Toll Fraud Detection of Voip Services via an Ensemble of Novelty Detection Algorithms

Abstract

Communications fraud has been dramatically increasing with the development of communication technologies and the increasing use of global communications, resulting in substantial losses to telecommunication industry. Due to the widespread deployment of voice over internet protocol (VoIP), the fraud of VoIP has been one of major concerns of the communications industry. In this paper, we develop toll fraud detection systems based on an ensemble of novelty detection algorithms using call detail records (CDRs). Initially, based on actual CDRs collected from a Korean VoIP service provider for a month, candidate explanatory variables are created using historical fraud patterns. Then, a total of five novelty detection algorithms are trained for each week to identify toll frauds during the following week. Subsequently, fraud detection performance improvements are attempted by selecting significant explanatory variables using genetic algorithm (GA) and constructing an ensemble of novelty detection models. Experimental results show that the proposed framework is practically effective in that most of the toll frauds can be detected with high recall and precision rates. It is also found that the variable selection using GA enables us to build not only more accurate but also more efficient fraud detection models. Finally, an ensemble of novelty detection models further boosts the fraud detection ability especially when the fraud rate is relatively low.