Federated embedding models for privacy-preserving anomaly detection in bank transactions

The Featurespace project team will develop a transformational federated solution for financial crime prevention that leverages insights across financial institutions (FI), whilst guaranteeing that personal information of individuals and accounts stays private to their trusted FI. Current dominant approaches in the industry rely on machine learning models that are trained on datasets private to each bank, or consortium models that generalise insights derived from dominant FIs.

These approaches are limited in missing the bigger picture in increasingly faster and cross-border payments, or may be vulnerable to privacy attacks.

The project will deliver an innovative collaborative deep learning solution that builds over existing in-house technologies, currently protecting major banks, payment processors and payment infrastructure providers worldwide. The system will take a hybrid approach to input and output privacy preservation, using a protective architectural design in combination with de-facto private data publication technologies, including de-identification and differential privacy.

In this setting, data processing pipelines will be distributed across participants in a federated system, avoiding parameter-pooling or score-averaging, and producing unified outputs that jointly contribute towards combining originator and beneficiary insights in financial transactions. The proposed system significantly reduces the risk of malicious actors being able to reverse-engineer privately identifiable information about individual account holders.

The result is a flexible collaborative learning system, which effectively balances confidentiality requirements and privacy preferences of banking institutions, along with effective financial crime detection.

The project team will place special focus on developing a system that reduces barriers for adoption across a broad spectrum of institutions and is able to deal with the large variety in real-world data formats and data quality observed across FIs. The solution will have limited expectations on the format of data inputs and is well-positioned to be productionised in modern enterprise financial crime management platforms.

This project builds on Featurespace's experience in delivering fraud and anti-money laundering solutions to major banks and payment infrastructure providers in the UK and around the world, as well as the research of Prof Derek McAuley (University of Nottingham) and Prof Richard Mortier (University of Cambridge) on user-centric systems and security undertaken at the Horizon Digital Economy Research centre.