How to detect problematic signers for automatic signature verification

Automatic Signatures Verification (ASV) is a pattern recognition application where the patterns are obviously the signatures of different signers. To correctly characterize a given signature, one must use several specimens from each signer and therefore, the intrinsic variability of some signers become a fundamental issue. Another fundamental issue of ASV system design is the unpredictable nature of true forgeries. This emphasizes the problem already mentioned but also points out another aspect of ASV systems that is usually not taken into account by system designer: the intrinsic difficulty of signatures to be Imitated by potential forgers. An experiment specifically designed on signature imitations has been conducted. Eight subjects were trained with the help of an electro-acoustical set-up in order to imitate, both visually and dynamically, eight reference signatures. First, these signatures have been analyzed to estimate the difficulty of each one to be imitated. This has been done with the help of a parametric coefficient based on data from experimental psychology and handwriting generation models. These signatures have also been analyzed to estimate their intraclass dissimilarity by an automatic comparison algorithm. The results of the experiment show to some extent, that problematic signers are those with signatures which are both instable (high intrinsic variability) and easy to imitate