Permissioned Blockchain-Based XGBoost for Multi Banks Fraud Detection

Abstract

Fraud detection is one of the financial institution problems which can utilize Machine Learning (ML). However, the fraud activity is hard to detect since the occurrence is relatively low compared to the actual transaction. Several banks can collaborate to gather more fraudulent transactions from their data. However, the collaboration can cause data leakage from each bank, where the customer data should be confidential. Decentralized ML is one of the approaches to tackle the privacy-preserving aspect. This work proposed a fully decentralized environment using a permissioned blockchain to detect multiple banks’ fraud. The training process utilizes a continual eXtreme Gradient Boosting (XGBoost) model. We provided the architecture of blockchain implementation for multiple banks, where it is conducted as batch and streaming data processing. As we compared our approach with the centralized, individual, and federated GBDT models, it maintains a good prediction performance and fulfills the environment of a fully distributed system.

This work was supported in part by Japan Society for the Promotion of Science (JSPS) KAKENHI under Grant JP20K11826, and in part by Japan Science and Technology Agency (JST) AIP Accelerated Program under Grant JPMJCR22U5.