Dynamic Features Extraction For Online Signature Verification

Dynamics features Extraction for on-Line Signature verification Flor Ramírez Rioja, Mariko Nakano Miyatake, Héctor Pérez M., Karina Toscano M. Escuela Superior de Ingeniería Mecánica y Eléctrica Unidad Culhuacan, IPN Sección de estudios y Postgrado e Investigación Av. Santa Ana No.1000, Col. San Francisco Culhuacan, 04430 México DF.TEL.57296000 Ext.: 73207/Fax: 56562058 e-mail: [email protected].

Abstract Actually great inters to develop a robust on-line signature verification system has been increased. In this field, three kinds of forgeries: random forgery, simple forgery and expert forgery must consider. In this paper a dynamic features extraction for an on-line signature verification system is presented. The dynamic features are extracted from authentic and forged signatures witch relatively low computational cost. In the proposed system, all kind of forgeries included expert forgeries are considered to detect as forged signatures. The computer simulation results show us a desirable performance of the proposed system. Index terms. Signature verification, dynamics verification, extraction of dynamics characteristics, forgery detection and off-line signature verification

1. Introduction The signals processing based in manuscript traces, like signatures has been used to the personal identification or verification to authenticate official documents. Such as checks, officials letter, etc. The signatures have been used in occidental countries since a few centuries ago, together with fingerprint patterns. Actually the signatures verification or recognition has been realized in manual way by authorized person in financial system and governmental office. While number of verification operations is increased, also the automatization is required. In the signature verification and recognition, basically there are two methods. Which are on line or dynamic method and off-line or static method, to realize the first method, the dynamic information of signature, such as writing speed and acceleration, graphical momentum, etc. is required. In the proposed system, we extracted the following dynamic features of signature. x x x x

Initial and final point signature Writing order. Pressure change in axis X and Y Speed change in axis X and Y

x x

Region with high pressure in the signature Region with high and low speed in axis X and Y.

To get these data, we used a digital table, in which we can get pen position and pen pressure in each sampling period. On the signature verification field, three kinds of forgeries random forgeries, simple forgeries and expert forgeries must be considered. The random forgeries means any another signature. In which the forged signature is produced without any knowledge about authentic signature. Obviously, this kind of forgeries is very easy to verify. The simple forged signature is generated by a forger, who has only knowledge about authentic signar name, this type of forgery can be the same of random one, when the signature is very personalized and it doesn’t represent signer name. In Europe also Mexico, generally signatures are very personalized [4], however in the USA or Canada, in many cases the signature represents signer name [5], In this case the simples forged signature can be very similar to the authentic one, Also signatures of countries like China, Korea, Japan, some signatures are equal to the handwritten signer names with by Chinese characters [1]. Finally in expert forgery, forger uses the authentic signatures to practice many times and also trace the authentic signature. This kind of forgeries occurs in use of traveler checks [5]. The expert forged signature detection is a very hard task. Though more than 80 % forgeries are random and simple forgeries in the financial system, the expert forgeries must be considered in the robust signature verification system. The signature verification systems considered expert forgeries in the literature; the verification performance is approximately 70 % or 75 %. This percent is not sufficient for a practice application. 2. DEVELOPMENT.

Proceedings of the 14th International Conference on Electronics, Communications and Computers (CONIELECOMP’04) 0-7695-2074-X/04 $ 20.00 © 2004 IEEE

The on-line signature verification system including dynamics features extraction task consists of the following modules. x Data acquisition module. x Preprocessing x Dynamic feature extraction x verification.

In the proposed system the following preprocessing are used. x x

Noise elimination Redundant data elimination

The figure 3 and 4 show the input signal in axis X and Y and these frequency responses, respectively.

Figure 3 the input signal and its frequency response in X axis.

Figure 1 The main structure of proposed system.

2.1 Data acquisition. Figure 4 the input signal and its frequency response in Y axis.

To acquire data in the dynamic verification system, we used a digital tablet, which capture dynamic information of signature such as initial and final point, writing order, writing speed and pressure of pen by each sampling period. In this system, we use a WACOM digital tablet with size 9 x 12 inches and resolution 2540 ppi. The figure 2 shows the digital tablet and the ergonomic pen used in the proposed system. [6]

Figure 2. Digital tablet and ergonomic pen used in the proposed system.

Basically, the tablet captures three data in each sampling period, which are pen position in axis X and Y and pen pressure.

2.3 Noise elimination. The data captured using digital tablet include noise caused by little vibration and movement by the hand. This noise produces many short traces, which is interpreted as signature traces and caused erroneous verification result. Generally this movement or little vibration of the hand is a high frequency signal, and then we can use a low-pass filter to eliminate this signal. Observing the figure 3 and 4, we determined the cut frequency of low-pass filter is 20Hz. In the proposed system, we used a infinite impulse response (IIR) filter, due to it’s better performance with less filter order compared with FIR filter. The figure 5 shows the frequency response of an IIR Butterworth filter with a cut frequency 20Hz and order 5.

2.2 Preprocessing. Generally the preprocessing will be realized to eliminate noise introduced in data acquisition process, reduce redundant samples and normalization in size, position, etc.

Proceedings of the 14th International Conference on Electronics, Communications and Computers (CONIELECOMP’04) 0-7695-2074-X/04 $ 20.00 © 2004 IEEE

Figure 5 Frequency response of a IIR Butterworth filter.

2.4 Redundant data elimination. In the tablet, we can change the sampling frequency in order to capture more quantity of data. Really we captured data with maximum sampling frequency. However in some traces, we can observe some redundant samples. These redundant samples can cause high complexity in features extraction module and signature verification module. The figure 6 shows a re-sampling example, with (a) original data, (b) is a uniform re-sampling result and (c) is a correct re-sampling result.

(a) (b) (c) La Fig.7ª (a) Original signature Firma original, (b) Signature with important and not important points selected , (c) Signature reconstructed using only important points.

2.5 Feature extraction. The dynamic features extracted from the signature are followings. a) b) c) d) e) f)

Initial and final points Writing order Writing speed in X and Y axis Pen pressure High pressure region High and low speed region in X and Y axis

3. RESULTS. The dynamic features extraction has been realized by computer simulations, these results are shown as follows. 3.1. Initial and final points.

Figure 6 (a) Original data, (b) is a uniform re-sampling result, (c) is a correct re-sampling result.

To conserve the signature shape, we have to realize a correct re-sampling, in which only redundant sampling points are eliminated and important points must be conserved. To conserve the important points we assume the following conditions.

In Mexico generally very personalized signatures are used, in which to find initial and final points is not so easy. Therefore in many cases initial, final points or both points of forged signature are different from these of the authentic one. The figure 8 shows an authentic signature with its initial and final points and figure 9 shows forged one with its respective initial and final points. We can observe from the figures, the initial and final points are different.

1. - Initial and final points are important points. 2. - If a distance between a point and a generated line by two neighbor points (captured before and after sampling period), is greater than a threshold value defined previously, the point is considered an important point. Using above condition, we can eliminate not important points. The figure 7 shows a correct re-sampling result, where the blue points represent points not important and red points represent important points.[2][3]

Figure 8 an original signature with initial and final points, Pi. Indicate the initial point and Pf indicate the final point.

Figure 9 a forgeries signature with initial and final points, Pi. Indicate the initial point and Pf indicate the final point

3.2. writing order. When a signature has continuous traces (the signer doesn’t pen up during sign), probably forger can know the writing order, once he knew the initial and final points of the signature. However for signatures with separated traces, it is not easy to know the writing order correctly. The figures

Proceedings of the 14th International Conference on Electronics, Communications and Computers (CONIELECOMP’04) 0-7695-2074-X/04 $ 20.00 © 2004 IEEE

10 and 11 represent an authentic signature and a forged one. We can observe that writing order of the authentic and forged signature is different.

(a)

Figure 10 Order of traces of an original signature.

(b)

Figure 11 Order of traces of a forgeries signature

3.3 Writing speed in X and Y axis. Generally the signer sign his signature with varying writing speed, for example, some person begin with slow speed, after increment his writing speed and. Finally lower again. This speed variation during sign could be a good factor to discriminate authentic signature from forged one. The speed variation in X and Y axis are shown by figures 12 and 13, in which speed variation of two authentic signatures and that of the authentic and forged signatures are compared. From these figures, we can observe that the speed variation of two authentic signatures is very similar; however this variation of the authentic and forged signature is very different. Original signature 1

Figure 13 Speed variation in X and Y axis of authentic signature and forged one.

In the figure 13,(a) shows speed variation in X axis and (b) shows speed variation in Y axis. 3.4 Pen pressure. The pen pressure is a useful feature to verify a forged signature. The figure 14 shows pressure changes during sign of two authentic signatures. The figure 15 shows a pressure variation compared between an authentic and forged signature. Original signature 1

Original signature 2

Original signature 2.

(a)

(b)

Figure 14 Pen pressure variation of two authentic signatures.

In the figure 14, blue line shows the pen pressure variation of authentic signature Original Signature

forgeries signature

Fig.12 Speed variation in X and Y axis of two authentic signatures.

In the figure 12 (a) Show speed variation in X axis and (b) shows speed variation in Y axis. Original signature

forgeries signature

Proceedings of the 14th International Conference on Electronics, Communications and Computers (CONIELECOMP’04) 0-7695-2074-X/04 $ 20.00 © 2004 IEEE

Figure 15 Pen pressure variation compared with an authentic signature and forged one.

In the figure 15, Pen pressure variation compared with an authentic signature and forged one. 3.5 High pressure region.

the authentic signatures are same, however these regions are different between the authentic and forged signature. 3.6 High and low velocity region in X and Y axis. The figure 18 shows high and low speed regions Original signature 1

The high pressure region is also a useful feature to verify a forged signature. The figure 16 shows high pressure region of two authentic signatures, thick trace represents high pressure region. The figure 17 shows the comparison of high pressure region between an authentic signature and forged one.

original signature 2

Low and high speed region for each original signature. Signature 1

Original signature 1

Signature 2

original signature 2

(a) (b) Figure 18 Measurement of low and high speed region of two original signatures. (a)

(b)

Figure 16 High pressure regions of two authentic signatures. Thick trace represents high pressure region.

The figure 18(a), show to original signature 1 with red points for its high speed region and green points for its low speed region. The figure 18(b) show to original signature 2 with red points for its high speed region and green points for its low speed region, both regions have the same region. Original signature 1

Original signature

forgeries signature

forgeries signature

Low and high speed region for an original and forgeries signature. Original signature -

forgeries signature

(a)

(b) Figure 19 Measurement of low and high speed region of an original and forgeries signature. Figure 17 High pressure region of an authentic signature and forged one. Thick trace represents high pressure region.

In the figure 16, (a) and (b) show high pressure region of the authentic signature 1 and 2, respectively, and in the figure 17, (a) and (b) shows high pressure region of an authentic signature and forged one. From figure 16 and 17, we can observe the high pressure regions of

The figure 19 (a) show to original signature with red points for its high speed region and green points for its low speed region. The figure 19(b) show to falsificated signature with red points for its high speed region and green points for its low speed region, both regions lie in different places.

Proceedings of the 14th International Conference on Electronics, Communications and Computers (CONIELECOMP’04) 0-7695-2074-X/04 $ 20.00 © 2004 IEEE

Using extracted features, we constructed on-line signature verification system. Firstly we captured 100 authentic signatures and 50 forged signatures by three Forgers per signer. Data acquisition is realized during one month to capture intra-person variation. All forged signatures are expert forgeries, three forgers have knowledge about the signature shape and they can practice until they can sign perfectly. Figure 21 Recognition

In the signature verification system, we use back propagation neural networks per each signer, shown by figure 20. The network parameters used in the system are given by table 1. The verification results obtained using ten neurons in hidden layer were shown by figure 21. We get type I error (negative false error) is 1.8% and type II error (positive false error) is 2 %. We use the back propagation network for detecting. The Figure 20 show a network and its parameters in table1

the percent graphic.

The future Works will be to put a base of date with 1400 original and professionals falsificated signatures on line in a system for yours recognition. 4. Conclusions The dynamic features extraction allows us to realize on-line signature verification in efficient manner. In the proposed system, we get type I error (negative false error) is 1.8% and type II error (positive false error) is 2 %. This error ratio is fairly well, compared with other systems proposed in literature. Each extracted feature is examined its contribution for verification performance. In many signers, the speed variation in X axis and pressure variation have more contribution to discriminate authentic signature from forged one. 5.Acknowledgement. This work is supported by the national council of science and Technology (CONACyT) in Mexico and the National Polytechnic Institute (IPN) of Mexico.

6.References.

Figure 20 Network structure for signatures recognition. Tip de Network: back propagation Signature number for recognition: 72 Neural number in the input layer (include BIAS): 251 Neural number in the hidden layer (include BIAS): 11 Neural number in the output layer: 2 Number Patters in train: 60 miu: 0.1 Epsilon: 0.0001

Table 1 Parameters of back propagation network.

The results obtained with ten neuronal in hide layer were: 98.2% original signatures recognized and 1.8% falsificated signatures not recognized.

[1]D.Sakomoto,H.Morita, T.Ohishi,Y.Komiya,T.Matsumoto, “On-line Signature Verification Algorithm Incorporating, Pen posiction,Pen pressure and Pen inclination trajectories”, Proc. of. 2001 IEEE international conf. Acoustics, Speed and signal processings, Vol. 2. Page. 993-996, 2001. [2]Taik H.Rhee,Sung J. Cho , Jin H. Kim “On-line Signature Verification Using Model-Guided Segmentation and Discriminative feature Selection for Skilled Forgeries”, Proc. Of. Sixth International conf. On Document Analysis and recognition, page. 645-649, 2001. [3]K.W.Yue and W.S.Wijesuma, “Improved Sementation and segment Association for On-line Signature Verification”, Proc. Of. 2000 IEEE International conf. On Systems Man and Cybernetics, Vol. 4, page. 2752-2756 , 2000. [4]R. Plamondon, Sargur N. Srihari, “On-line and Off-line Handwriting Recognition A comprehensive Survey”, IEEE transaction on patter analysis and machine intelligence, Vol. 22,No.1, page. 63-78, January 2000 [5]Charles C. Tappert, Ching Y. Suen , and T. Wakahara, “The State of the art in On-line Handwriting Recognition”, IEEE Transactions on pattern analysis and mavhine intelligence, Vol. 12, No. 8, page 787-804, August 1990 [6] Tablet use manual Wacom Mod. XD0912

Proceedings of the 14th International Conference on Electronics, Communications and Computers (CONIELECOMP’04) 0-7695-2074-X/04 $ 20.00 © 2004 IEEE