Digital administrative documents forgery detection using One-class SVM

Abstract

In everyday life, we rely on digital documents in almost every transaction. And with the widespread use of these documents in all jurisdictions, a new crime has emerged. This is the forgery of documents using a scanner, and advanced tools for images editing that allow a simple user to alter a digital document and change its content. To prevent such a crime and to fight against fraudsters who are constantly developing new methods, many researchers have tried to develop automatic methods for fraud detection using image processing techniques and machine learning. Most of these works use the binary classification to classify documents as authentic or forged and need to use samples of the two classes together for training.

In this work we propose a method to detect digital administra- tive documents forgery in cases where only authentic documents are available. The proposed method extracts background, text and stamp characteristics and build a decision model using One- class SVM. The obtained model was tested on a set of generated administrative documents using several kernel functions and parameters. Results show a high recognition rate, thus proof the effectiveness of our model.