Considerable interest has recently been generated in finding a reliable method of automated personal identification, largely due to increasing fraud involving computer systems such as those used for banking. Research indicates that automated signature verification systems based on capturing the actual dynamics of the signing process promise to be both reliable and economical to implement.
A wide variety of systems have been reported in the literature, differing both in methods of acquiring the dynamical information as well as in subsequent processing and recognition. The results of these research efforts are often difficult to compare due to differences in factors such as sample size, motivation of subjects etc. A recent comparison of recognition methods applied to the same data set has not given completely unequivocal support for any single method.
The sequence matching paradigm is a theoretically attractive framework for recognition of signatures, balancing considerable discriminatory power with reasonable computational demands. A variety of these techniques are applied to this problem. A number of novel approaches are tested with encouraging results.
A data set comprising 520 signatures by 13 different subjects was recorded using a simple and robust pad designed for this purpose, allowing detection of forces in three dimensions. A variety of local distance measures were tested, verification error for any single channel being lowest for pressure information. Performance was found to be significantly better for slope-constrained Dynamic Time Warping (DTW) than for DTW without slope constraints. A significant improvement also resulted from amplitude scaling of the signals prior to time warping. The best results obtained were for a local distance measure on all three of the recorded channels. The equal error rate was 1.3% (on the basis of one signature per verification attempt), while a 0% false acceptance error was attainable at the cost of 11.3% false rejection rate. It was found that, in agreement with other research, most of the errors resulted from extremely poor consistency of a few signers. The majority of subjects were distinguished with very low error rates, with the signatures for 8 out of the 13 subjects being recognised with 0% personal equal error rate.
A dynamic programming correlation method, as well as a string matching method are also investigated. A new amplitude warping algorithm is proposed.