Fraud Detection in Online Auctions

Abstract

Online auctions are auctions held over the internet. Online auctions avoid some of the problems of traditional auctions, such as geographical and time restrictions, and limited audiences. However, this has also made them a target for fraud. The majority of pre- vious approaches for reducing auction fraud make use of machine learning techniques to identify suspicious users and auctions. However, many of these approaches have generally encountered the same challenges during implementation and evaluation. These challenges include lack of good quality datasets, imbalance in the number of normal and fraudulent examples, heterogeneity of both normal and fraudulent behaviours, and behaviours which evolve over time. In this thesis, we explore methods of dealing with some of these chal- lenges. To address the lack of good datasets, we implement an agent-based simulation for online auctions as a means of generating synthetic auction data. The simulation models the behaviour of normal bidders and sellers, based on real online auction data. The synthetic data was evaluated using three methods, and results show that the synthetic data is similar to real data. We then applied the simulation to evaluate an existing fraud detection method to show that the simulation can be used as a source of data for evaluating fraud detection algorithms. To address the di culty in creating detection models for di erent fraud behaviours and strategies, we demonstrate using supervised learning methods with our simulation to easily create models for detecting arbitrary types of fraud. Models created using this approach were shown to have higher accuracy compared to an existing fraud detection method even after tuning. The caveat is the fraud type of interest must be explicitly de ned. To deal with the limitations of the previous approach and to avoid the need for model retraining when fraud behaviours change, we propose an unsupervised method based on anomaly detection. Since the method uses an anomaly detection approach, the model can adapt to changes in user and fraudulent behaviour: the method will identify users who behave di erently to the majority of other users. The method makes use of additional net- work information to identify groups of users that appear suspicious. Extensive evaluation using synthetic data shows that it has higher accuracy than other related approaches. When applied to a real dataset, our method nds a reasonable number of potentially fraudulent users who exhibit unusual characteristics when compared to normal users.