Do synthetic generated signatures reflect the subject motor programs? A pilot study

Handwritten signature is a biometric trait used for verifying a person’s identity. Automatic signature verification systems typically require a lot of specimens in order to model the signing habit of a subject but, in a real scenario, few signature samples are available. To overcome this problem, methods for creating human-like duplicated signatures using one real signature per subject and based on sigma lognormal decomposition have been proposed in literature. In this paper, we evaluate if duplicated signatures show the same amount of variability observed in real signatures by detecting and analysing signature stability regions. In particular, we investigate if real and duplicated signatures could be the instances of a similar motor program. Experimental results on a standard dataset show that in some cases duplication methods introduce a variability that is greater than the writer's variability to such an extent to generate motor programs that do not belong to the writer's repertoire. Results suggest that a connection exists between trajectory plan and motor plan parameters, which cannot be modified independently one from the other in order to generate synthetic signatures that reflect the writer’s motor program repertoire.