An Online Signature Verification System for ... - Semantic Scholar

An Online Signature Verification System for Forgery and Disguise Detection Abdelˆ aali Hassa¨ıne and Somaya Al-Maadeed Computer Science and Engineering Department College of Engineering, Qatar University Doha, Qatar {hassaine,s_alali}@qu.edu.qa

Abstract. Online signatures are acquired using a digital tablet which provides all the trajectory of the signature as well as the variation in pressure with respect to time. Therefore, online signature verification achieves higher recognition rates than offline signature verification. Nowadays, forensic document examiners distinguish between forgeries, in which an impostor tries to imitate a given signature of another person and disguised signatures, in which the authentic author deliberately tries to hide his/her identity with the purpose of denial at a later stage. The disguised signatures play an important role in real forensic cases but are not considered in recent literature. In this paper, we propose a new system for online signature verification for both forgeries and disguised signatures. This system extract features from both the questioned and the reference signature. The combination of the features is performed using several classifiers and achieves high performances on several signature databases. Keywords: Online signature verification; Online signature database; Feature extraction; Forgeries; Disguised signatures.

1

Introduction

Signature verification is a very active research field. It consists in comparing a questioned signature with a set of one or several reference signatures. Signature verification systems can assist forensic experts in deliberating about the authenticity of a questioned signature. Such systems can also be used in banks as security check. Two kinds of modalities are considered when dealing with the verification of signatures, the offline modality, in which scanned copies of the signatures are available for the verification, and the online modality, in which the signatures are acquired through digital tablets. The online modality provides more information about the signatures (trajectory, speed, pressure...ect), therefore, this modality provides generally better verification results than the offline modality. Nowadays, forensic document examiners distinguish between forgeries, in which an impostor tries to imitate a given signature of another person and

2

An Online Signature Verification System for Forgery and Disguise Detection

disguised signatures, in which the authentic author deliberately tries to modify his/her signature with the purpose of denial at a later stage. A first survey of existing methodologies in signature verification has been done in [13]. A second survey has been done some years later [8]. A recent survey has also been conducted some years ago [7]. Out of these surveys, we can say that signature verification systems range into two categories of methods: 1)Global methods, in which features are extracted from both the questioned and the reference signatures and then compared. According to the difference between these features a decision is taken on the authenticity of a certain signature. 2)Local methods, in which features are also extracted but they are considered locally, in order to allow a localized comparison. Local methods generally achieve better results than the global methods. Note that most of the work which has been done in the field of signature verification deal with the detection of forgeries. A recent study by Malik et al. suggests that local features achieve better performances in detecting disguised signatures than global features [11]. This study however only considered the offline modality. Furthermore, several international competitions on signature verification have been organized. A first one in 2004 dealing only with online signatures [14]. Since 2009, the international competition on signature verification has been organized on a yearly basis [1, 10, 9]. Note that only the 2010 edition competition dealt with disguised signatures [10], however this competition only considered offline signatures. In this paper, we describe a new system for online signature verification which is able to detect forgeries but also disguised signatures. We study the performance of several online signature verification features and we propose several classifiers for combining those features. The reminder of this paper is organized as follows: section 2 gives the details of the acquisition of online signatures. Section 3 presents details of the method and the proposed features. Section 4 presents results of individual features as well as their combinations. Section 5 concludes this paper and draws some future work perspectives.

2

Data acquisition

Online signatures contain a set of samples, each sample corresponds to the point coordinates on the digitizing tablet along with the corresponding pressure (Xt , Yt , Pt ) where t correspond to time (cf. figure 1). In this study, we used online signatures acquired using a WACOM Intuos4 digitizing tablet with a sampling rate of 200 Hz, a resolution of 2000 lines/cm and a precision of 0.25 mm. The pressure information is available in 1024 levels. 194 volunteers participated in the data collection process. They were instructed to provide occurrences of their natural signatures, and then to change their signatures in order to deny their identity at a later stage. Other volunteers were then asked to produce a simulation of the genuine signatures that they


3

(a)

X

y

Time

(b)

Pressure

Time

(c)

Time

(d)

Fig. 1. Example of a signature and the corresponding X, Y and Pressure signals.

could see. Figure 2 shows a screenshot of the software which has been developed for the acquisition of online signatures.

Fig. 2. Acquiring genuine, forgeries and disguised signatures.

3

Method description

Similarly to any automated verification process, signature verification involves two main steps: 1) Feature extraction: In this step, features are extracted from the questioned signature as well as the reference signature. A feature is said to be discriminative if its intra-writer variability is smaller than its inter-writer variability, 2) Matching: In this step, a decision or a score indicating how probable it is that the questioned signature is genuine. This score is generally obtained by a combination of several features. Each of these steps is detailed below.

4


3.1

Feature extraction

From the three basic signals which are X, Y and Pressure, several features (or signals) can be extracted from the online signatures [12]. The following list gives a description of the features used in this study. – Distances: The euclidian distance is pcomputed between each successive X and Y coordinates of signature dt = (xt − xt−1 )2 + (yt − yt−1 )2 (cf. figure 3(a)). – Angles: The angle between the X axis and the line formed with the first 0 signature point and the current point αt = atan XYtt −Y −X0 (cf. figure 3(b)). – Speeds: The difference between successive distances St = dt − dt−1 . – Angular speeds: The difference between successive angles ASt = αt −αt−1 .

Y3

Y3

Y2

Y2

dist3 dist2

Y1

Y1 dist1

Y0

α1

Y0

X0

X1

X2

X3

(a)

X0

α3

α2

X1

X2

X3

(b)

Fig. 3. Computing distances and angles.

3.2

Matching

In order to compare the questioned signature with the reference signature, we compare the differences between each pair of features. For the sake of simplicity, this is illustrated for the distance feature in figure 4. Two differences are considered: the difference at the signal level ∆S and the difference at the histogram level ∆H . ∆S is computed after resizing (using a cubic interpolation) the two signals to the same number of samples which has been empirically set to 100 in this study. i=1 ∆S (d1 , d2 ) = Σ100 |d1 (i) − d2 (i)| .

∆H is computed after computing the distribution histograms of the two signals with the same number of bins (which has also been set empirically to 100).


Questioned signature

dQ

Reference signature

dR

ΔS(dR,dQ), ΔH(dR,dQ)

5

Fig. 4. Comparing the distance feature of the questioned and the reference signature.

i=1 ∆H (d1 , d2 ) = Σ100 |Histogramd1 (i) − Histogramd2 (i)| .

Note that distribution histograms are widely used in the field of writer identification [5]. Those formulas apply as well on the other features, resulting in the list of differences shown in table 1 for each comparison.

Table 1. List of differences for each comparison Difference

Description

Difference

Description

∆S (X)

Signal difference of X

∆H (X)

Histogram difference of X

∆S (Y )

Signal difference of Y

∆H (Y )

Histogram difference of Y

∆S (P )

Signal difference of pressures

∆H (P )

Histogram difference of pressures

∆S (d)

Signal difference of distances

∆H (d)

Histogram difference of distances

∆S (α)

Signal difference of angles

∆H (α)

Histogram difference of angles

Signal difference of speeds

∆H (S)

Histogram difference of speeds

∆S (S) ∆S (AS)

Signal difference of angular speeds ∆H (AS) Histogram difference of angular speeds

All the above differences have a different discriminative power. In the next section, we present the performance of each of them separately and propose several ways of combining them.

4

Results and discussion

Two online signature databases have been considered in this study: ICDAR2009 signature verification competition dataset: This dataset [1] contains two separate sets, one for training and one for evaluation. The training set contains signatures of 12 authentic persons, each of them provided 5 signatures. Moreover, 32 persons provided 5 forgeries of the genuine signatures. In sum, each signature has 5 genuine occurences and 620 forgeries. This set has also been used for training purposes.

6


The evaluation set contains signatures of 77 persons, each of them provided 12 genuine signatures. This set contains a total of 640 forgeries. QU online signature database: This dataset was acquired as described in 2. It contains signatures of 194 persons, each of them provided at least 3 reference signatures and 3 disguised signatures. For 50 persons, at least 3 skilled forgeries have also been provided. The signatures of 50 random persons of this dataset have been used for training. Evaluation on this dataset has been performed by considering the full testing set in a first step, by disregarding disguised signatures in a second step and by disregarding forgeries in a third step. Table 2 shows the Area Under the ROC Curve (AUC) and the Equal Error Rate (EER) of the presented features on these databases. Table 2. AUC and EER measures for the presented features ICDAR2009 Feature AUC

EER

QU all AUC

EER

QU forgeries

QU disguised

AUC

AUC

EER

EER

∆H (X) 0.8365 0.2546 0.6561 0.3959 0.8730 0.2088 0.6299 0.4097 ∆S (X) 0.8266 0.2152 0.7148 0.3192 0.8155 0.2598 0.7027 0.3305 ∆H (Y ) 0.8347 0.2551 0.5832 0.4337 0.7490 0.3295 0.5631 0.4504 ∆S (Y ) 0.6776 0.3917 0.6690 0.3909 0.6885 0.3790 0.6666 0.3924 ∆H (P ) 0.9705 0.0866 0.7315 0.3242 0.9526 0.1188 0.7048 0.3480 ∆S (P ) 0.8499 0.1969 0.7690 0.2920 0.9860 0.0471 0.7428 0.3094 ∆H (d)

0.8464 0.2345 0.5959 0.4182 0.7232 0.3316 0.5805 0.4271

∆S (d)

0.6687 0.3957 0.5406 0.4782 0.5495 0.4737 0.5395 0.4787

∆H (α) 0.6449 0.3851 0.5641 0.4525 0.6008 0.4343 0.5596 0.4562 ∆S (α)

0.6478 0.4069 0.5458 0.4752 0.5643 0.4650 0.5435 0.4765

∆H (S) 0.6800 0.3519 0.6435 0.4102 0.6083 0.4409 0.6477 0.4063 ∆S (S)

0.6794 0.3627 0.5351 0.4777 0.5570 0.4642 0.5324 0.4794

∆H (AS) 0.5527 0.4842 0.6023 0.4388 0.6818 0.3759 0.5927 0.4462 ∆S (AS) 0.5857 0.4453 0.5115 0.4936 0.5407 0.4776 0.5079 0.4956

Several classifiers have been tried in order to combine the presented features, including Random Forest [2] with 2000, 5000 and 10000 random trees, logistic regression [6], linear regression, Multivariate Adaptive Regression Splines (MARS) [4] and neural networks with 2, 5 and 10 hidden neurons using a logistic output unit [3]. Table 3 shows the performance of those classifiers on the presented features. Notice that all the classifiers achieve comparable results. However, neural networks and random forests slightly outperform the other classifiers, the latters are specifically prefered in the case of disguised signatures. Finally, although the


7

Table 3. AUC and EER measures of combining features using several models ICDAR2009

QU all

QU forgeries

QU disguised

Random Forest (2000 tree) 0.9685 0.0924 0.8602 0.2243 0.9847 0.0459 0.8451 0.2359 Random Forest (5000 tree) 0.9688 0.0932 0.8604 0.2233 0.9845 0.0454 0.8454 0.2359 Random Forest (10000 tree) 0.9699 0.0905 0.8604 0.2250 0.9849 0.0449 0.8453 0.2368 Logistic Regression

0.9523 0.1149 0.8375 0.2334 0.9728 0.0570 0.8211 0.2490

Linear Regression

0.9470 0.1108 0.8424 0.2323 0.9743 0.0576 0.8265 0.2461

MARS

0.9472 0.1153 0.8590 0.2261 0.9829 0.0523 0.8440 0.2368

Neural Nets (2 nodes)

0.9650 0.0784 0.8486 0.2359 0.9761 0.0505 0.8332 0.2460


0.9588 0.0912 0.8525 0.2301 0.9791 0.0507 0.8372 0.2425


0.9733 0.0782 0.8499 0.2355 0.9739 0.0544 0.8349 0.2482

performance in detecting disguised signatures is acceptable, the performance in detecting forgeries is much more accurate which suggests that further research is needed for detecting disguised signatures.

5

Conclusion

We presented a new system for online signature verification which deals with both forgeries and disguised signatures. This system extracts several features of the signatures and compares them at the histogram level and the signal level. Several classifiers have been tested for combining these features, with neural networks and random forests generally prefered. It is planned to study new methods of matching these features using dynamic time warping as well as extending this method for the case of multiple reference signatures. Acknowledgments. This publication was made possible by a grant from the Qatar National Research Fund under its Undergraduate Research Experience Program and Qatar University under Student fund QUST-CENG-FALL-11/1210. Its contents are solely the responsibility of the authors and do not necessarily represent the official views of the Qatar National Research Fund or Qatar University.

References 1. Blankers, V.L., Heuvel, C.E.v.d., Franke, K.Y., Vuurpijl, L.G.: Icdar 2009 signature verification competition. In: Proceedings of the 2009 10th International Conference on Document Analysis and Recognition. pp. 1403–1407. ICDAR ’09, IEEE Computer Society, Washington, DC, USA (2009) 2. Breiman, L.: Random forests. Machine Learning 45, 5–32 (2001)

8


3. Brierley, P.: Tiberius, predictive modelling software. http://www.tiberius.biz (2011) 4. Friedman, J.H.: Multivariate adaptive regression splines. Annals of Statistics 19(1), 1–67 (1991) 5. Hassa¨ıne, A., Al-Maadeed, S., Alja’am, J., Jaoua, A., Bouridane, A.: The ICDAR2011 Arabic Writer Identification Contest. In: Document Analysis and Recognition, 2011. Proceedings. Eleventh International Conference on. Beijing, China (September 2011) 6. Hosmer, D.W., Lemeshow, S.: Applied Logistic Regression, 2nd Edition. Wiley (October 2000) 7. Impedovo, D., Pirlo, G.: Automatic Signature Verification: The State of the Art. Systems, Man, and Cybernetics, Part C: Applications and Reviews, IEEE Transactions on 38(5), 609 –635 (sept 2008) 8. Leclerc, S., Plamondon, R.: Automatic signature verification: the state of the art1989-1993. Intl. Journal of Pattern Recognition and Artificial Intelligence 8(3), 643–660 (1994) 9. Liwicki, M., Malik, M., van den Heuvel, C., Chen, X., Berger, C., Stoel, R., Blumenstein, M., Found, B.: Signature Verification Competition for Online and Offline Skilled Forgeries (SigComp2011). In: Document Analysis and Recognition (ICDAR), 2011 International Conference on. pp. 1480 –1484 (sept 2011) 10. Liwicki, M., Heuvel, C.E.v.d., Found, B., Malik, M.I.: Forensic signature verification competition 4nsigcomp2010 - detection of simulated and disguised signatures. In: Proceedings of the 2010 12th International Conference on Frontiers in Handwriting Recognition. pp. 715–720. ICFHR ’10, IEEE Computer Society, Washington, DC, USA (2010) 11. Malik, M.I., Liwicki, M., Dengel, A.: Evaluation of Local and Global Features for Offline Signature Verification, pp. 26–30. CEUR (2011) 12. Nalwa, V.: Automatic on-line signature verification. kluwer International series in engineering and computer science pp. 143–164 (1997) 13. Plamondon, R., Lorette, G.: Automatic signature verification and writer identification the state of the art. Pattern Recognition 22(2), 107 – 131 (1989) 14. Yeung, D.Y., Chang, H., Xiong, Y., George, S., Kashi, R., Matsumoto, T., Rigoll, G.: Svc2004: First international signature verification competition. In: Proceedings of the International Conference on Biometric Authentication (ICBA), Hong Kong. pp. 16–22. Springer (2004)