A Novel Approach for Online Signature Verification Using Fisher Based Probabilistic Neural Network
Abstract
The rapid advancements in communication,networking and mobility have entailed an urgency to furtherdevelop basic biometric capabilities to face security challenges.Online signature authentication is increasingly gaining interestthanks to the advent of high quality signature devices. In thispaper, we propose a new approach for automaticauthentication using dynamic signature. The key featuresconsist in using a powerful combination of linear discriminantanalysis (LDA) and probabibilistic neural network (PNN)model together with an appropriate decision making process.LDA is used to reduce the dimensionality of the feature spacewhile maintining discrimination between users. Based on itsresults, a PNN model is constructed and used for matchingpurposes. Then a decision making process relying on anappropriate decision rule is performed to accept or reject aclaimed identity. Data sets from SVC 2004 have been used toassess the performance of the proposed system. The resultsshow that the proposed method competes with and evenoutperforms existing methods.
Keywords-Online Signature Verification, ProbabilisticNeural Network, Linear Discriminant Analysis
I. INTRODUCTION
Biometric technology is becoming one of the mostimportant supports providing security [1]. It is mainlyrequired for authentication purposes. Biometrics is usuallyused in conjunction with other security alternatives likepasswords and tokens to further improve security providedby the authentication system. The motivations thatencouraged the use of biometrics are their reliability andalso the difficulty to steal, to copy or relatively to forgebiometric information. Biometric authentication consists inestablishing identity using human traits like physiologicalcharacteristics (fingerprint, hand geometry, irispatterns…etc.) and behavioural ones (signature,keystroke,voice..etc). Selecting a particular biometricdepends upon the targeted application, user preference andattitude, practicality issues, accuracy and also technologicalissues and level of security required. With the advent of highquality signature capture devices, signature is attractingmore attention as a biometric to develop practicalapplications. Targeted applications are numerous andinclude banking, e-commerce, access control and egovernmentamong others. The challenge is to develop anauthentication system that is trustworthy and accurateenough. The difficulties inherent to signature basedauthentication are related to the great variability ofsignatures. Furthermore, forgers can reproduce signatureswith high resemblance to the user’s signatures. Forgeriesrange from simple to skilled. Generally, signature basedauthentication systems address two issues: signatureverification and signature recognition. Signature verificationrefers to the process of accepting or rejecting a claimedidentity given an input signature. Depending on the waysignatures are represented and more precisely on theavailability of time related information, methods forautomatic signature based authentication (ASA) fall into twobroad categories namely off-line methods and on-linemethods [2]. In the first case, a signature is represented byan image obtained once the writing process is over bydigitizing the signature written on a paper. In another way,an off-line process deals with a signature as a static twodimensional image that's why it is also known as staticprocess in the literature [3]. On-line methods operate ondynamic features that are captured during the writingprocess using a specialized device. In this case a signature isviewed as an ordered list of points defined each by a list offeatures like x and y point coordinates, pen pressure,azimuth angle of the pen with the digitizing tablet and thealtitude of the pen with the device. Other local and globalfeatures can be derived from these basic features such asvelocity, acceleration, center of gravity…etc.In quest of effective methods for online signatureverification, we describe in this paper a method that relieson the following ideas. First signatures are represented bynormalized dynamic features related to pressure, azimuth,altitude and distance to center of gravity which is a derivedfeature. Second, a linear discriminant analysis (LDA) isused to reduce the dimensionality of the feature space whilemaintaining discrimination between classes. LDA has beenlargely investigated with Principal Component Analysis(PCA) in the context of appearance-based object recognitionand especially face recognition [4]. LDA focuses on thediscrimination between classes whereas PCA describes datawithout paying any attention to the underlying classstructure. PCA has been investigated for off-line signaturerecognition and verification in [5]. The third and core ideaunderlying our work consists in the use of a ProbabilisticNeural Network (PNN) for effective verification


Download full report
http://coeia.edu.sa/images/stories/paper...etwork.pdf