Fraud detection in online community trading sites

Abstract

Online community trading sites have been the target of various fraudulent schemes, resulting in considerable monetary losses. Maintaining reputation through user feedback is a system often employed by contemporary trading sites in an attempt to measure the trustworthiness of a seller. However, these mechanisms can be easily foiled through the creation of multiple fake accounts, aimed at artificially inflating the reputation of fraudsters. The aim of this project is to explore and apply promising methods of detecting fraud in this context. To this end, two fraud detection approaches are adopted, which focus on compensating for the drawbacks in current reputation systems. The first approach, stemming from the field of anomaly detection, is based on the intuition that by modelling user behaviour in a hypercube, fraudulent individuals can be distinguished as outliers from clusters of legitimate users. The second approach utilises a graph-based technique to leverage known fraudulent interactions between users specifically employing the use of fake accounts. A challenging aspect of this study involves the acquisition of real user information, which is crucial in evaluating these approaches. To accomplish this task, we developed a user auction data retriever component specifically focused on retrieving live user information from an online trading site. Our results demonstrate that the anomaly-based approach is a viable technique which achieves a high detection rate, albeit with a relatively high rate of false-alarms. However, this technique exhibits weak support for early fraud detection, as it requires a degree of negative feedback ratings for effective classification. On the other hand, the graph-based approach also achieves a high detection rate and is capable of early fraud detection, although with the caveat that it needs to consider a much larger portion of the community's behaviour as compared to the anomaly-based approach.