Optimization of a Signature Verification System Using Neural Networks [chapter]

Luan Ling Lee, Toby Berger
1993 ICANN '93  
Signature verification is a two-pattern classmcation problem or a hypothesis testing problem. A signature vermcation system for poinkf-sales applications requires an extremely low false rejection rate of genuine signature due to the fact that the satisfaction of customers is a top priority [1]. On the other hand, a system for purpose of granting entry to a secure facility must operate at or near zero false acceptance of forgery in order to be assured of barring any intruder. This paper describe
more » ... This paper describe two neural network (NN) based approaches for optimization of a signature verifi.cation system. which uses a majority classifier: 1) mjnjmizing the total frequency of misclassmcation using the pocket perceptron learning algorithm [3], and 2) simultaneously minimizing the false rejection error and the total frequency of misclassifi.cation using a modiftcation of the pocket algorithm. A majority rule based signature verifi.cation system employs a set of 42 personalized parameter features consisting of both dynamic and static features extracted from sampled signatures by using a graphic tablet [2] . A majority classifi.er assigns weight one or zero to each feature according to a defined condition. The architecture of the networks is inspired by that of the single-cell model proposed in [3] . The NN classifter is a feed-forward network with one-hidden layer. The output unit computes a weighted sum of its binary inputs obtained from normalized feature values after hard limiters. The pocket algorithm and its modiftcation, because of their positive feedback, are well-behaved even for nonseparable problems. The basic principle of the both algorithm is to keep currently the best set of weights in a "pocket" while training the network until a better set is found in the sense of a higher percentage of correct classifi.cation. Two sets of experiments were carried out. Simulated and real signature data were used in the first set and second set of experiments, respectively. Neural network methods are particularly attractive when statistical properties of the feature set are unknown. In all experiments, the neural networks perform better than the majority classifter. The NN classifi.er trained by the pocket algorithm is considerably sensitive to the size of genuine and forgery training set. This fact is confirmed by the receiver operating characteristic (ROC) mathematical model in [4] . However, the classifter has the smallest number of misclassiftcations which is the real virtue of the pocket algorithm. The modifted pocket algorithm is less sensitive to the size of training sets and is capable of providing better asymptotic performance than the NN trained by the pocket algorithm. We believe that the proposed learning algorithms and the NN classifi.ers can extent their applications to other practical problems of pattern classifi.cation. References [1] L. L. Lee and T. Berger, "Adaptive method and system for real time verifi.cation of dynamic human signature," United States Letters Patent,
doi:10.1007/978-1-4471-2063-6_275 fatcat:azkawf5wmbdmniqglidqjviqqi