Dynamic Signature Forgery andSignature Strength Perception Assessment

Stephen Elliott & Adam HuntPurdue University

ABSTRACT

Dynamic signature verification has many challengesassociated with the creation of the impostor dataset. Theliterature discusses several ways of determining theimpostor signature provider, but this takes a differentapproach - that of the opportunistic forger and his or herrelationship to the genuine signature holder. Thisexamines the accuracy with which an opportunistic forgerassesses the various traits of the genuine signature, andwhether the genuine signature holder believes that his orher signature is easy to forge.

INRODUCTION

Dynamic signature verification (DSV) has long been usedto authenticate individuals based on their signingcharacteristics, such as speed, pressure, and graphical output.Approaches to DSV have been discussed in detail in theliterature. Popular applications, such as documentauthentication, financial transactions, and paper-basedtransactions have all, at one time, used the signature toconvey the intent to complete a transaction [1, 2]. DSV is asubset of a larger science called biometrics. Biometrics aimsto authenticate an individual based on either behavioral orphysiological traits, (or a combination of both), includingface recognition, iris recognition, and fingerprint recognition,to name a few. Many of these modalities are made up of bothbehavioral and physiological attributes, with variousproportions of each. Within the continuum, the signature is astrong behavioral biometric. The signature's unique traitsmake it harder to test and evaluate than some of the otherbehavioral biometrics, such as voice or face recognition.Challenges to testing and definitively evaluating the signature

Author's Current Address:S. Elliott, Ph.D., and A. Hunt, Associate Professor, Department of Industrial Technology,College of Technology, Purdue University, 401 N. Grant Street, West Lafayette. IN 47906,USA.

Based on a presentation at Carnahsan 2006.

0885/8985/08/ USA $25.00 0 2008 IEEE

include the fact that a signature is learned over time (andevolves over time as the owner and his or her handwritingmatures), it contains variant measures (such as pressure,speed, etc., that can be changed), can be changed by theowner (depending on the ceremony of the transaction), andmay have several versions (for example, at work and homemay have different signatures).

A discussion of DSV invariably raises a number ofconcerns. The first concern is that people acknowledge theirfailure to sign consistently, and the second is that mostpeople have attempted, irrespective of degree of success, toforge someone's signature at some time. In fact, a straw pollconducted in a class of 80 undergraduate college studentsrevealed that at least 90% of them have attempted to forge asignature at one time. When asked for more details about theforgery attempt, in the majority of cases, the subject of theattack is someone who is known to the forger (typically aparent or close relative) and the signature is easily available.In these cases, it is more than likely that the forger has hadseveral chances to practice the signature and that thesignature is not rigorously checked by the receiver of thedocument being forged. These two conditions correlate to alow chance of getting caught - this is the scenario of theopportunistic forger.

Another important consideration has to do with whether agenuine signature holder believes that his or her signature isdifficult to forge, and whether the imposter also believes thatto be the case. The approach, proposed herein, is tounderstand whether the impostor can actually makewell-informed decisions on the measurable variables of thegenuine signature. For example: Can the forger determine thespeed of the signature, as well as the handedness of thegenuine signer? If the forger can determine these most basicof attributes, then he or she might then achieve some level ofsuccess to forge some of the additional variables within DSV.

VARIABLE CHARACTERISTICS OF THE DSV

DSV's numerous variables are calculated using the inputgathered from a digitizer. These variables include x and y(Cartesian) coordinates, pressure (p) or force, and time (t)[3]. This output from the digitizer is used to create the global

IEEE A&E SYSTEMS MAGAZINE, JUNE 2008 113

Authorized licensed use limited to: Purdue University. Downloaded on December 14, 2009 at 17:26 from IEEE Xplore. Restrictions apply.

Table 1. DSV Studies & Error Rates

Study Authors Type I (FAR)

Achemlal, Mourier, Lorette & Bonnefoy (1986)Beatson (1986)Bechet (1984)Bonnefoy & Lorette (1981)Bault & Plamondon (1981)Collantier (1984)Crane & Ostrem (1983)Debruyne (1985)Hale & Pagnini (1980)Herbst & Liu (1979)Herbst & Liu (1979)Ibrahim & Levrat (1979)Lam & Kamis (1979)Lorette (1983)Mital, Hmn & Leng (1989)Parizeau & Ilamondon (1989)Sato & Kogure (1982)Worthington, Chainer, Williford

& Gundersen (1985)Zimmerman & Varady (1985)Cordelia, Foggia, Sanson & VentoWirtz (1997)Dimauro, Impedovo & Pirlo (1993)Naiwa (1997)Naiwa (1997)Nalwa (1997)Mingming & Wijesoma. (2000)Mingming & Wijesomna (2000)Mingining & Wijesoma (2000)Hamilton, Whelan, McLaren & Macintyre (1995)Hamilton, Whelan, McLaren & Macintyre (1995)Hamilton, Whelan, McLaren & Maclntyre (1995)Hamilton, Whelan, McLaren & Macintyre (1995)Martens & Clausen (1997)Chang, Wang & Suen (1993)Higashino (1992)Minot & Gentric (1992)Lucas & Damper (1990)Tseng & Huang (2002)Lee, Berger & Aviczer (1996)Lee, Berger & Aviczer (1996)Lee, Mohankrshnan & Paulik (1998)Han, Chang, Hsu & Jeng (1999)Komiya & Matsumoto (1999)Cardot, Revenu, Victorri & Revillet (1993)Cardot, Revenu, Victorri & Revillet (1993)

11.0%1.0%5.0%0.6%1.2%3.5%1.5%3.0%1.5%1.7%2.4%19.0%0.0%6.0%0.0%4.0%1.8%

1.8%30-50%

0.03-20.82%

1.7%

34.0%22.0%12.0%7.0%1.5%2.0%8.0%2.0%5.6%

12.5-28.8%1.0%5.0%0.9%4.0%

2.0%0.9%

14 IEEE A&E SYSTEMS MAGAZINE, JUNE 2008

Error RatesType 11 (FRR)EER

8.0%2.0%5.0%

1.0%

1.5%2.0%

1.2/2.5%0.4%3.4%5.5%2.5%

0.0%

0.0%

0.28-2.33%4-12%5.7%

10%1.2%

3%2%5%5%7%9%

26.0%18.0%10.0%6.0%1.3%2.5%0.6%4.0%4.5%

5.0-12.5%20.0%20.0%0.7%7.2%

2%4.0%7.4%

14

Authorized licensed use limited to: Purdue University. Downloaded on December 14, 2009 at 17:26 from IEEE Xplore. Restrictions apply.

and local features described in various accounts [4, 51. Theseglobal and local features, derived from the basic set of datathat a digitizer provides, vary significantly across algorithms.[6] outline 61 features, [7] note over 40 features, [8] discuss44 features. In [2], describes the x and y coordinates of thepen's motion. Also [9], observes that the temporal"characteristics of the production of an on-line signature arethe key to the signature's verification" (p. 5). From thesemany approaches and features, the number of variablesassociated with a dynamic signature can be synthesized toseveral major statistical features. These major statisticalfeatures include pressure, time, horizontal, and verticalcomponents of position, velocity, acceleration, and force, allmeasured against time. An alternative approach tocharacterizing a signature involves analysis of the "stroke,"that is, the up and down movements of the pen on thedigitizer. The many dynamic traits collected by the digitizerduring the act of signing are said to make an impostorsignature easier to detect than that of a traditionalpaper-based impostor signature.

HIMPOSTOR SIGNATURE GENERATION

Impostor datasets are created in numerous ways and havethe effect of changing the respective performances ofalgorithms. This change can be done through the differentgeneration of the impostor signatures. A review of theliterature shows various performance results from severalstudies, all of which have different methodologies forcollecting impostor signature datasets. Table 1 outlines thevarious studies and their respective error rates (false accept,false reject, and the equal error rate where appropriate).

The variances in error rates shown in Table 1 (0% to 50%false accept rate and 0% to 20% false reject rate) can beexplained by a number of factors, one of which has to dowith how an impostor signature dataset is created. [10]'sstudy is particularly interesting. This study had a databaseconsisting of 293 genuine signatures and 540 forgerysignatures from eight individuals. Although the study did notexplain how the forgery took place (in terms of training,payment, etc.), eight individuals created the impostor dataset.[11] study dataset was comprised of 496 original signaturesfrom 27 people. Each person signed I11 to 20 times. Thedatabase contained 48 forgeries that "fulfill the requirementon the visual agreement and the dynamic similarity with theoriginal signature " (p.5). [12] trained the algorithm using250 signatures per writer; of these 250 signatures, 100 wereauthentic signatures and 150 were random forgeries,classified as the genuine signatures of other writers. [ 13] used27 people in their study, with the participants writing theirown signature. The study also used 4 people who imitated thesignatures of these 27 people. Unfortunately, no furtherinformation is provided on the selection of the impostor or onwhat knowledge the forgers possessed in order to forge thesignatures.

[1] used genuine signatures from other individuals asforgeries. In addition, a group of synthesized signatures was

created by distorting real signatures through the addition oflow-level noise and dilation or erosion of the variousstructures of the signature. [14] motivated the forgers byoffering a cash reward. [15] examined people's signaturesover a four-month period to assess variability over time. Inthe Signature Verification Competition, genuine signerscreated signatures other than their own [16]. In [17], theauthor uses a number of different methodologies to generatethe impostor distribution, with the majority of impostorsusing some form of practice. In [ 18], the authors definedthree different levels of forgeries: the simple, staticallyskilled, and timed (p. 643). [19] used signatures that "oncasual visual inspection would pass as authentic" (p. 201).[20] provides three characteristics of forgery: the randomforgery, defined as one that belongs to a different writer ofthe signature model; simple forgery, represented by a similarshape consistency with the genuine signer's shape; and theskilled forgery (p. 2).

PERCEIVED STRENGTH OF SIGNATURE (PSS)

The purpose of this paper is to assess the basic attack on asignature by an opportunistic forger and to determine theperceived strength of the signature (PSS). PSS is a conceptthat indicates that an opportunistic forger will not forge asignature that is difficult to forge, as their success at thepoint-of-sale may be not as high as the forgery of an easysignature. This is more the trademark of an opportunisticforgery than of a more sophisticated attack on the signature,as outlined in previous research. For this study, anopportunistic forger is analogous to an opportunistic thief,that is, one who works on his or her own without anyequipment [211]. This definition is further enhanced by theabsence of occasion to practice forging the signature. Thestudy outlines a basic truth involving the genuine signatureowner's perception of the strength of their signature and triesto understand whether the owner of a genuine signature hasthe same or different perception of the signature than that ofthe forger.

In order to understand the basic truth of the perceivedstrength of the signature, each of the genuine signatureowners was asked for the following information about theirsignature:

1. How easy their signature was to forge (rated ona Likert scale).

2. How fast or slow they signed their signature(rated on a Likert scale from slow to fast).

3. Handedness (right- or left-handed;ambidexterity was not an option captured by thesurvey).

The objective in obtaining these three pieces ofinformation was to assess whether the forger was able to

IEEE A&E SYSTEMS MAGAZINE. JUNE 2008 115

Authorized licensed use limited to: Purdue University. Downloaded on December 14, 2009 at 17:26 from IEEE Xplore. Restrictions apply.

predict the speed and forgeability of genuine signatures,which was typically centered on the dynamic traits of speedor velocity, time, and graphical outline or complexity of theshape. Furthermore, subjects were asked to sign their nameon a digitizer so that feature variables could be extracted toestimate whether there were any correlations betweenvariables and the respective PSS categories.

Table 2. T-Test for Difficulty Groups

METHODOLOGY

In order to assess PSS, two separate groups wereorganized. The dataset of genuine signatures was collectedfrom consent forms signed by the genuine signature owners.The consent forms were used to maintain a level ofceremony, since the consent form is a document that requiresa signature with a level of intent, as opposed to a randomsignature with no intent. This signature was subsequentlyused as the target signature. In order to estimate whether anyof the dynamic signature verification variables were the samefor each group (those who ranked their signature within thesame Likert classification), each subject signed his or hername on a digitizer three times. In order to obtain aconsistently precise signature, the study utilized an InterlinkElectronics ePad-ink ProTM1 device, which has 100-400 reportsper second and 300 dots per inch [22].

The device was connected to forensic signature softwareto extract the raw data from the digitizer, but the subjectscould not see the signature or the information on the PCmonitor as they signed. The digitizer provided an inkeddisplay of the signature as the subject signed his or her name.The three signatures were then processed and the resultantvariables averaged across the signatures.

The impostor group consisted of individuals other thanthose who owned the original signatures. Members of theimpostor group were asked for information about what theyobserved while looking at the signed consent form of eachindividual in the genuine group:

1. How easy the genuine signature was to forge(rated on a Likert scale).

2. How fast or slow was the genuine signatureestimated to be made (rated on a Likert scalefrom slow to fast).

3. What was the handedness of the genuine subject(right- or left-handed; ambidexterity was not anoption captured by the survey).

The results were analyzed statistically to determinewhether any significant differences existed between thegenuine signature owners and the impostors regarding theirassessments of the signature.

RESULTS

For the genuine dataset, a total of 60 subjects participated,of which 1 was female and 59 were males. Of these 60subjects, 36 signed the digitizer. The remainder did not sign(or dropped out of the study). This represents a retention ratein the study of 60%. For the impostor dataset, there were 9individuals who ranked the genuine signatures using theparameters previously described.

A t-test was used to determine whether the mean of thegenuine and impostor groups were statistically significantfrom each other with regards the three questions posed: therank of the perceived level of difficulty, velocity, andhandedness. A level of 0.05 was selected for determiningstatistical significance. In the study, the data were normallydistributed, and there were no outliers. The data were rankedfrom 1 to 5, with 1 being easy to forge and 5 being difficultto forge.

Table 3. Signature Speeds

When assessing the PSS, a t-test result shows parallelsbetween genuine subjects and forgers when the level ofdifficulty was assessed as "neutral." Likewise, the differencein the "very difficult" ranking has a p-value of 0.053. Othercategories (levels 1, 2, and 4 in the Likert scale) exhibitedsignificantly different means. It is difficult to determinewhether the groups were statistically significant. Furtherrefinement of the question is needed (and will be undertaken

IEEE A&E SYSTEMS MAGAZINE, JUNE 2008

Speed p-value

1 0.002

2 0.0113 0.060

No group 4 N/A

5 0.035

Authorized licensed use limited to: Purdue University. Downloaded on December 14, 2009 at 17:26 from IEEE Xplore. Restrictions apply.

Table 4. Handedness

Handedness p-value

Left 0.000

Right 0.000

in a subsequent study) to examine the PSS in more detail.Table 2 outlines the results of the PSS category.

When analyzing speeds (see Table 3, where 1 denotes fastspeed and 5 denotes slow speed), both groups had significantdifferences, except for the neutral standing. Neither groupassigned a rank of 4 as a speed.

The results indicate that neither the impostors nor thegenuine users could determine the speed of each other'ssignatures. This is particularly interesting, as speed (orvelocity) is an often used a statistical feature in DSValgorithms.

When analyzing handedness, the impostor group could notconsistently determine the handedness of the genuinesignature owners. Only 3 out of 8. 37.5%, of the forgerscorrectly identified a left-handed signature, while 3 of theleft-handed signatures were not correctly identified at all.Comparably, 7 of 49 right-handed signatures were correctlyidentified by all forgers. However, the least accurate resultsshowed that 5 of 8 forgers incorrectly identified a signatureas left-handed when it was, in fact, right-handed. Theseresults are represented in Table 4

The last question posed is whether the dynamic featuresextracted from the digitizer were similar for each group ofthe PSS categories. For example: Do those in theeasy-to-forge category exhibit the same speed? Is there anunderlying dynamic variable within these groups that areselected by impostors as easy to forge?

An Analysis of Variance (ANOVA) test was conductedover all of the individual variables that were extracted fromthe digitizer. At a = 0.05, none of these variables weresignificantly different across each difficultly group. For theforger group, the ANOVA showed no significance with theseextracted variables and difficulty group. There were someinteresting correlations, however; speed was negativelycorrelated with difficulty (-0.191, with p-value 0.273), aswere the number of strokes and difficulty of -0.314. Theforger groups had a slightly positive correlation with speedand slightly negative correlation with segments (0.081).

CONCLUSION

The purpose was to assess whether genuine and impostorgroups could successfully predict variables that could aid inthe successful forgery of the genuine signature. Variablesincluded the perceived strength of the signature, speed, and

IEEE A&E SYSTEMS MAGAZINE, JUNE 2008

handedness. The results indicate that genuine signers andimpostors did not rank signature within the same strengthcategories, that the impostors could not determine speed ofthe genuine signature, and that the impostors could notdetermine handedness. Furthermore, there were no commoncharacteristics of the signature variables within the groups.Further research should be undertaken to examine whetherthese attributes change as the forger gains more knowledgeabout and experience with the signature.

REFERENCES

[I] D.J. Hamilton, J. Whelan, A. McLaren, 1. Maclntyre and A. Tizzard,Low cost dynamic signature verification system,

presented at 1995 European Convention on Security andDetection, Brighton, UK, 1995.

[2] WA. Nelson and E. Kishon,Use of dynamic features for signature verification,

presented at 1991 IEEE International Conference on Systems,Man, and Cybernetics, Decision Aiding for Complex Systems,Charlottesville, VA, 1991.

[31 C. Vielhauer, R. Steinmetz and A. Mayerhofer,Biomnetric hash based on statistical features of online signatures,

presented at 1 6' International Conference on PatternRecognition, 2002.

[4] F. Leclerc and R. Plamnondon,Automatic Signature Verification: The State of the Art - 1989-1993,

Singapore: World Scientific Publishing Co., 1994.

[5] J.-J. Brault and R. Plamiondon,A Complexity Measure of Handwritten Corves: Modelingof Dynamic Signature Forgery,

IEEE Transactions on Systems, Man, and Cybernetics,Vol. 23, pp. 400-413, 1993.

[6] M.C. Fairhurst and S. Ng,Management of access through biometric control: A case studybased on automatic signature verification,

Universal Access in the Information Society,Vol. 1, pp. 31-39, 2001.

[71 A. Kholmatov and B. Yanikoglu,Biometric Authentication Using Online Signatures,

presented at W9, International Symposium on Computer

and Information Sciences - ISCIS 2004,Kemer-Antalya, Turkey, 2004.

[8] H.D. Crane and J.S. Ostem,Automatic Signature Verification using a Three-axisForce-Sensitive Pen,

IEEE Transactions on Systems, Man, and Cybernetics,Vol. 13, pp. 329-337, 1983.

[9] V. Nalwa,Automatic On-Line Signature Verification,

Proceedings of the IEEE, Vol. 85, pp. 215-239, 1997.

[10] M. Komiya Y.T.,On-line pen input signature verification PPI (pen-position/

17

Authorized licensed use limited to: Purdue University. Downloaded on December 14, 2009 at 17:26 from IEEE Xplore. Restrictions apply.

pen pressure / pen inclination),presented at IEEE International Conference on Systems,Man, and Cybernetics, 1999,IEEE SMC '99 Conference Proceedings. 1999Tokyo, Japan, 1999.

[I I] C. Schmidt and Kraiss, K-F.,Establishment of Personalized Templates for AutomaticSignature Verification,

presented at International Conference on DocumentAnalysis and Recognition, 1997.

[12) W.S. Lee, Mohankrishnan, N. and Paulik, M.,Improved Segmentation through Dynamic Time Warping ForSignature Verification using a Neural Network Classifier,

presented at the 1998 International Conference on ImageProcessing, 1998. ICIP 98. Proceedings, 1998.

[13] Q.-Z. Wu, Jou, I-C. and Lee, S-Y,On-Line Signature Verification using LPC Cepstrum andNeural Networks,

IEEE Transactions on Systems, Man, and Cybernetics,pp. 148-153, 1997.

[141 M. Mingming and Wijesoma, W.,Automatic On-Line Signature Verification Based onMultiple Models,

presented at Computational Intelligence in FinancialEngineering Conference, 2000.

[15] P.-C.C. Chin-Chuan Han., Chao-Chih Hsu, and BorShenn Jeng,An on-line signature verification system using multi-templatematching approaches,

presented at Security Technology, 1999 Proceedings,IEEE 33'ý Annual 1999 International CarnahanConference, 1999.

[16] D-Y. Yeung. H. Chang, Y. Xiong, S. George, R. Kashi,T. Matsumoto and G. Rigoll,SVC 2004: First International Signature Verification Competition,

presented at First International Conference on BiometricAuthentication, ICBA, Hong Kong, China, 2004.

[17] S. Elliott,A Comparison of On-Line Dynamic Signature Trait Variablesvis-A-vis Mobile Computing Devices and Table-Based Digitizers,

in Third Workshop on Automatic Identification AdvancedTechnologies. Tarrytown, NY: IEEE, 2002.

[18] L. Lee, Berger, T. and Aviczer, E,Reliable On-Line Human Signature Verification Systems,

IEEE Trans. Pattern Analysis Machine,Vol. 18, pp. 643-647, 1996.

[19] W. Nelson and Kishon, F.,Use of Dynamic Features for Signature Verification,

presented at Proceedings IEEE International Conference onSystems, Man, and Cybernetics,Charlottesville, VA, 1991.

[20] E.J.R. Justino, Bortolozzi, F. and Sabourin, R.,The Interpersonal and Intrapersonal Variability Influences onOff-Line Signature Verification using HMM,

presented at Proceedings of the XV Brazillian Symposiumon Computer Graphics and Image Processing,(SIBGRAPHI'02), 2002.

(2 1] D. Cvrcek, Krhovjak, J. and V. Matyas,PIN (& Chip) or signature -beating the cheating?

Bmo, Czech Republic, 2005.

[22] 1. Electronics, 2006,E-Pad Signature Pad Specification Sheet. A

18 IEEE A&E SYSTEMS MAGAZINE, JUNE 200818

Authorized licensed use limited to: Purdue University. Downloaded on December 14, 2009 at 17:26 from IEEE Xplore. Restrictions apply.