Fraud detection in telecommunication using pattern recognition method / Mohd Izhan Mohd Yusoff

Mohd Yusoff, Mohd Izhan (2014) Fraud detection in telecommunication using pattern recognition method / Mohd Izhan Mohd Yusoff. PhD thesis, University of Malaya.

Preview	PDF Download (6Kb) | Preview
Preview	PDF (Full Text) Download (1098Kb) | Preview

Abstract

Fraud activities have reached to critical point causing millions of ringgit of losses to telecommunication companies, and as a result, forcing them to employ applications or systems (such as Telekom Malaysia Berhad’s Next Generation Fraud Detection System) to detect the said activities. We introduce a new algorithm that could detect fraud activities in telecommunication industry (e.g. intrusion fraud which occurs when legitimate account is comprised by an intruder who makes or sells calls on this account) that uses Gaussian Mixed Model (or GMM), a probabilistic model normally used in fraud detection via speech recognition. Due to the complexity of GMM, we use Expectation Maximization (or EM) algorithm by Dempster et al. (1977) to obtain the maximum likelihood estimates of the GMM parameters. Together with Kernel method (see Silverman, 1986), we improve the process of finding the number of components in GMM. In addition, we have also successfully derived the likelihood ratio test in the determination of the number of components in GMM and the comparison of its results with those of Akaike Information Criteria (AIC) will also be highlighted in this thesis. The said algorithm uses similarity coefficient to classify the real data based on the log-likelihood function and it’s extended to detect incoming fraud calls as suspected by the telecommunication company. The new algorithm is tested on simulated and real data where the results show it is capable of detecting fraud activities. The real data, which included call charging and duration, are collected from Telekom Malaysia Berhad’s exchanges and they are believed to be contaminated by fraud activities. As the original data are clearly not in the format that is generally used for speech recognition, they are reformatted prior to testing and analysis. The new algorithm is in agreement with those suspected by the company.

Actions (For repository staff only : Login required)