Towards the design of an offline signature verifier based on a small number of genuine samples for training

Signature verification has remained one of the most widely accepted modalities to authenticate an individual primarily due to the ease with which signatures can be acquired. Being a behavioral biometric modality, the intra-personal variability in signatures is rather high and extremely unpredictable. This leads to relatively higher error rates as compared to those realized by other biometric traits like iris or fingerprints. To address these issues, this study investigates run-length distribution features for designing an effective offline signature verification system. The objective of this research is to enhance the capabilities of automatic signature verification systems allowing them to work in a realistic fashion by training them using positive specimens (genuine signatures of each person) only without access to any forged samples. Classification is carried out using One-Class Support Vector Machine (OC-SVM) while the evaluations are performed using GPDS960 database, one of the largest offline signature corpus developed till date. Experimental results show the potential of the proposed system for detection of skilled forgeries, especially for the challenging case of a single reference signature in the training set.