Paving the way for secure and private financial crime detection

This spring, the US and UK governments jointly hosted the ‘Summit for Democracy’ – aimed at promoting and protecting democratic values. A part of this summit was the PETs Prize Challenge, designed to develop privacy-enhancing technologies to protect users in our data-driven society whilst also tackling the challenges of that increasingly digital society, such as financial fraud. In this international competition, the team of TU Delft won the second prize, developing the most effective model for detecting financial fraud such as money laundering.

Privacy-enhancing technologies are critical in today's data-driven society, where enormous amounts of data are being collected, processed, and shared every day. PETs are designed to enhance privacy and security by protecting this sensitive data, while allowing useful information to be extracted from it. PETs can thus help us analyse privacy-sensitive networks such as international finance and biomedical research data, while preserving privacy and ensuring confidentiality.

The international challenge was initiated to lay the groundwork for useful new technologies in this field. The team in which the TU Delft partook – called PPMLHuskies, with Delft researchers Zeki Erkin, Jelle Vos and Célio Porsius Martins participating – developed a privacy-preserving machine learning model to detect anomalous payments in the international SWIFT banking system. By combining this machine-learning model with a new cryptographic protocol, they could safely perform computations over the synthetic SWIFT dataset. By focussing primarily on spotting the transactional loops of money laundering – laundered money tends to end up back in the same space – they could detect anomolies without unencrypting the data of private citizens. Additionally, to prevent the machine-learning model from memorizing instances from the training data, thus making it possible to deduce sensitive information, the team trained the model with an algorithm that ensures differential privacy.