Mobile network Fraud Detection Using Artificial Neural Networks
Abstract
The past decade has witnessed the rapid deployment and evolution of mobile cellular networks,
which now support billions of users and a vast diverse array of mobile devices from smartphones,
tablets, to e-readers and smart meters
Due to this high number of mobile devices and low mobile phone service connection rates Mobile
phone communication is now faced with two major threats which are Voice-related security
threats, ranging from conventional voice scams similar to those on landlines, e.g., stealing
customers privacy information or defrauding users of money through various social engineering
techniques, the new forms of voice fraud that utilize the data functionality of smartphones for
voice-related trickeries. The other threat is the SMS-related security threats which range from
sending threatening messages to other mobile phone users to extort money from them to sending
‘false win’ messages to other subscribers and demand funds in return.
Detecting and rooting out voice-related and SMS-related fraud activities, is not an easy task, due to
the large user population, the vast phone number space and limited data which is recorded when a
call is made or an SMS is send.
The objection of this study was to determine the performance of Artificial Neural Networks in
classifying and detecting the fraud rent activities. We developed a system that uses Artificial Neural
Network to classify phone numbers and detect the once being involved in the fraud rent activities
using the call attributes captured by the mobile service provider. The system was tested using the
data captured in a span of three months and the results compared with actual fraud rent cases
reported to the service provider. In the model different time variant datasets were used to train the
network and perform the classification. Also training the Network using different size variant
dataset was performed. This was to examine the correct data size and age that is optimum and
accurate in the classification.
We found out that Artificial Neural Network was an optimum tool when it comes to classifying
these fraud rent activities due to its ability to dynamically learn fraud rent patterns that change day
by day. We also found out that Training data size and age were major factors that affected the
accuracy in classification.
Citation
Master of Science in Computer Science,2014Publisher
University Of Nairobi