The Use of Static Biometric Signature Data from Public Service Forms Emma Johnson and Richard Guest School of Engineering and Digital Arts, University of Kent, Canterbury, UK {ej45,r.m.guest}@kent.ac.uk Abstract. Automatic signature verification/recognition is a commonly used form of biometric authentication. Signatures are typically provided for legal purposes on public service application forms but not used for subsequent biometric recognition. This paper investigates a number of factors concerning the use of signatures in so-called static form (an image of a completed signature) to enable the sample to be used as stand-alone or supplementary data alongside other biometric modalities. Specifically we investigate common sizes of unconstrained signatures within a population, assess the size of application form signing areas with respect to potential constraints and finally investigate performance issues of how constrained and unconstrained enrolment signature data from forms can be accurately matched against constrained and unconstrained verification data, representing the full range of usage scenarios. The study identifies that accuracy can be maintained when constrained signatures data is verified against other constrained samples while the best performance occurs when unconstrained signatures are used for both enrolment and verification. Keywords: Signature biometrics, static feature analysis, form design. 1 Introduction The human signature is a widely used and accepted form of personal authentication with application areas spanning a multitude of everyday domains (including retail and legal) [1],[2],[3]. Signature also has widespread use within automatic biometric analysis solutions alongside other modalities such as fingerprint, face and iris recognition [4]. Despite comparable performance in terms of verification error rates, signature does not enjoy the market share of other modalities within large scale deployment applications such as national identity cards or border documentation (for example passports or visas) for which face and fingerprint are often primary choices [5]. These modalities require specific capture and enrolment equipment often meaning that enrolees have to attend a specific site (in the case of fingerprints) or donate samples to a particular predefined specification (in the case of facial images). Applications for official identify documentation are almost always made using a paper form on which the applicant has to sign (usually for legal purpose regarding use of data and consent). Using this signature data in a static/image format may be fully integrated C. Vielhauer et al. (Eds.): BioID 2011, LNCS 6583, pp. 73 82, 2011. Springer-Verlag Berlin Heidelberg 2011
74 E. Johnson and R. Guest into the verification process as additional data readily available to enhance the biometric process alongside the primary biometrics collected for the application [6]. This study aims to assess several key aspects of the use of this supplementary static biometric signature data collected on application forms. Specifically, we firstly focus on the design of common application form signing areas with respect to the range of human signature sizes (both Western and non-western) to investigate if constraints are imposed which will affect normal signature production. By using a specially collected signature dataset we will secondly investigate verification performance rates of constrained and unconstrained static signatures using a common static signature verification engine. This second experiment will lead to an understanding of performance issues as to how both constrained and unconstrained (enrolment) signature data from forms can be accurately matched against constrained and unconstrained verification data, representing the full range of usage scenarios. 2 Methodology To investigate the effects of signature size, form-based constraints and verification performance, three separate stages to the experiment were conducted. 2.1 Stage 1 Unconstrained Signature Size Analysis Throughout the investigation a new dataset was used which comprised of 150 signers who were asked to donate four separate instances of their signature on individual blank A4 sheets of paper, which, when scanned, formed an unconstrained subset of static signature images. The data collection took signature samples in a single session, and the images were scanned using an HP Scanjet 8250 at a resolution of 600dpi and a bit depth of 24. The images were stored in jpeg format. The signers donating to the dataset were from both Western and non-western first writing languages. Out of 150 donors, 15 had a non-western first writing language, and 135 a Western first writing language. The first writing languages of the subjects in the datasets were: English (122 participants), Russian (6 participants), Chinese, German, French and Greek (2 participants each), and Kurdish, Italian, Polish, Basque, Spanish, Thai, Hebrew, Arabic, Hindi, Albanian, Dutch, Welsh, Portuguese and Romanian (1 participant each). As an initial investigation, the physical height and width of all 600 (4 signatures x 150 signers) unconstrained signatures was measured by assessing the ink pixel extents in the x and y axes and converting into mm. In this way it was possible to analyse the average and range of signatures across a representative population. Furthermore, by separating the 150 signers according to whether their first writing language was Western and non-western it is possible to investigate the broad effects on ethnicity, which can be important meta-information within signature assessment. 2.2 Stage 2 Form Signing Areas Having established typical ranges of unconstrained signature sizes, the second stage of experimentation assessed a variety of public service application forms to note the box/area sizes provided for signature donation. Whilst it is acknowledged that the
The Use of Static Biometric Signature Data from Public Service Forms 75 signatures donated in these areas are purely for legal completion of the application, in the context of this work by reviewing the sizes for signature donation it will be possible to ascertain whether boxes are constraining typical signature sizes (through a comparison with the result from Stage 1). In this experiment a total of 98 signature donation boxes/spaces from 56 public service application forms were measured. These forms include UK, EU and US application forms for services such as entry visa application, naturalisation, passport and driving license applications. The forms were chosen as they are in the public domain and are for a range of secure applications that could benefit from the use of biometric signature recognition. Signing box areas were obtained by physically measuring the areas on each form. If no physical bounding box was present, the extents were defined by text or other bounding objects on the form. 2.3 Stage 3 Constrained Static Signature Verification To assess the effects of form constraints on static signature verification, signers donating to the dataset outlined in Stage 1 (Section 2.1) also signed eight sheets (with a single signature per sheet) with a signature donation box with dimensions 80mm by 30mm. This box size, when compared with the average box sizes on public service forms, is rather large, suggesting that any effects of constraining the signature would be amplified when using some of the forms measured in Stage 2 (Section 2.2). The constrained signatures were gathered using an identical method to the unconstrained signatures and were scanned using the same device and resolution. They were also taken in the same session as the unconstrained images and then stored in jpeg format. The constrained signatures were measured in the x and y direction, as the unconstrained signatures had been, to assess how much constraining the signature affected the size of the static signature sample. The results of this can be found later in the paper. 2.4 Stage 4 Four Way Matching Experiments Having both constrained and unconstrained signatures from 150 subjects enabled all four possible real-life combinations of static assessment: Scenario 1: Unconstrained enrolment vs. unconstrained verification Scenario 2: Unconstrained enrolment vs. constrained verification Scenario 3: Constrained enrolment vs. unconstrained verification Scenario 4: Constrained enrolment vs. constrained verification Four separate experiments were conducted representing the above scenarios. The experiments followed the same methodology: Four signatures were used to form an enrolment template against which a series of four genuine signatures per scenario were verified. This was followed by four false signatures verified against the enrolment template. The false signatures were unskilled forgeries taken from other users signatures so as to maintain the scenarios of constrained or unconstrained images. The verification algorithm used outputted a distance metric, which was used to create a ROC curve and identify the error rates for each scenario. The static ASV
76 E. Johnson and R. Guest system used for these experiments was an algorithm which calculates the signature bounding envelope and interior strokes using a range of geometric polar and Cartesian features as inputs to a Euclidean distance measure [7]. This method was used to assess the signatures because of the proven high performing nature of the algorithm. For these experiments it was treated as a black-box in that the algorithm was not optimised for performance. 3 Results 3.1 Stage 1 Unconstrained Signature Size Analysis In the first set of experiments the sizes of four constrained signatures from 150 subjects were determined. Table 1 and Figure 1 show the statistics for unconstrained images, divided into all signatures, those submitted by people whose native writing language was of a western style, and those whose native writing language was of a non-western style. Table 1. Mean, minimum and maximum values for unconstrained signatures Writing Language Axis Mean (mm) Min (mm) Max (mm) All x 45.47 11.47 104.70 y 14.84 5.03 34.45 Western x 45.80 14.64 104.70 y 14.37 5.03 33.82 Non-Western x 42.48 11.47 78.78 y 18.97 5.88 34.45 Signature size in y axis (mm) 60 50 40 30 20 10 0 Non Western Western 0 50 100 150 Signature size in x axis (mm) Fig. 1. x and y size values of unconstrained signatures
The Use of Static Biometric Signature Data from Public Service Forms 77 As can be seen, there is considerable variation between signature sizes, which may make optimum form design difficult. However, if the most important issue is that the signatures do not get deformed when scaled down, an average ratio between the sizes of the signature could be useful so as to design a box of the appropriate proportions. In assessing the signatures of Western participants in the dataset the ratio would be approximately 1:3 (height to width across all signatures). If this ratio is adhered to, deformation of constrained signatures could possibility be minimized. 3.2 Stage 2 Form Signing Areas With reference to above unconstrained signature sizes, the box/signing area sizes found on 98 signature donation boxes/spaces from 56 public standard public service forms were analysed to see how they compared. The statistics for the forms examined are shown in Table 2 and Figure 2. Table 2. Signing area sizes found on standard public service forms Axis Mean (mm) Min (mm) Max (mm) x 80.59 23.00 158.00 y 8.33 3.00 18.00 Box size in y axis (mm) 60 50 40 30 20 10 0 0 50 100 150 Box size in x axis (mm) Box Sizes Fig. 2. x and y size values of constraining boxes on public service forms As can be seen, the smallest size of a box in the y direction was a mere 3mm, which is smaller than even the minimum constrained signature height. The boxes exhibited a range of proportionality, some were very long and narrow, (Figure 3, IAP-66), while others were short and wide (Figure 4, I-17). More examples of boxes found on standard forms are shown in Figures 5, 6 and 7, which all come from the same form, (N644) showing that, not only are these forms not standardised across an agency or country, but in some cases, they are not even standardised on the same form, leading to an inability for these boxes to reliably produce proportional signature data if used as multiple static signature enrolment data within the same enrolment session.
78 E. Johnson and R. Guest Fig. 3. Signature box from US form IAP-66 Exchange Visitor Visa Application Fig. 4. Signature box from US form I-17- Petition for Approval of School for Attendance by Non-Immigrant Student Fig. 5. Signature box from US form N644 Application for Posthumous Citizenship Fig. 6. Signature box from US form N644 Application for Posthumous Citizenship Fig. 7. Signature box from US form N644 Application for Posthumous Citizenship 3.3 Stage 3 Constrained Static Signature Verification Having established that constraining boxes are often smaller or differently proportioned to an unconstrained signature, it was necessary to identify if constraining the signature
The Use of Static Biometric Signature Data from Public Service Forms 79 significantly distorted the signature to the point of affecting the accuracy of automatic signature verification. The size statistics of signatures in constraining boxes are shown in Table 3 and Figure 8, divided into sub-categories of all signatures, those submitted by people whose native writing language was of a Western style, and those whose native writing language was of a non-western style. As can be seen from these data, the unconstrained signatures were generally of a larger size than the constrained. An ANOVA analysis determined that in the x direction, the differences were not statistically different, but in the y direction, there was a significant variation. This shows that not only does the y direction vary in size depending on whether the signature is constrained or unconstrained, but that it varies more than in the x direction, meaning that the signature itself is being deformed to different proportions when restricted to a box. This data is from signatures constrained in boxes that were fairly large (80mm x 30mm) compared to the majority of the boxes found in standard forms that are publically used, so this effect would be even higher for many of the boxes currently used. Table 3. Mean, minimum and maximum values for constrained signatures Nationality Axis Mean (mm) Min (mm) Max (mm) All x 44.5 10.77 133.86 y 13.96 4.9 52.76 Western x 44.62 14.52 89.26 y 13.61 4.9 46.87 Non-Western x 43.33 10.77 133.86 y 17.08 5.31 52.76 Signature size in y axis (mm) 60 50 40 30 20 10 0 Non Western Western 0 50 100 150 Signature size in x axis (mm) Fig. 8. x and y size values of constrained signatures
80 E. Johnson and R. Guest It should also be noted that there are some outlying values within this data on examining these outlying images it was discovered that some participants had attempted to fill the box with their signature, introducing a further donation method of distorting the signature. As constraining a signature will deform the static image, a performance assessment was undertaken to examine whether this distortion would affect the accuracy of a static signature verification system. Four-way matching was applied examining the results of unconstrained enrolment images with unconstrained verification images (Scenario 1), unconstrained enrolment images with constrained verification images (Scenario 2), constrained enrolment images with unconstrained verification images (Scenario 3) and constrained enrolment images with constrained verification images (Scenario 4). The results of these experiments can be found in Table 4 and Figure 9. Table 4. Performance Evaluation of Constrained vs. Unconstrained Static Signature Verification Using a Four-Way Match Error Rate Scenario 1 Scenario 2 Scenario 3 EER 5.9 7.9 12.5 6.9 FAR 5.9 5.3 11.8 8.6 FRR 5.9 10.5 13.2 5.3 Scenario 4 1 True Positive Rate 0.8 0.6 0.4 0.2 0 Scenario 1 Scenario 2 Scenario 3 Scenario 4 0 0.2 0.4 0.6 0.8 1 False Positive Rate Fig. 9. ROC Curves to Show Performance Evaluation of Constrained vs. Unconstrained Static Signature Verification Using a Four-Way Match From these results, we can see that the constraining of a signature had a fairly minor effect on system accuracy, when both the enrolment and verification images are both constrained to the same level. However, as soon as one image was constrained
The Use of Static Biometric Signature Data from Public Service Forms 81 and the other unconstrained, particularly if enrolment was constrained but verification unconstrained, as in Scenario 3, the accuracy was severely affected. 4 Conclusions In this study it has been determined that when people are constrained as to the amount of space they have to sign, their signature will deform, particularly in the y axis. This occurs even when the constraining box is larger than the person s unconstrained signature. It is also possible to ascertain that when signature images are constrained, static verification system accuracy is adversely affected. While, in our experimental system both enrolment and verification images are constrained in the same size box, the equal error rate is 6.9%, the accuracy of the system is improved to an equal error rate of 5.9% when neither image is constrained. When either the enrolment or verification samples are constrained, and the other samples unconstrained, accuracy ratings drop considerably (especially when the enrolment template is constrained and the comparison unconstrained, where the error rate rises to 12.5%). The constrained sample results in this work were obtained using a constraining box of the size 80mm x 30mm. The size of this box in the x axis approximates to the mean of the boxes found on public forms in common usage, however, the size of the box in the y axis is significantly larger than even the largest box found on a public form. This form constraint is likely to cause even more reduction in accuracy for a biometric system as the y axis was the dimension varying the most between constrained and unconstrained signatures. While constraining signatures on boxes will allow for more space on a form, and, if the signature is being stored in a raw image format will lead to a smaller template size, careful thought should be given to the use of constraints on any signature that is to be used for biometric authentication as the accuracy of the system will be adversely affected by constraining the user s signature. In terms of the practical application of signature systems utilising static signature data to supplement primary authentication data (biometric or otherwise), this work has identified a number of key findings: Firstly, we have identified average signature size (and ranges) from a large population and surveyed how these compare with signing areas on a cross-section of common public service application forms. By noting the relative performance of constrained and unconstrained signatures we can identify that accuracy is maintained if constrained signatures are used for both enrolment and verification. Using form constrained signatures for enrolment and unconstrained for verification (Scenario 3 a typical implementation using forms at enrol time) produces sub-optimal results. Importantly the results show the possibility of form-based signatures either as a primary or supplementary biometric modality. References 1. Impedovo, D., Pirlo, G.: Automatic Signature Verification: The State of the Art. IEEE Trans. SMC-C 38(5), 609 635 (2008) 2. Plamondon, R., Lorette, G.: Automatic signature verification and writer identification - the state of the art. Pattern Recognition 22(2), 107 131 (1989)
82 E. Johnson and R. Guest 3. Leclerc, F., Plamondon, R.: Automatic signature verification: the state of the art - 1989-1993. Intl J. Pattern Recognition and Artificial Intelligence 8(3), 643 660 (1994) 4. Jain, A.K., Ross, A., Prabhakar, S.: An Introduction to Biometric Recognition. IEEE Trans. Circuits and Systems for Video Technology, Special Issue on Image- and Video-Based Biometrics 14(1), 4 20 (2004) 5. Kwon, T., Moon, H.: Multi-modal Biometrics with PKI Technologies for Border Control Applications. In: Kantor, P., Muresan, G., Roberts, F., Zeng, D.D., Wang, F.-Y., Chen, H., Merkle, R.C. (eds.) ISI 2005. LNCS, vol. 3495, pp. 99 114. Springer, Heidelberg (2005) 6. Vielhauer, C., Scheidat, T.: Multimodal Biometrics for Voice and Handwriting. In: Dittmann, J., Katzenbeisser, S., Uhl, A. (eds.) CMS 2005. LNCS, vol. 3677, pp. 191 199. Springer, Heidelberg (2005) 7. Ferrer, M.A., Alonso, J.B., Travieso, C.M.: Offline Geometric Parameters for Automatic Signature Verification Using Fixed-Point Arithmetic. IEEE Trans. PAMI 27(6), 993 997 (2005)