Frameworks for Fraud Detection in

Fraud is costing the mobile communications industry millions of pounds a year. A rapid solution is needed to reduce fraudulent activity in analogue networks and preventative measures are required to protect GSM and later UMTS. A joint European project`Advanced Security for Personal Communications Tech-nologies' (ASPeCT), part of the ACTS programme 1 , has been formed to research security issues in mobile communications networks. Part of this project is to investigate how Artiicial Intelligence can be used by a network operator to detect fraudulent activity in a real-time environment. We discuss ways to characterize a user's behaviour by computing user prooles over sequences of Toll Tickets. We show that with a neural network fraud detection system we can monitor user behaviour patterns through both diierential and absolute usage. A diierential analysis enables us to detect changes in behaviour associated with a mobile telephone which could indicate fraudulent usage after a theft or through cloning, more common under the TACS system. The GSM system is still relatively secure. The most common fraud is that of subscription fraud. We propose that our systems would detect this through absolute usage. We describe a number of potential systems architectures using Neural Networks each varying in complexity that would enable us to implement these ideas. We suggest how clustering techniques and Non Linear Dimension Reduction could be used to form natural classes of users. Separate systems could be run in parallel to process Toll Tickets relating to each user class. We discuss how an oo-line adap-tive critic could monitor current user trends and update individual component systems working on separate user classes. Finally we discuss the beneets and drawbacks of a number of the most promising systems and suggest a way forward.