Decentralised finance has seen explosive growth, but this has introduced new challenges in detecting illicit activity on public blockchains. My dissertation explored whether it's possible to build a real-time risk assessment system for DeFi transactions using Graph Neural Networks to identify suspicious patterns before transactions are processed. I analysed over 70 major crypto exploits and developed a labelled dataset of laundering behaviours, including mixer usage and peeling chains. Several machine learning models were tested, with Graph Isomorphism Networks showing the strongest performance. While results were promising, they also revealed practical limitations: even small false positive rates could disrupt millions of legitimate daily transactions. As a result, I propose hybrid AI-human systems and post-transaction monitoring as more viable near-term solutions. This talk will walk through my journey building the system, what worked, what didn’t, and the future of ML for DeFi compliance and blockchain security.
