Bank-to-bank messaging network Swift has reported promising results from tests combining AI and cross-border data sharing, successfully doubling real-time fraud detection during trials involving 13 leading banks and ten million test transactions.
The experiments employed privacy-enhancing technologies (PETs) to facilitate secure sharing of fraud insights across borders. One notable application allowed participants to verify intelligence on suspicious accounts in real-time, potentially expediting the identification of complex international financial crime networks and preventing fraudulent transactions.
In another scenario, participants utilized both PETs and federated learning—an AI approach that enables local data training without sharing customer information—to identify anomalous transactions. Models trained on synthetic data from ten million artificial transactions proved to be twice as effective in detecting known fraud cases compared to those trained on individual institution datasets.
Financial institutions involved in these trials include ANZ, BNY Mellon, and Intesa Sanpaolo, along with technology partners like Google Cloud.
Rachel Levi, head of AI at Swift, stated, “The industry loses billions to fraud each year, but by enabling the secure sharing of intelligence across borders, we’re paving the way for a significant reduction in this figure and allowing fraud to be stopped in a matter of minutes, not hours or days.”
Following these successful trials, Swift plans to broaden participation before launching a second phase that will utilize real transaction data to demonstrate the technologies’ effectiveness against real-world fraud.
The cooperative is actively exploring AI’s potential to address challenges in cross-border payments and currently has over 50 use cases spanning proof of concept, pilots, and live applications. Earlier this year, Swift launched an AI-enhanced Payments Controls Service to help small and medium-sized financial institutions accurately flag suspicious transactions in real time.