
• Privacy-first: Local deployment ensures sensitive data security.

Summary Fraud-Detector is an excellent open-source starting point for quickly building fraud detection capabilities, suitable for teams needing real-time risk control.