AI-Based Fraud Detection Using Graph Neural Networks

Abstract

Fraud detection is an important problem in today’s digital world due to the rapid increase in 
online transactions. Traditional methods often fail to identify complex patterns of fraud. This project 
proposes an AI-based fraud-detection system using Graph Neural Networks (GNNs) to improve accuracy. 
In this system, transaction data is represented as a graph where nodes represent users or accounts and 
edges represent transactions between them, and a GraphSAGE-based GNN analyses these relationships to 
detect unusual patterns and identify fraudulent activities effectively. The system is implemented in Python 
using PyTorch and PyTorch Geometric for the GNN, with NetworkX for graph construction, scikit-learn 
for preprocessing, and a Streamlit interface for visualisation and prediction. It models a user–merchant 
transaction graph for edge-level fraud classification, incorporates realistic fraud indicators such as 
unusual transaction timing, device changes, and abnormal behaviour, supports uploading real UPI 
transaction CSVs or using built-in synthetic data, and displays risk scores together with explanations. The 
approach is designed to handle large amounts of data and provide faster, more reliable fraud detection, 
reducing false alerts while strengthening decision-making in real-time financial systems. The prototype 
was validated through functional test cases for dataset upload, missing-value handling, model training, 
fraud prediction, and visualisation, all of which passed. Overall, the system demonstrates how graph-based 
deep learning can significantly enhance fraud detection, making it more robust, scalable, and suitable for 
real-world financial applications.