OpenAI Fraud Detection Analysis System

📋 Overview

This system is an end-to-end R-based pipeline designed to simulate, analyze, and detect fraudulent patterns in AI API usage. [cite_start]It generates a synthetic dataset of 10,000 records, performs Exploratory Data Analysis (EDA), and trains Machine Learning models to identify various types of abuse.

🚀 Features

• [cite_start]Synthetic Data Generation: Creates 10,000 records with features like API call frequency, token usage, unique IPs, and geographic risk factors.
• [cite_start]Risk Scoring Engine: Implements a weighted heuristic to label fraud based on behavioral thresholds (e.g., failed auth attempts, rate limit violations).
• [cite_start]Exploratory Data Analysis: Visualizes fraud distribution, risk score frequency, and usage patterns using ggplot2.
• Machine Learning Models:
  • [cite_start]Random Forest: A high-accuracy ensemble model for robust detection.
  • [cite_start]Decision Tree: An interpretable model using rpart to understand the logic behind fraud flags.
• [cite_start]Real-Time Detection Function: A utility function to input live metrics and receive an immediate risk assessment (Low, Medium, or High Risk).

🛠 Tech Stack

• [cite_start]Language: R
• Core Libraries:
  • [cite_start]tidyverse & lubridate: Data manipulation and temporal analysis.
  • [cite_start]caret, randomForest, rpart: Machine learning and model evaluation.
  • [cite_start]ggplot2, plotly, corrplot: Advanced data visualization.

📁 Project Structure

The script is organized into 10 logical sections:

1. [cite_start]Data Generation: Building the synthetic user base.
2. [cite_start]Labeling: Creating fraud categories based on risk factors.
3. [cite_start]Summary Statistics: Initial dataset audit.
4. [cite_start]Visualization: Plotting usage patterns and risk scores.
5. [cite_start]Correlation: Mapping feature relationships.
6. [cite_start]Modeling: Splitting data (70/30) and training Random Forest.
7. [cite_start]Comparison: Evaluating models against Decision Trees.
8. [cite_start]Real-Time Logic: The detect_fraud function.
9. [cite_start]Usage Examples: Testing legitimate vs. high-risk scenarios.
10. [cite_start]Persistence: Saving results to .rds and .csv files.

📊 Fraud Types Detected

The system classifies fraudulent behavior into specific categories:

• [cite_start]API Abuse: Excessive calls per hour.
• [cite_start]Account Takeover: Rapid changes in location and high number of unique IPs.
• [cite_start]Credential Stuffing: Repeated failed authentication attempts.
• [cite_start]Content Fraud: High suspicious keyword counts and content similarity.
• [cite_start]Payment Fraud: Frequent changes to payment methods.

🚦 Getting Started

Prerequisites

Install the required R packages:

install.packages(c("tidyverse", "caret", "randomForest", "e1071", "ggplot2", 
                   "plotly", "lubridate", "scales", "corrplot", "rpart", "rpart.plot"))