Fingerprinting devices to identify fraud

Client information / Background

Spatialedge designed and implemented a fraud reduction solution for a client in the finance sector. This solution utilised advanced data science methods to match point of sale devices based on customer behaviour.

The Challenge

The client faced issues with fraud related to loans on credit card machines. Vendors and sales agents were exploiting the system; vendors by obtaining loans on multiple devices fraudulently, and sales agents by earning commissions on the installation of these devices. This was compounded by the challenge of linking devices back to individual vendors due to poor data capture.

The Solution

The solution was the development of a "fingerprinting" model designed to identify similarities in behaviour across devices to infer if they were operated by the same vendor. This model allowed for the clustering of devices based on usage patterns and behaviours, thereby identifying potential fraudulent activities. 

Data and Approach

The model utilises data generated from the operation of card machines, which, although not explicitly detailed, includes transaction volumes, frequency, patterns, and types of transactions conducted. This data is critical in identifying the unique "behavioural fingerprint" of each vendor.

The system doesn't directly link devices to specific individuals or vendors due to inadequacies in data capture. Instead, it looks for patterns and similarities in the behaviour of different devices to infer if they are operated by the same individual or entity.

Clustering and Pattern Recognition

The fingerprinting model employs clustering algorithms, a type of unsupervised learning, to group devices showing similar operational behaviours. These algorithms might consider various features, such as transaction frequency, average transaction value, geographic location consistency, and temporal patterns in device usage.

By analysing these clusters, the model can identify devices that share a high degree of similarity in their operational patterns, suggesting they are operated by the same vendor.

Decision Process and Outcomes

When a loan application is made for a device, the model assesses the device's operational data against its learned patterns to determine the likelihood that the application is fraudulent. If the application's device exhibits behaviours highly similar to those of devices previously identified as fraudulent, the loan application will be automatically declined.

The Result

Early investigations showed promising results, with a preliminary accuracy of 74% in identifying fraudulent activities. The model's deployment is expected to significantly reduce fraud by automatically declining high-risk loan applications. 

This proactive approach to fraud prevention aims to safeguard the company's financial interests and reduce the operational costs associated with investigating and addressing fraud. Additionally, it provides a scalable and efficient method for identifying and mitigating fraud risks.