U.S.-U.K. PETs Prize Challenge: Anomaly Detection via Privacy-Enhanced Federated Learning.

Area of Science:

• Computer Science
• Cybersecurity
• Data Science

Background:

• Collaborative analytics using large datasets is crucial in finance, healthcare, and national security.
• Privacy Enhancing Technologies (PETs) are vital for enabling data collaboration without compromising privacy.
• The U.S. and U.K. governments launched a PETs prize challenge in 2021 to address financial crime and pandemic response.

Purpose of the Study:

• To present the Rutgers ScarletPets approach for privacy-preserving federated learning.
• To identify anomalous financial transactions within a payment network system (PNS).
• To develop a solution for the PETs prize challenge focused on financial crime prevention.

Main Methods:

• A two-step anomaly detection methodology was employed.
• Features were mined from account-level data and labels.
• A privacy-preserving encoding scheme augmented features within the PNS.
• The PNS learned a highly accurate classifier from the augmented data using federated learning.

Main Results:

• The ScarletPets approach achieved no noteworthy drop in accuracy compared to centralized methods.
• The federated learning model demonstrated high accuracy in identifying anomalous financial transactions.
• The solution won first prize in the U.S. PETs prize challenge for its privacy, utility, efficiency, and flexibility.

Conclusions:

• The Rutgers ScarletPets approach offers a flexible and accurate method for privacy-preserving anomaly detection in financial networks.
• This federated learning model enables continuous improvement without additional computational or privacy burdens on participating banks.
• The approach effectively addresses the need for secure collaborative analytics in sensitive domains like finance.