Supplier Impersonation Fraud Detection using Bayesian Inference

In this paper, we introduce ProbaSIF, a supplier impersonation fraud detection system that relies on a Bayesian model to perform the classification of a new transaction as legitimate or fraudulent. ProbaSIF is divided in two parts: an intra-company analysis that aims to recreate the vision of a specific client about the legitimacy of the account used in a transaction with one of its supplier, and an inter-company analysis that uses all the accounts used to pay a supplier to model the supplier's payment behavior and take into account transactions issued by other clients. We use a dataset composed of more than 2 million transactions issued by real companies, provided by the SiS-id platform, to fit our Bayesian model, and evaluate the classification results of ProbaSIF using an other set of 108,000 transactions labeled by SiS-id expert system. Our study of a representative client shows that both of the approaches described in ProbaSIF show good precision (0.927 and 0.836) for the 255 transactions tested. Results also shows that ProbaSIF gives results consistent with the expert system provided by SiS-id. Finally, after evaluating ProbaSIF approaches on all the clients available in our dataset, we demonstrated that our classification system was accurate for a wide set of different clients.