A performance evaluation of a new signature verification algorithm using realistic forgeries

A neural network architecture for carrying out signature verification was developed and tested in an earlier study using a segment-based autoregressive characterization of the signatures. In this work the model and classifier are subjected to a more rigorous test using an extended database of realistic forgeries. While there is some deterioration in performance, it is shown that the proper selection of the modalities of training and the inclusion of time of execution of the signature as an additional feature make the model fairly robust. False Acceptance and False Rejection error rates of 0.78% and 1.6% respectively were obtained in tests conducted using 1920 skilled forgeries.