Streaming Fraud Detection on Session Based Data

Abstract

Financial fraud is, within the banking world, a major source of expenses. Improving on timely detecting fraud is a constantly ongoing cat and mouse game between the financial institutions and criminal organisations. To be on the edge of fraud detection, new approaches have to constantly be developed. This research will tap into the large amount of data produced by systems around the actual financial transactions and attempts to use this data to timely, as quick as possible, react on events. This means that the system developed is able to receive an incoming stream of data, find relations in this data, classify instances as either fraudulent or not and return this information. The proposed algorithm performs as second when compared to three peer researched techniques on a widely scientifically used, topic related, publicly annotated intrusion detection dataset. The comparable cost of our algorithm equals 0.141 where the cost for our peers is 0.058, 0.254 and 0.376. 20000 real world cases lead to 0.16% found anomalies (32) of which expert review pointed out six suspicious cases where three were to be investigated. These results show the viability of our research on the to hand problem of timely detecting financial fraud in an ongoing stream of events.