Fraud detection from paper texture using Siamese networks

In this paper, we present a model for the fraud detection of documents, using the texture of the paper on which they are printed. Different from prior studies, we present a data generation process through which we generate a dataset of papers and propose a deep learning model based on Siamese networks that is trained with samples from the dataset to reliably detect fraud from the original. Toward this end, we introduced a new regularization parameter for the training that would reduce the likelihood of the network making a Type-II error (i.e., classifying a fraud document as original), while being more tolerant of Type-I error (i.e., classifying an original document as fraud). Our analysis has shown that, combined with a Meta Learner, the proposed model can provide better fraud detection performance than that obtained with the Local Binary Pattern method, Prototypical Networks, and Matching Networks as the baseline.