On-line systems for human signature verification

The design and implementation of an on-line dynamic signature verification system requires the solution to the following problems: data acquisition, feature extraction, feature selection, decision making and performance evaluation. Although much effort has been expended, past efforts have not resulted in adequate solutions to such pressing tasks as: construction of a reliable data base, selection of optimum feature sets with or without forgery data available, finding classifier independent feature selection procedures, obtaining reliable asymptotic global and individual performance, obtaining a suitable statistical model for signatures, minimizing the effects both of the inconsistency of genuine signatures and of the variety of forgeries, and adapting to practical limitations such as on-line response and limited memory size.; Our data acquisition system captures human signatures dynamically using a graphics tablet. The problem of data acquisition was addressed by compiling a reliable data base of more than ten thousand signatures.; For feature extraction we started with a 42-parameter feature set and advanced to a 49-normalized feature set. The 49-feature set is particularly effective in tolerating inconsistencies in genuine signatures without losing the discrimination power against forgeries. We proposed a statistical model for parameter feature vectors. Such a model allows estimation of asymptotic performance of a signature verification system even with a relatively small data base.; Our feature selection algorithms have the property of being classifier independent and provide optimum subsets of features with or without forgery data.; For decision making we proposed the majority classifier. The majority classifier, in addition to providing the performance suitable for point-of-sales application, performs better than many well-known classifiers in the sense of higher overall discrimination power between genuine signatures and forgeries.; In order to tolerate inconsistency in genuine signatures and at same time to be effective against forgeries, we introduced a new set of 49-normalized features and two modified versions of the majority classifier. Performance of as low as 2.5% equal error rate and, more importantly, an asymptotic performance of 7% Type II error at zero Type I error is achieved with the modified majority classifier using only 15 parameter features.