Next Generation Machine Learning Based Real Time Fraud Detection

Abstract

In telecommunications there are several schemes to defraud the telecommunications companies causing great financial losses. We can considerer three major categories in telecom fraud based on who the fraudsters are targeting. These categories are: Traffic Pumping Schemes, Defraud Telecom Service Providers, Conducted Over the Telephone. Traffic Pumping Schemes use "access stimulation" techniques to boost traffic to a high cost destination, which then shares the revenue with the fraudster. Defraud Telecom Service Providers are the most complicated, and exploit telecom service providers using SIP trunking, regulatory loopholes, and more. Conducted Over the Telephone, also known as "Phone Fraud", this category covers all types of general fraud that are perpetrated over the telephone. Telecommunications fraud negatively impacts everyone, including good paying customers. The losses increase the companies operating costs. While telecom companies take every measure to stop the fraud and reduce their losses, the criminals continue their attacks on companies with perceived weaknesses. The telecom business is facing a serious hazard growing as fast as the industry itself. Communications Fraud Control Association (CFCA) stated that telecom fraud represented nearly $30 billion globally in 2017 cite{telecomengine}. Another problem is to stay on top of the game with effective anti-fraud technologies. The need to ensure a secure and trustable Internet of Things (IoT) network brings the challenge to continuously monitor massive volumes of machine data in streaming. Therefore a different approach is required in the scope of Fraud Detection, where detection engines need to detect risk situations in real time and be able to adapt themselves to evolving behavior patterns. Machine learning based online anomaly detection can support this new approach. For applications involving several data streams, the challenge of detecting anomalies has become harder over time, as data can dynamically evolve in subtle ways following changes in the underlying infrastructure. The goal of this paper is to research existing online anomaly detection algorithms to select a set of candidates in order to test them in Fraud Detection scenarios.