Markov Modelling of Simple Directional Features for Effective and Efficient Handwriting Verification

Abstract

Signature verification has long been a traditional means of authenticating an individual, with this process now being automated via a number of research activities. The problem with automated signature verification systems is that they can be susceptible to forgery as it is often possible to obtain a copy of an individual’s signature. The work described here alleviates this problem to some degree in that a signature is not used but rather a user-determined “password”.

This approach is centred on the fact that people don’t write according to a standard penmanship and deviation from the norm is dependent on the individual. Additionally, individuals tend to deviate in a similar way from one instance to the next. This similarity of deviation can be exploited to allow identification of the author by analysing a small sample of handwriting.

The handwriting is captured and digitized in real-time using a graphics tablet, so no physical evidence of the handwriting sample is ever recorded (eliminating the possibility of the password being stolen). The samples are modelled in the system using a Markov model with five states. The state transitions of the model are determined by first segmenting the handwriting sample into a series of “strokes” (pen path between consecutive minima in the pen-tip velocity). The next step involves obtaining the “net direction” for each stroke by positioning the beginning of the stroke at the origin. The current state is then assigned a value corresponding to the quadrant in which the stroke end-point lies (or a fifth state representing a “pen-up” occurrence).

This system takes advantage of multiple security schemes in that users would benefit from the protection of password defenses as well as aspects of signature verification. A potential forger does not automatically gain access to a resource simply by finding out (or guessing) the user’s password, they also need to be able to forge the writing style of that user. The opposite is also true - if a forger is familiar with the user’s writing, that forger must also know the user’s password if they are to break into the system.

The results obtained from a database of almost 1000 handwriting samples from 47 writers include an error rate of 0.64% when the potential forger does not know the genuine user’s “password”.