End-to-End Real-time Architecture for Fraud Detection in Online Digital Transactions

Abstract

The banking sector is witnessing a fierce concurrence characterized by changing business models, new entrants such as FinTechs, and new customer behaviors. Financial institutions try to adapt to this trend and invent new ways and channels to reach and interact with their customers. While banks are opening their services to avoid missing this shift, they become naturally exposed to fraud attempts through their digital banking platforms. Therefore, fraud prevention and detection are considered must-have capabilities. Detecting fraud at an optimal time requires developing and deploying scalable learning systems capable of ingesting and analyzing vast volumes of streaming records. Current improvements in data analytics algorithms and the advent of open-source technologies for big data processing and storage bring up novel avenues for fraud identification. In this article, we provide a real-time architecture for detecting transactional fraud via behavioral analysis that incorporates big data analysis techniques such as Spark, Kafka, and h2o with an unsupervised machine learning (ML) algorithm named Isolation Forest. The results of experiments on a significant dataset of digital transactions indicate that this architecture is robust, effective, and reliable across a large set of transactions yielding 99% of accuracy, and a precision of 87%.