fraud detection performance

Fraudulent payroll accounts are a significant issue in the financial and payroll services market. About 75 percent of employees have stolen from their employer once, 20 percent have stolen at least twice, and half of this figure has stolen at least three times or more. The increased use of direct deposit has made it easier to steal the identity of a small payroll customer. A fraudulent account is set up with apparently legitimate credentials, and the account’s data set is changed after the customer’s login authentication returns an all-is-ok status. The direct deposits are then quickly withdrawn. It’s an ongoing game of cat-and-mouse: after a bank improves or alters its fraud detection system, criminals quickly adapt to avoid it. Small-scale business contract fraud is accepted and paid as normal business expenses, and there is extremely little fraud detection technology in place to pick up the kickback schemes. This type of fraud can bring a company to its knees. (B) Automated fraud detection in payroll and financial management services harnesses sophisticated deep learning architectures. Computers do the heavy lifting for identifying patterns, obtaining insights, making decisions, and taking action. Unless large fraudulent datasets are already commercially available, the models are unable to understand or predict fraud when trained only on legitimate transaction data. Constructing artificially oversampled imbalanced data sets leads to flawed models. A self-training active learning ensemble stack of models using transaction authentication data is described. It directly leverages fraud patterns in a mostly-unlabeled data set and requires minimal retraining when a pattern changes. Model performance is measured with area under the ROC curve, and this method outperforms existing techniques.