Handwritten signature verification using complementary statistical models

Abstract

There is considerable interest in computerised personal identification and in particular in biometrics, a branch of identification that deals with verifying physical or behavioural characteristics of human beings. This thesis is concerned with the development of the particular biometric of handwritten signature verification which is superior in many ways to other biometric authentication techniques that may be reliable but are much more intrusive. Specifically this project involves the use of two complementary artificially intelligent systems in the form of neural networks and hidden Markov models. Five sample signatures are used to build a reference in each of the independent models and experimentation and testing is done using an extensive database of almost 4000 genuine signatures and forgeries. The confidence levels from each model are then combined and tested on unseen signatures resulting in an equal error rate of 1.1%. Further experimentation is performed and includes analysis of different verification scenarios, error contribution and the importance of visual feedback when signing. Finally, experiments are conducted exploring the possibility of "signing" handwritten passwords, with the developed system resulting in an equal error rate of 0.7% in the worst case.