Federated Tabular Data Synthesis A Collaboration between Banks

Important results from the project

The project has had two objectives: Generate high-fidelity synthetic data. Explore federated model training between banks. We have developed an open source tool to create synthetic transaction networks, adapted to real data or known performance metrics. This enables systematic work to improve anti-money laundering. The tool has been used for federated learning between two banks. Although the results so far are limited, the project has laid an important foundation for further studies.

Expected long term effects

Today, the ability to detect money laundering is limited. i) Banks rarely cooperate and have different risk strategies, which results in different approaches. ii) Regulators lack tools to assess effectiveness – sanctions are almost exclusively related to KYC. With synthetic data, banks and regulators can discuss transaction monitoring on a common, non-sensitive basis. This could be a key to increased collaboration.

Approach and implementation

The project was driven by weekly meetings to track progress and address issues. AI Sweden led the development of the data generator in close collaboration with the banks, who provided domain expertise. Major challenges were addressed through dedicated workshops. Federated learning was developed jointly with the banks’ data science, IT, and innovation teams to meet technical and security needs. We also collaborated with Scaleout AB, a startup focused on FL.

External links

Verktyg för att generera syntetiska transaktionsdata