🏦 FRAUDSENSE : Advancing Fraud Detection through Tabular GAN Synthetic Transactions

This project focuses on enhancing fraud detection systems by generating privacy-preserving synthetic bank transactions . The synthetic data augments real-world transactions to address class imbalance and data scarcity, leading to more robust and generalizable fraud detection models.

Generated privacy-preserving synthetic bank transaction data [ to improve fraud detection research and model benchmarking.

---

Website Link (Live Action): FraudSense

📁 Folder Structure

Synthetic-Fraud-AI-Project/
│
├── README.md
├── Final_Report.pdf
├── Fraud_Detection_Slides.pptx
│
├── Model/
│   ├── preprocess_pipeline.pkl          
│   ├── best_model_augmented.pkl  
│   └── catboost_best_model.pkl  
│
├── Data/
│   ├── Bank_Transaction.csv          
│   ├── Augmented_data.csv       
│   └── Synthetic_Bank_Data.csv     
│
└── Notebooks/
    ├── Exploratory Data Analysis (EDA).ipynb                
    ├── Synthetic_Data_Analysis.ipynb     
    ├── Model_Training.ipynb    
    ├── Visualization.ipynb   
    ├── app.py   
    ├── preprocess.py   
    └── catboost_info/
          ├── catboost_training.json
          ├── learn_error.tsv
          └── time_left.tsv

🎯 Objective

The main goal of this project is to:

* Generate synthetic financial transactions using CTGAN and Finance TVAE.

* Combine synthetic and real data to balance the dataset and reduce bias.

* Train and evaluate multiple machine learning models to measure how synthetic augmentation impacts fraud detection performance.

* Ensure data privacy and integrity by replacing sensitive data with realistic synthetic equivalents.

---

🧩 Problem Definition

Dataset

• File: Data/Bank_Transaction.csv

• Target Variable: Is_Fraud

  • 0 → Legitimate Transaction

  • 1 → Fraudulent Transaction

• Class Distribution:

  • Legitimate: 189,912 (≈94.96%)

  • Fraudulent: 10,088 (≈5.04%)

This severe imbalance leads to poor model recall and limited fraud detection capability.

Features Overview

Categorical Columns:

- Customer_ID, Gender, City, Bank_Branch, Account_Type, Transaction_ID,
Transaction_Type, Merchant_Category, Transaction_Device, Transaction_Location,
Device_Type, Transaction_Currency, Customer_Email, Transaction_Description, etc.

Numerical Columns:

Age, Transaction_Amount, Account_Balance

🧠 Approach & Methodology

1. Data Preprocessing

   • Handled missing values, encoding, and scaling.

   • Split into train-test sets using Stratified Sampling to preserve class ratios.

   • Implemented a custom preprocessing pipeline (preprocess_pipeline.pkl).

2. Synthetic Data Generation

   • Used EVAE Tabular to generate realistic synthetic transactions.

   • Merged synthetic and real data to create the Augmented Dataset.

   • Evaluated synthetic quality metrics (score ≈ 8.5) for realism and privacy.

3. Model Training and Tuning

   • Models were trained on both original and augmented datasets.

   • Hyperparameter tuning performed via RandomizedSearchCV .

   • Evaluation metrics: Accuracy, Precision, Recall, and F1-score.

---

🤖 Top 3 Performing Models (on Augmented Data)

Rank	Model	F1 Score	Recall	Precision
🥇 1	CatBoost Classifier	0.2868	0.8807	0.1713
🥈 2	XGBoost Classifier	0.2861	0.8302	0.1729
🥉 3	LightGBM Classifier	0.2667	0.5772	0.1734

---

🔍 Key Findings

* CatBoost and XGBoost achieved highest recall (>0.83) — ideal for minimizing missed frauds.

* LightGBM offered a better balance between recall and precision, making it suitable for production scenarios.

* Linear models (Logistic Regression, SGD) performed decently but struggled with extreme imbalance.

* Balanced Random Forest, despite high accuracy, failed to detect minority class effectively.

---

💻 Technologies Used

Category	Tools / Libraries
Data Processing	pandas, numpy, scikit-learn, joblib
Synthetic Data Generation	Tabluar CTGAN, gretel-synthetics
Machine Learning	XGBoost, LightGBM, CatBoost, DecisionTree, LogisticRegression, SGDClassifier
Visualization	matplotlib, plotly, seaborn
App Interface	Streamlit
Environment	Python 3.11, Anaconda, Jupyter Notebooks, VS Code

---

🚀 Results & Conclusion

* Synthetic data generation using CTGAN improved fraud recall by over 40% compared to models trained on raw data alone.

* CatBoost emerged as the most recall-efficient model for fraud detection.

* Combining real and synthetic datasets improved robustness, diversity, and fairness in model training.

* This approach can serve as a framework for banks and financial institutions seeking privacy-preserving data for fraud detection research and benchmarking.

---

👨‍💻 Author

Payal Dhokane
📍 B.E CS, International Centre of Excellence in Engineering & Management
🔗LinkedIn Profile
📧 payaldhokane282@gmail.com
Kaggle: (https://www.kaggle.com/payaldhokane)