Improving Signature Forgery Analysis through Deep Learning Classification Algorithms

Abstract

This project seeks to develop a machine learning algorithm to identify a forgery from a legitimate signature, for use in signature verification in low profile financial crimes. This model will be trained on data collected on specific handwriting characteristics used by professional document analysis experts. Signature forgery in financial institutions was recently brought to light in the Wells Fargo fake accounts scandal, where employees opened 3.5 million unauthorized accounts, for which 190,000 customers were unwittingly charged fees. This product will help protect individuals from exploitation by providing a verification tool for company managers and executives. The hypothesis predicted if characteristics such as length-to-height ratio and relative slant were taken into account, then the accuracy of the model would be greater than 90%. Each data point consisted of 6 signatures, of which 5 were ‘true’ (produced by the same person) and the final was forged. The data was fed to a processing program that used mathematical formulae (standard deviation from the mean) to account for, and negate human error. Finally, using the python Scikit machine learning library, multiple models were trained on the data sample, using k-fold analysis. The most successful model, XGBoost Classifier, had an accuracy rate of 94.55%.