Performance Evaluation of Unsupervised Learning Techniques for Enterprise Toll Fraud Detection

In recent years, the impact of telecom frauds is increasing and their behavior is changing through time which makes them to remain the main challenge for telecom service providers in terms of revenue loss and degradation of quality of service. Toll fraud occurs whenever a criminal uses cheating or dishonest means with the intention to use telephony services free of charge, reduced rate or to make money. This study focuses on detection of toll fraud committed through enterprise PBX hacking by analyzing Call Detail Record (CDR) data. Unfortunately, this data is mostly unlabeled, meaning no indications on which calls are fraudulent or nonfraudulent exist. A clustering model was developed and tested with CDR collected from ethio telecom. In order to test the model with big dataset additional synthetic CDR was also used in the research. Two user profile are constructed by summarizing the data on daily basis. WEKA machine learning tool has been used to come up with a model for predicting fraudulent activities. The experimentation result showed that, the model from the K-means algorithm exhibited higher accuracy level, and can be applied in toll fraud detection. Since it was able to detect unusual changes in calling patterns which are highly likely as a consequence of fraud. The implementation of the model will enable telecom operators in general and ethio telecom in particular to detect such fraud at minimum cost of operation. Moreover, it can also be used to support enterprise customers by showing security vulnerability of their Private Branch eXchange (PBX).