A Machine Learning Approach to Off-Line Signature Verification Using Bayesian Inference

A machine learning approach to off-line signature verification is presented. The prior distributions are determined from genuine and forged signatures of several individuals. The task of signature verification is a problem of determining genuine-class membership of a questioned (test) signature. We take a 3-step, writer independent approach: 1) Determine the prior parameter distributions for means of both "genuine vs. genuine" and "forgery vs. known" classes using a distance metric. 2) Enroll n genuine and m forgery signatures for a particular writer and calculate both the posterior class probabilities for both classes. 3) When evaluating a questioned signature, determine the probabilities for each class and choose the class with bigger probability. By using this approach, performance over other approaches to the same problem is dramatically improved, especially when the number of available signatures for enrollment is small. On the NISDCC dataset, when enrolling 4 genuine signatures, the new method yielded a 12.1% average error rate, a significant improvement over a previously described Bayesian method.