Toll fraud detection in voip networks using artificial neural networks

Abstract

Toll fraud occurs whenever a perpetrator uses deception or dishonest means with the intention to receive telephony services free of charge or at a reduced rate. The introduction and increasing popularity of the internet has brought new possibilities like Voice over Internet Protocol (VoIP). Majority of the voice communication systems in most organizations nowadays have Internet Protocol (IP) connectivity to enable them leverage on the benefits of VoIP. A growing number of these organizations are however being targeted by toll fraud leading to losses estimated to run in billions of United States (US) dollars annually. Toll fraud therefore poses one of the largest threats to enterprise voice systems today. This study presents a VoIP fraud detection model that identifies Artificial Neural Networks as an enabling tool to classify VoIP calls as either fraudulent or legitimate based on the attributes of the call. A multi-layer feed-forward neural network with back propagation learning algorithm was used to build the model. Actual Call Detail Records (CDR) collected from an operational VoIP system in Kenya was used. A total of 15,000 call records were sampled for the study. The data set contained time series of call records from both fraudulent and legitimate callers. Feature selection was performed on the data in order to eliminate redundant variables and select the attributes that would best describe fraudulent behavior. The sampled data was then labeled as being fraudulent or genuine based on the attributes of the call. The data set was partitioned as follows: 70% of the data set was set aside for training, 15% for validation and the remaining 15% was the testing set. The destination phone numbers fell into the following categories: on-net (Internal) and off-net (Mobile, National and International).The implementation of the Artificial Neural Network was based on the Matlab Neural Network toolbox. The trained network achieved 100% classification performance on the test data set. A web application for reading and classifying the call records in new CDR files was developed as part of this study. It utilized the trained neural network and provided a visualization of the classified records. The study established that Artificial Neural Network are a successful technology that can be applied in VoIP fraud detection since it was able to detect abrupt changes in established calling patterns which may be as a consequence of fraud. The implementation of the fraud detection tool will be a big step towards detection and mitigation of VoIP fraud. Keywords: VoIP, Call Detail Record (CDR), Internet Protocol (IP), Artificial Neural Network,Matlab.