00000478

7/27/2019 00000478

1/4

Computerized Signature VerificationSystemDinesh F? Mital, C h o o Pee Hin, and Wee Kee Leng

l

a +rn Scanner9 preprocessor

--j Hadamard + rocessing _j Verificationtransformnd- T

ABSTR ACT: The fine structure of the mus-cIc forces that are exerted during the writingof a signature is consistent and well definedfo r most people. Based on this observation,an experimental system that utilizes a per-sons signature waveform dynamics foridentification has been proposed. T he systemis intended to be used for on-line signatureverification. It can successfully verify a per-sons identity and can also detect forgeries.The acceptance rate for random forgeries,i.e., accidental matching of two differentpersons signatures, is very low.

IntroductionWith the development of widely dispersed

networks of computer terminals, automatictellers, and data banks. there has been a cor-responding increase in computer crime anda growing need to protect sensitive infor-mation. An important aspect of the problemis personal identification, that is, the abilityto ensure that only authorized people get ac-cess to computer resources. A method ofpersonal identification that cannot be lost,stolen. or forgotten is required for control ofcomputer access, building access. or auto-mated banking. An effective method of on-line signature verification will have manyimportant applications. Because the signa-ture is the normal and customary method ofidentifying an individual, it has many naturaladvantages over other techniques such as fin-gerprints or voice verification.

There has been considerable research insignature verification [ 1]-[4], and, basically,there are two ways to obtain such a repre-sentation. One way is to scan the signatureoptically after it has been written. This tech-nique is similar to optical character recog-nition. However, the optical scanningmethod is costly and time-consuming, and isnot suitable for real-time applications. Amore attractive and useful approach is toPresented at the 1987 IEEE International Confer-ence on Systems. Man, and Cybernetics, Alex-andria. Virginia, October 20-23, 1987. Dinesh P .Mital is a staff member of the School of Electricaland Electronic Engineering. Nanyang Technolog-ical Institute, Singapore. Choo Pee Hin and WeeKee Leng are working with Housing DevelopmentBoard Corporation and Real Time Systems. Inc..Singapore. respectively.

generate electric signals representative of thesignature during the signing process [5]-[8].

There have been mixed results using thesetechniques. The algorithms have been com-plex, which discourages microprocessor im-plementation and on-line applications. How -ever, a major advantage of this approach isthat signature verification is based on the dy-namics of the signature, which are not visi-ble and, therefore, are very difficult to forgeor copy. In this respect, a static image of thesignature on a card or document is almostuseless to the forger, because the dynamicsof the forged signature are usually com-pletely different from that of the true signa-ture.This paper describes an on-line signatureverification system based on pressure wave-form measurements. The algorithm is basedon extracting features from the signals gen-erated by the instrumented pen. The advan-tages of this method are computational effi-ciency, minimal reference or templatestorage per user (typically, 75-100 samples) ,and easy accommodation in a low-cost,stand-alone microprocessor system. T he ver-ification time is also very low (4-6 sec).More sophisticated verification systems 161may yield somewhat higher performance buttheir practical implementation is very diffi-cult. The system presented here has alreadybeen tested with excellent results. The blockdiagram of the verification system is shownin Fig. 1.

Signature Verification TechniqueA general signature verification system

consists of four parts: transducer, compara-

tor, reference file, and decision logic. Themost striking aspect of signature dynamicsis that the time interval for writing a signa-ture, measured from start to finish, remainsremarkably consistent. The time for succes-sive signatures frequently differs by as littleas 10msec. V ariation in signature dynamicsdoes take place as years pass, but this changecan be handled by dynamic updating of therecords. Even muscle pressure variation onthe writing surface is consistent during thesignature process. Based on this consiste ncy,we propose an automatic verif ication schemeusing these concepts. The com plete processof verification consists of the following fourparts. (The interested reader is referred to[4]-[6] for more detailed mathematical treat-ment .)

Time interval for writing signature. Fo rdecision making, a duration variation of10-15 percent is permitted. T he nominaltime for a signature is taken to be 4. 0sec.Number of pressure pe aks in signature.In [SI, t was shown that the number ofpressure peaks and valleys are well de-fined in a signature. T he numbe r of peaksalong with the mean an d standard valuesis selected as another important featurefor the verification criterion.Distance measure of peaks and valleys.Distance measure refers to the differencein coordinates between the peaks andvalleys. This criterion involves dividingthe waveforms into segments of differentlengths according to the peaks and val-leys found in the waveforms. This cri-

r-5eferenceFig. 1. Block diagram of an automatic verif ication system.

5 A0272-170818810600-0054 $0 1 00 0 988 IEEE I E E E Control Systems Mogozine

7/27/2019 00000478

2/4

tenon is composed of four submeasures.Val lej row is the difference in the rowcoordinates of all the valleys between asample signature and a reference signa-ture. Val ley column is the difference inthe column coordinates of the two sig-natures. Peak row an d peak column ar edefined similarly.

(4 ) Area measure between peaks and val -l e y . Area measure refers to the differ-ence between slopes or segments be-tween a peak an d a valley. Th is criterionwas developed using the principle ofsegmentation and correlation. Thismethod of waveform correlation is wellknown and gives good results. To im-prove performance, overlapping seg-ments can be used as an alternative.

For an on-line and interactive verificationsystem. the correlation algorithm must beable to verify within a few seconds. T o iden-tify a person, the system would first ask fora personal identification number, or PI Nnumber. After the person has entered the PINnumber correctly, a signature request will bemade. The algorithm uses the proposed ver-ification criterion for waveform correlation.

Threshold SettingThreshold setting is done individually for

all the verif ication parameters. T he thresholdis set to maximize the percentage acceptanceof true signatures and, at the same time, tomaximize the percentage rejection of forger-ies.

To m aximize the percentage acceptance oftrue signatures, the Less Than CumulativeDistr ibution Function of data is gatheredfrom true signatures. This curve gives thepercentage of the signees having signatureswith parameter variation less than the nom-inal value. Similarly, to maximize the rejec-tion of forgeries, the More Than C umulativeDistr ibution Function of data gathered fromforgeries is plotted. T he objective is to max-imize both functions. A typical plot andthreshold setting for one of the verificationparameters are shown in Fig. 2. This typeof threshold is determined for all the verifi-cation param eters. These threshold values arethe maximum allowable tolerances in veri-fication parameters during the verificationprocess. Rese arch work currently in progresswill be required for optimal selection of thethreshold values of these verification param-eters. Computer-generated outputs for thesignature waveform and for one of the ver-if ication parameters are shown in Figs. 3an d4 , respective ly .

Threshold setting (third criterion)Submeasure Valleykolumn100

80 -YI___-70-55-

0 500 1000 1500 2000 2500 3000 3500 4000Deviation in column coord. (valley)-True signa ture cumulative distributionfunction.-Forgery CDF

Fig. 2 . Threshold setting.

12 5

20

I 5

Fig. 3. Pressure waveform of a signature.

/ = 3 J = l

Irea difference = 96.42 JFig . 4 .slopes.

Area difference of two rising

System HardwareFrom a functional point of view, the hard-ware of the system may be partitioned into

three main subsystems.(1 ) Microcomputer as Central Controller ,

(2 ) On-Line Acquisition System, and(3) System Interface Unit (SIU) .

The microcomputer selected for this sys-tem is the IBM-PCIAT. The SI U is the in-terface between the computer and the on-lineacquisition system. I t handles bidirectionalinformation flow between the two functionalunits. The on-line acquisition system is theterminal by which the end user enters thesystem, I t comprises the pressure transducercharge amplif ier ( to capture the signature),the key pad (PIN entry), and liquid-crystal-display and light-emitting-diode modules(status indicators) , A piezoelectric trans-ducer was used to sense the pressure wave-form and a charge amplif ier was used forsignal conditioning. The analog-to-digitalconversion (ADC) circuitry was centered onthe ZN488 8-bit successive approximationchip, which wa s wired for bipolar operation.This was necessary because the piezoelectrictransducer responds to dynamic pressurevariations, giving bipolar readings.

System SoftwareThe software was developed to implement

the following six functions:(1) Capture sampled values of the signature

waveform followed by simple data pro-cessing.

(2 ) Devise a file system to store captureddata .

(3) Process data using a correlation algo-rithm to match the reference signaturewaveform.

(4) Report various status messages.(5) Create a data base for storing and re-(6) Plot waveforms, whenever necessary.

tr ieving, whenever necessary.

At the center of the software domain ofthe system is the on-line, menu-driven pro-gram MAIN. The funct ion o f MAIN is tocreate a user-fr iendly environment and co-ordinate activities among various modules.Functions of a few of the important subpro-gram modules are described here briefly.

Datu Cupture SubprogramThe task of this subroutine is to capture

the signature waveform. The first task of theroutine is to initialize the COMI module,which accepts data transfer at 600 bps; thesecond task is to act iva te ADC to conver tthe analog pressure waveform into discreteform. On the average, the system can acceptapproximately 75- 100 samples over a 5-secperiod. The sample values are concatenatedto form a character str ing. Fo r each signature

June 1988 55

7/27/2019 00000478

3/4

reference, three reference waveforms areneeded.

Rrwirrl Fetch SubprogrrrtnThe task of this routine is to retrieve ap-

propriate records from the data base. Theaccess is through PIN number.

Dutrr Cot1 Yr f s l l b / J I ' ~ J , q'll I ?1This subprogram reforniats the record\ ofa person into a fo m i witable for funher dataprocessing. This routine is called t o perfomia best-fit reference uaveforni o r to preparea signature for correlation u th the be\t-fitreference.

o/l~;~lll2. lCIl l l Slrt?/?ro,qI'cllllThe menu option ih used t t i obtain a

graphic plot of the reference \ignature. toupdate the signature records. o r to add a newsignature to th e data b a x Fo r the latter ap-plication. three reterence signatures are re-quired.

Drift1 P I ' O C.~ \ . \~ i h p r o , q r r m ~This subprogram p e t f o r n i ~ orrelation be-

tween two wavefornis. The criterion of cor-relation is based on the tollowing verificationparameters. described previously: time in-t e n d l o r a signature. number o f peaks inthe signature, distance measure of peaks andvalleys. and area measure between peaks andvalleys. A s an option-to impro ve perfor-mance-the featurelike mea n. standard de-viation. and the number of zero crossings ofa third verification parameter maq be used.

Experimental ResultsFor the experimental system. 200 signa-

tures were collected from SO volunteers. Tosimulate a real application environment. eachindividual was asked to sign four times toprovide sample signatures needed for the ref-erence data base. In subsequent sessions,each individual was asked to sign again forverification. Forgers were also encouraged.The forgeries were not casual. but part of anattempt t o exercise the verif ication schemevigorously. The s ystcm could verify true in-dibidual signatures without any failure.

Figures 5 an d 6 provide output w aveformslo r a true signature (average) and a besttorgecl 5ignature. respectively. In Fig. 5 .pressure waveforms of a person's true sig-natures are plotted. From four sample sig-natures. the nornial values of verification pa-raineter\ are calculated and stored in the,ignature data base. When this person wants

a c ~ ' e \ \ his system again. an additional.ipnature i'r required. The new verification?.lr;nieten ar e compared with storcd values.

Reference signature pressure waveforms Reference signature pressure waveforms50

40

30

2010

0

0 1 2 3Time scale (sec)- est-fit referen ce - rial-run signature -10-1 2 3Time scale ( sec)= rial-run signatures

Fig. 5 . Reference s ignatu re waveform - Fig . 6 . Reference s ignatu re wavefom s-true signature. forged signature.

If the difference in verification parameters iswithin tolerance values, as shown in Table1. the person is accepted by the system, andthe test parameters will be updated for futureuse. If the parameter values are above tol-

erance values, the person is rejected. Criticalcases are asked to sign again. A rejected per-son may be given a maximum of threechances.

A forger may try to copy the visible sig-

Table 1Correlation Results of True Signature

Correlation SessionDate : 02-05- 1987Time: 16:03:56Account NumberTime Duration of Sample, secTime Duration of Reference, secNo. Peaks in SampleNo. Peaks in ReferenceDistance [Valley (Row)] [2000]Distance [Peak (Row)] [3000]Distance [Valley (Co l)] [5001Distance [Peak (Col)] [5001Area Measure of Sample [2000]

Signature Accepted

21002.802.50

99

936.0079.50135.00

102.00803.60

Nominal Threshold ValuesSignature Time Variation 15 Percent of

NominalValue

15 Percent ofNominalValue

Difference in PeaksiValleys

Other Values A re Specified in B rackets Along withMeasured Values

I Control Sys tems Mogorine

7/27/2019 00000478

4/4

Table 2Correlation Results of Forged Signature

grateful to the Applied Research Fund forproviding financial assistance.

Correlation SessionDate: 02-05-1987Time: 16 :24 :16Account Number 2100Time Duration of Sample, sec 3.33

2 . 5 0ime Duration of Reference, secNo. Peaks in SampleNo. Peaks in Reference

79

Distance [Valley (Row)] [2000] 3049.0Distance (Valley (Col)] [500] 5012 .25D i m n c e [ Pe ak ( Row ) ] [3000] 62 4 1 OODistance [Peak (Col)] [500] 4 167 OOArea Measure of Sample [2000] 3299.31

Signature Rejected

nature. but the forger can never get close tothe nominal parameter values of the dynamicpressure of the true signature waveform.Therefore, the forgery will never be ac-cepted by the system. Of course, by acci-dent, one person's verif ication parametersmay come within the tolerance range of an-other person's param eters. Howeve r, in suchcases, the two people will have difTerent PINnumbers. The probability of matching thePIN number as well as the nominal param-eter values is quite low.

During testing, the forgers could nevercome close to the true signature verif icationparameters. In fact, during our testing ses-sion, no forger could break into the system.A person with the wrong identity is given amaximum of three chances before access isdenied. In Fig. 6, a forger 's pressure wave-forms are shown, which are completely dif-ferent than the true signature waveform inFig . 5 . The nominal verif ication parameter

True signatures Forged signaturesFig . 7 .tures.

Samples of true and forged signa-

values for the forged signature are shown inTable 2.

Some samples of true and forged signa-tures are shown in Fig . 7 . It is our belief thatthis verification technique will work for sig-natures in any language.

ConclusionThis paper describes an on-line signature

verification system used to identify a person.The signature verif ication technique is cou-pled with a personal identification number.An important aspect of the verification tech-nique is that it depends on the dynamic pres-sure of the signature waveform, w hich is dif-ficult to forge. The verification algorithm hasbeen simplified, so that it is possible to ver-ify a signature in approximately 4-5 sec. Theone-time enrollment session requires threereference signatures and lasts about 30 sec .

The system was tested ov er a six-week pe-riod. with a total of 200 signatures collectedfrom 50 individuals. The on-line system wasable to identify true signatures and forgerieswithout error. Despite many intrinsic hu-man-factor problems, w e believe our perfor-mance figures are sufficiently encouraging toindicate the feasibility of signature verifica-tion as a means of personal identification.

AcknowledgmentsThe authors would like to thank Professor

Brian Lee, Dean, EEE, Nanyang Techno-logical Institute, for providing facilities forcarrying out the project. The authors a re also

ReferencesG. Lorrette, "On-Line Handwritten Signa-ture Recognition Based on Data-Analysisand Clustering." Laboratorie de GenieEletrique de Creteil. Universite Paris, C re-teil, France 1984.W . F. Nemcek and W . C. Lin, "Experi-mental Investigation of Automatic Signa-ture Verification.'' IEEE Truti.7. Syst., M u t i ,Cyberti . , Jan. 1974.C . N . Lin and N. M. Herbst. "AutomaticSignature Verification Based on Acceler-ometry." IB M J. Res. Det,c,lop., vol. 21 .no. 3 , p. 245, May 1977.C. N . Lin, N. M. Herbst. and N . J . An-thony, "Automatic Signature Verification:System Description and Field Test Re-sults." IEEE Trun.\. Sysr. , Muti. Cyberr i . .vol. SMC-9. no. I . Jan. 1979.H. D. Crane and J . S . Ostrem, "AutomaticSignature Verification Using a Three-AxisForce Sensitive Pen," IEEE Trutis . Syst . ,Man, Cyber t i . . vol. SMC-13, MaylJune1983.J . P. De Figueiredo and C-L. Hu . "Wave-form Feature Extraction Based o n Tauber-ian Approximation." IEEE Trutis. Purr.A n d y . Machine Intell., vol. PAMI-4, no.2, Mar. 1982.Y-C. Cheng and S-Y. Lu . "Wavefomi Cor-relation by Tree Matching," IEEE T r m s .Pur:. Atiuly. M u d i i t i e Ir itc41. , vol. PAMI-7,no. 3, May 1985.S . Y. Lu . "A Tree-Matc hing AlgorithmBased on Node Splitting and Merging."IEEE Trans. Putt. A i iu /y . Machitic, l r i r c J l l .vol. PAMI-6, no. 2 , Mar. 1984.

Dinesh P. Mital receivedthe B.Tech. degree inelectrical engineeringfrom IIT. Kanpur. India,in 1968. Thereafter. hereceived the M.S. andPh.D. degrees f rom theState University of NewYork, Stony Brook, NewYork, in 1970 and 1974.respectively. From 1974to 1976, he worked forNCR Corporation, Day-ton. Ohio, as an Advanced Development Engi-neer. Thereafter. from 1976 to 1983. he workedas Assistant ProfessoriProfessor at University ofRoorkee, Roorkee, India. Currently. he is work-ing as a Senior Lecturer at Nanyang Technologic alInstitute, Singapore. His current areas of researchinclude microprocessor applications. digital con-trols. robotics, and artificial intelligence. He haspublished over 40 technical papers in related areas.

Ju n e 1988 57