On-Line Signature Verification: Comparison and Fusion of Feature Based and Function Based Classifiers

In this chapter the authentication performances achievable by on-line signature recognition systems employing feature based and function based classifiers are analyzed. Specifically, three different stand-alone matchers, namely Hidden Markov Models, Parzen Windows classifier, and Symbolic classifier, are compared. An ensemble of the considered matchers, which significantly improves the system verification performances, is also proposed. Moreover, the usefulness of a feature selection process, and of a random feature subspaces approach for on-line signature verification, are also evaluated. The present work also investigates the authentication performance achievable when protected on-line signature based recognition systems are considered, by exploiting the BioConvolving template protection scheme. An extensive set of experimental results, obtained by using the public MCYT on-line signature subcorpus, which comprises 100 users, for each of which 25 genuine signatures and 25 skilled forgeries have been captured, are also presented.