Machine learning-based approach for designing and implementing a collaborative fraud detection model through CDR and traffic analysis

Abstract

Fraud in telecommunications networks is a constantly growing phenomenon that causes enormous financial losses for both the individual user and the telecommunications operators. We can denote many researchers who have proposed various approaches to provide a solution to this problem, but still need to be improve to ensure the efficiency. Detecting fraud is difficult and, it's no surprise that many frauds schemes have serious limitations. Different types of fraud may require different systems, each with different procedures, parameter adjustments, database interfaces, and case management tools and capabilities. This article uses the K-Means algorithm to handle fraud detection based on Call Detail Record (CDR) and traffic analysis in a telecommunication industry. Our algorithm consists to compare traffic and CDR generated in the network and check if there is abnormal behavior and if yes, our model is used to confirm if users suspecting of fraud are really fraudster or not. To build our model we used real word CDR data collected in November 2021. Our model associates the Differential Privacy model to encrypt users' personal information, and the k-means algorithm to group users into different clusters. Those clusters represent non fraud users having similar characteristics based on criteria used to build the model. Users having abnormal behavior that can be assimilated to fraudsters are those who are far from the different clusters center. Thanks to a representation in a plan, we better visualize user’s behavior. We validated our model by evaluating our segmentation method. The interpretation of the results shows sufficiently that our approach allows to obtain better results. Our approach can be used by all telecommunications operator to reduce the impact of fraud on internet services.