Fraudhunter : a supervised fraud detection tool for Internet banking transactions

Nowadays online bank services are used by customers every day. The huge numbers of transactions and accounts are a very rich target for fraudsters, which steal billions of euros every year. For this reason financial institutions must adopt fraud detection systems as countermeasures to react to frauds in time. Our goal is to develop a supervised fraud detection tool that helps the transaction analyst to investigate over suspect transactions. The problem can be defined as a classification task, but particular challenges exist in the domain of Internet banking fraud detection: it is dificult to obtain real large datasets, fraudulent transactions behaviours change over time, frauds are rare, misclassification costs are different. In this work we explore supervised algorithms and techniques in order to test and compare these methods on a real bank transactions dataset that we have obtained from an important Italian bank group. We focus on the optimization of a cost function which gives higher weights to fraudulent transactions. The result is a dynamic system that adapts itself. During training, it aggregates attributes per user for each transaction and builds a Random Forest classifier on the basis of analyst's feedback. We implement FraudHunter, a fully working prototype and a web application that can be effectively used in real contexts. We obtain overall good validation results and our system has very good metrics when compared with other works in literature. In particular it can detect up to 87% of frauds with a very low false positive rate of 0.3%.