Scalable and Privacy-Preserving Financial Fraud Detection
Status:
The project started in February 2026 and is still ongoing.
Researchers:
Industry Partner:
Swissquote
Description:
Financial institutions are legally obligated to screen customers and transactions for fraud under anti-money-laundering regulations. Fraud markers identified by one bank are valuable to others, since illicit funds are typically funneled through accounts at multiple institutions, but banks rarely share these markers with their peers. Banks would be willing to share fraud markers under strict data minimization properties: a bank should learn only which of its own events match, without learning the broader marker set, and without transmitting any of its transaction data to third parties. Private set intersection protocols meet the core matching requirement but are inherently interactive, requiring the receiving bank to send an encoding of its transactions to the marker holder and thereby leaking how often and when it screens against shared markers. Non-interactive alternatives, such as distributing hashed markers, avoid this leakage but cannot currently bound how many times a marker may be matched or let issuing banks revoke markers after distribution.
This project designs and analyzes non-interactive protocols for cross-institutional fraud-marker sharing. It aims to construct mechanisms that let a bank locally match its transactions against externally issued fraud markers, with no exchange of transaction data and no interaction-based leakage, while enforcing a bound on how many times each marker can be matched and allowing issuers to revoke markers after the fact. The resulting privacy-preserving protocols should meet the latency and scale requirements of real-world transaction volumes.