Application of mahalanobis distance in a peer profiling model for fraud detection

Abstract

With the increasing volume of mobile and internet banking money transactions in today's era, fraud detection has become a major concern in ensuring a secure and trustworthy financial system. The rise in both digital money transactions has led to convenience and posed risks in opening gaps for fraudulent activities. Due to the changes in the digital era, fraudsters are looking for new techniques to exploit vulnerabilities in digital banking systems which leads to financial losses. The failure to stop fraud can lower the consumer benefit and financial inclusion gains in the businesses. There is a need to implement applications that accurately detect fraud as digital transactions in both mobile and internet banking has become the most used for fraud and other criminal activity. The traditional methods involve rule-based systems to sort the data depending on the scenario that a human will term as a fraud. The human experts in the fraud domain will explore different transactions and customer accounts to note which ways swindlers use to perform suspicious activities. This will be then replicated as a rule in case that is done again the transaction will be termed as suspicious. There are a lot of limitations in this as there are infinite rules to be created to detect fraudulent activities. The systems that use have limitations that hinder efficient detection thus limited adaptability. Current systems and approaches use different machine learning techniques that are based on pattern recognition which work well but the models must be retrained after a short time as they have a limited shelf life. The other problem is that swindlers and fraudsters also use machine learning in their arsenal of tricks to perform fraud; since they are smart people thus eloping fraud tool defenses as they understand how the pattern recognition models work thus fraud happening. This is because fraud evolves, the fraudsters evolve, and technology advances thus a better need for fraud detection ways. The goal is to develop a tool that incorporates the use of the peer profiling method through clustering. The method involves the grouping of a dataset into different groups based on similar attributes and features of the peers. Afterward, new transactions can be classified from where the peer resides and use mahalanobis distance to calculate if the distance is away from the peer to be classified as an outlier. This proposed research aims to use the peer-based profiling and clustering algorithm to improve the accuracy of fraud detection. This will assist in reducing false positives and increase the accuracy of the frauds detected compared to the traditional way of detecting fraudulent activities.

Keywords: peer profiling, mahalanobis distance, fraud detection