Neurashield: A GNN Based Approach to Credit Card Fraud Detection

Credit card fraud has been increasing with the rise of technology. This has become a major issue for institutions and individuals, leading to substantial economic losses and diminished trust in credit card transactions, especially which occurs online. Due to the rise of fraudsters who use new technology and ways to do fraudulent activities, traditional fraud detection methods are less effective. The existing techniques and approaches that have been implemented have often failed to identify complex patterns and relationships of fraudulent activities in this evolving world. The current machine-learning and deep learning studies have shown potential limitations due to the use of older datasets with less data variety while current Graph Neural Network (GNN) models use complex architectures and high number of features. Neurashield is a novel Edge Graph Convolutional Network (Edge-GCN) approach to enhance the credit card fraud detection using latest dataset and less number of features. NeuraShield shows the ability to analyse the complex patterns and relationships of data by representing transactions as a graph, with nodes indicating cards and merchants while edges representing relationships. This approach first preprocesses the data, then applies feature engineering and at last develops a highly accurate GNN model which was trained for different class ratios using a real-world mimicking credit card transaction dataset from IBM that has 24,386,900 unique credit card transactions. The results show that the Edge-GCN model performed better compared to the Multi-Layer Perceptron (MLP) model and also showed an increment of accuracy when the size of the dataset was increased while addressing class imbalance. The final results were captured as a testing accuracy of 80.67%, 0.7138 of F1 score, and 0.8691 of ROC AUC. The evaluation shows that NeuraShield has the potential to detect more complex credit card fraud activity with a less information, than the traditional methods.