Feds pitch anti-fraud AI trained on COVID loan data

"The Engine was built with a bunch of modular components, including unsupervised machine learning models used to detect anomalies, supervised ML that identifies patterns consistent with pandemic fraud cases, and rules-based flags that catch invalid Social Security and employer identification numbers. According to Dieffenbach, such small anomalies can often identify hidden connections, like shared bank account numbers among supposedly independent applicants, that suggest fraud."

"A fraud-detection AI model trained on COVID-19 loan data could have flagged potentially tens of billions of dollars in payments before they went out, reducing the feds' pay-and-chase cleanup, the US government's Pandemic Response Accountability Committee told Congress on Tuesday. "The time is now to use this data to prevent fraud schemes before taxpayer dollars are lost and hold wrongdoers accountable," Dieffenbach told the Committee. And it can do it quickly: Dieffenbach said the AI can process 20,000 applications per second."

PRAC developed a Fraud Prevention Engine using roughly five million Small Business Administration COVID-19 Economic Injury Disaster Loan applications as training data. The model could have flagged potentially tens of billions of dollars in suspect payments prior to disbursement, reducing reliance on post-payment recovery. The system combines unsupervised anomaly detection, supervised machine learning for known fraud patterns, and rules-based checks for invalid Social Security and employer identification numbers. Small anomalies can reveal hidden links such as shared bank accounts among separate applicants. The prototype processes about 20,000 applications per second and demonstrates potential to prevent future government fraud.

Read at Theregister