Signature Based Authentication

Authors: Sam Krasnik, Ron Hose.

Currently, the viable authentication mechanisms in use are password or PIN based. A user who wishes to be authenticated enters a short string or a number which is supposedly easy to remember, but is easy to forget and the authenticating machine checks the hash of the password against a stored hash in a database. If a user were to forget his or her password or it was stolen by someone wishing to steal the user's identities, the identity of the user becomes endangered and easy to compromise.

A common solution to the problem is to implement a biometric authentication mechanism such as retina scans or fingerprint based authentication. The advantage of this mechanism is that a retina or a fingerprint cannot be forgotten and it is unique to each user. While such a system is extremely difficult to compromise, The cost of implementation restricts its deployment. In addition, if such biometric data was in fact compromised, the result would be disastrous for the user-- he or she cannot change fingerprints or retinas.

To solve both the problems of password based authentication (relatively easy to compromise, easy to forget) and biometric authentication (expensive, intrusive, and disastrous if compromised), we attempt to solve the problem of user authentication via Handwritten Signature Verification (HSV). The hardware necessary for HSV is inexpensive and already widely deployed in (e.g. credit card machines in stores, tablet PCs)

The problem consists of verifying a user's identity based on her signature by requiring the user to enter a signature and comparing the signature to a set of previously entered signatures. The system should have a very low false acceptance rate and a pessimistic false rejection rate. Our basic approach to solving the HSV problem uses on-line back-propagation. Both local and global features of the input signatures are used to train an Artificial Neaural Netowrk

Some related work in the field uses only off-line data, while others also use on-line data. Some use only global or only local features, while others use a combination. The work presented is one of the most general approaches to the problem.

Our results indicate that the back-propagation algorithm was able to converge on more than 90% of user data and achieved a pessimistic False Rejection Rate of around 30% and a low False Acceptance Rate of 2%. We conclude that the solution presented is successful in solving the Handwritten Signature Authentication Problem.