Privacy-Preserving Federated Learning for Detecting Fraudulent Financial Transactions in Japanese Banks
Sachiko Kanamori, Taeko Abe, Takuma Ito, Keita Emura, Lihua Wang, Shuntaro Yamamoto, Le Trieu Phong, Kaien Abe, Sang‐Wook Kim, Ryo Nojima, Seiichi Ozawa, Shiho Moriai
Abstract
To tackle financial crimes including fraudulent financial transactions (FFTs), money laundering, illegal money transfers, and bank transfer scams, several attempts have been considered to employ artificial intelligence (AI)-based FFT detection systems, particularly, deep learning-based ones. However, to the best of our knowledge, no federated learning systems using real transaction data among financial institutions have been implemented so far. This is because there are several issues to be addressed as follows: (1) it is difficult to prepare sufficient amount of transaction data for training by one financial institution (e.g., a local bank), and a small amount of dataset does not promise high accuracy for detection, (2) each transaction data contains personal information, and thus it is restricted by Act on the Protection of Personal Information in Japan to provide the transaction data to a third party. In this paper, we introduce out demonstration experimental results of privacy-preserving federated learning with five banks in Japan: the Chiba Bank, Ltd., MUFG Bank, Ltd., the Chugoku Bank, Ltd., Sumitomo Mitsui Trust Bank, Ltd., and the Iyo Bank, Ltd. As the underlying cryptographic tool, we proposed a privacy-preserving federated learning protocol which we call Deepprotect, for detecting fraudulent financial transactions. Briefly, Deepprotect allows parties to execute the stochastic gradient descent algorithm using a set of techniques for the privacy-preserving deep learning algorithms (IEEE TIFS 2018, 2019). In the demonstration experiments, we built machine learning models for detecting two types of financial frauds — detecting fraudulent transactions in customers/victims' accounts and detecting criminals' bank accounts. We show that our federated learning system detected FFTs that could not be detected by the model built using the dataset from a single bank and detected criminals' bank accounts before the bank actually froze them.