Limits of and Enhancements to NetProbe for Analyzing Online Auction Networks

Fraud is an ever present threat in online auctions. The anonymity of the Internet provides a hospitable environment to both buyers and sellers of disrepute. To combat this problem many solutions have been proposed which use belief propagation and social network analysis to identify fraudulent actors. In this paper, we mainly analyze NetProbe which uses belief propagation over Markov Random Fields to classify users in online auctions as honest, fraud and accomplices. However, NetProbe fails to address some properties inherent to online auction networks like eBay. We identify cases where NetProbe might misclassify users or how users can exploit vulnerabilities of the NetProbe algorithm. We then propose a modification to the belief propagation algorithm to address the vulnerabilities to a great extent. Our assumptions on the properties of the original eBay graph and the effectiveness of the modified belief propagation algorithm are validated by running it on an synthetic network of 1000 nodes and approximate 2500 edges and an eBay network of 50,074 nodes and 149,935 edges collected using the eBay API. We observe that there is a 3.8% chance of a fraudster being flipped to an honest user, a 4.4% chance that an accomplice is flipped to an honest user and a 0.1% chance of a honest user being misclassified as an accomplice or fraudster in our synthetic test network. Keywordsbelief propagation;auction networks;