Financial fraud costs the global economy billions annually. Traditional fraud detection systems primarily rely on tabular data—analyzing individual transactions independently—or rigid rule-based systems. These methods often fail to capture the complex relationships, such as layering and collusion, that are indicative of organized fraud rings.

This project studies financial fraud detection on two graph-based datasets:

1. The Elliptic Bitcoin transaction graph
2. The Elliptic++ transaction and address graph

In this project we compare a classical tabular baseline against graph neural network models and see whether richer graph structure improves illicit transaction detection.