Online handwritten signature verification system: A Review

Transcription

1 Online handwritten signature verification system: A Review Abstract: Online handwritten signature verification system is one of the most reliable, fast and cost effective tool for user authentication. This work examines the online handwritten signature verification system methodologies. Signatures are acquired from devices such as pressure sensitive tablets, digitizer, etc. The aim of this paper is to review the signature feature extraction algorithms, techniques and methodologies. This gives a brief description of the performance evaluation parameters. The performance of algorithms is compared using various factors which include the False Acceptance Rate (FAR), False Reject Rate (FRR) and Equal Error Rate (EER) etc. Keywords: Feature extraction, Segmentation, False Acceptance Rate, False Rejection Rate. 1. INTRODUCTION Many business applications (e.g. E-banking) depend on biometrics since using biometrics is the only way to pledge the presence of the owner when a transaction is made. The main benefit of using a biometrics can t easily be lost, stolen, hacked, and forged. The future of biometrics looks increasingly bright with the demand for security rising day by day. Application areas of online signature verification include protection of small personal devices (e.g. PDA, laptop, mobile, etc.); protection of data of computer users from unauthorized persons and authentication of individuals for access to documents, physical devices. The signatures of the same person can vary with time and state of mind. This factor is not considered in offline verification system, but in online verification system these factors are also considered to make the system more precise. Numbers of techniques have been proposed to authenticate forgeries. Some of them are discussed in this paper to provide the performance analysis in the field of online handwritten signature verification system. The paper is structured in six sections; Section 1 is the introductory part, Section 2 describes the challenges associated with online handwritten signature verification system. Section 3 presents the basic aspects of data acquisition and preprocessing stage. In section 4, the feature extraction stage is addressed. The Study of various existing methods of Online Signature verification systems brief description of papers of each category is given in Section 5.Finally the conclusion of the paper is reported in Section 6. Prathiba M.K. 1 and Dr. L. Basavaraj 2 1 Visvesvaraya technological University, ATME College of Engineering, Mysore , India. 2 Visvesvaraya technological University, ATME College of Engineering, Mysore , India. 2. CHALLENGES ASSOCIATED WITH ONLINE HANDWRITTEN SIGNATURE VERIFICATION SYSTEM One of the main challenges in signature verification is related to the signature inconsistency. Skilled forgers can perform forgery with high resemblance to the user s signature. The variability noticed in this case is known as inter-person variability. Significant differences have also been noticed with the signatures of the user taken at different times known as intra-person variability. It is a challenge to distinguish both and to get better result. The signing relies on a very fast, practiced, and repeatable motion. But, this motion may change over time, generating another completely different signature. The signatures signed with a digitize pen and unfamiliar signing surface may also affect the signing process. 3. DATA ACQUISITION AND PREPROCESSING Signature verification involves signature acquisition, preprocessing, feature extraction, enrollment, matching, and threshold selection. 3.1 Signature acquisition Online signature verification uses the signature which is captured by data acquisition devices like pressuresensitive tablet. The use of dynamic features makes the signature more unique and more difficult to reproduce. As a result the online signature verification is more reliable than offline signature verification. Commonly used data acquisition device is a digitizing tablet and digitize pen. 3.2 Preprocessing After data acquisition, preprocessing steps are commonly performed to improve the performance of a verification system. Typically, there are three stages in the preprocessing phase. These include smoothing, normalization, and segmentation Smoothing Low resolution tablet suffer from discretization errors, resulting in jagged signature trajectories. Extracting local features from jagged signature trajectories and then using them for verification lead to poor performance. Hence, smoothing is required for low resolution tablets Normalization In the systems where tablets of different active areas are used, signature size normalization is a frequently used as Volume 3, Issue 2 March April 2014 Page 263

2 preprocessing technique. Comparing two signatures having the same shape with different sizes would result in low similarity scores. Size normalization is applied to remove this effect, i.e. the characteristics of signature are preserved Segmentation Segmentation is an important preprocessing step, which influences all the successive phases of signature verification system. Signature segmentation is a complex task since different signatures produced by the same signer can differ from one another. Because of this, specific attention has been devoted to signature segmentation and several techniques have been proposed [11], [8]. 4. FEATURE EXTRACTION Performance of the online signature verification depend system the selection of features are very significant. The features are related to the shape of the signature and independent of the data acquisition device. The efficiency of a signature verification system mainly depends on Feature extraction stage. Feature extraction techniques should be fast and easy to compute so that system has low computational power. Selected features should discriminate between genuine and forgery signature. There are two types of features that validate a signature. Static features are those, which are extracted from signatures that are recorded as an image. Dynamic features are extracted from signatures that are acquired in real time which provides the information about the number and order of strokes, the overall speed of the signature, the pen pressure at each point etc. which make the signature more unique. Two main approaches in feature extraction process are parameters-based approach and functions-based approach. When parameter-based features are used, the signature is characterized as a vector of elements, and each one represents the value of a feature [10]. When function-based features are used, the signature is usually characterized in terms of a time function whose values constitute the feature set. In general, function-based features allow better performance than parameters-based features [2]. Parameter based features are generally classified into two main categories such as global and local. Global features are those which are extracted from the whole signature. They can be extracted easily, but it delivers only limited information about signature. Global features include signature size, signing time, number of pen ups, signature height to width ratio, etc. Local parameters concern features extracted from specific parts of the signature. Based on the extracted features, local parameters can be divided into component-oriented parameters, which are extracted at the level of each component (i.e., height to width ratio of the stroke, relative positions of the strokes, stroke orientation, etc.), and pixel-oriented parameters, which are extracted at pixel level (i.e., grid-based information, pixel density, gray-level intensity, texture, etc.). Generally, local features are computationally expensive but more accurate than the global features. Some parameters features, which are generally considered to be global features, can also be applied locally, and vice versa. 4.1 Performance evaluation parameters The performance of signature verification systems is typically described by two parameters; False Acceptance Rate and False Reject Rate. The percentage of genuine signatures rejected as forgery which is called False Rejection Rate and the percentage of the forgery signatures accepted as genuine is called False Acceptance Rate. These two errors are directly correlated, where a change in one of the rates will inversely affect the other. Generally signature verification system shall have an acceptable trade-off between a low FAR and a low FRR. A common alternative to describe the performance of system is to calculate the Equal Error Rate. EER corresponds to the point where the false accept and false reject rates are equal. 5. STUDY OF VARIOUS EXISTING METHODS OF ONLINE SIGNATURE VERIFICATION SYSTEMS Every Signature verification system is differing from one another, based on their feature selection method, decision making method or both. This section reviews some of the existing signature verification systems such as Tablet PC and digitalize pen, Hidden Markov Models (HMM s), modified Dynamic Time Wrapping Technique (DTW) and Neural Networks (NNs). 5.1 Tablet PC and digitize pen The online signature verification systems are using the Tablet PCs and its pen. Tablet PC is a stable and reliable signature capturing devices. It captures both the dynamic and static features of a signature at the same time. Commonly used data acquisition device is a digitizing tablet. However, the signing process using digitizing tablet is different because the signer instantaneously cannot see what he has written, which may cause inconvenience [3],[4].Digitize pens with touch-sensitive screens and digital-ink technologies are developed to avoid signer disorientation by providing immediate feedback to the writer. In a digitize pen several electronic components are used to detect pen motion, velocity, inclination, and other dynamic features which are extracted during the signing process. Alonso-Fernandez [3] developed a prototype of securing access and securing document application using Tablet PC system. Two different commercial Tablet PCs(Hewlett-Packard TC 1100with Intel Pentium Mobile 1.1 GHz processor and Toshiba Portege M200 with Intel Centrinto 1.6GHz processor)had been used. The signature verification systems based on the pressure statics of the signature. The experiments were tested against both random and skilled forgeries by using data base containing 3000 signatures. According to their reports F. Alonso-Fernandez et al. [4], the Toshiba tablet performed Volume 3, Issue 2 March April 2014 Page 264

3 better with the reporting result of 8.27% to 8.05% EER for Skilled forgeries and 3.20% to 2.76% EER for random forgeries. A. K. Jain [1] utilized a digitizing tablet to capture the spatial and dynamic information of signature. The digitizing tablet used was the IBM CrossPad from A.T. Cross Company. IBM CrossPad records the x-coordinates and y-coordinates of the points in the signature with a sampling rate of samples per second. Experiments were conducted using a database with over 1232 signatures collected from 102 individuals. The best result reported a false accept rate of 1.6% and a false reject rate of 2.8%. Thanin Maneechot[15] Proposed an algorithm for online signature verification, using N-tuple learning machine. The features are extracted from hand written signature on WACOM digital tablet. This captures dynamic information of signature such as X-Y position, pressure of pen and pen altitude angle. The verification employed N- tuple learning machine. The basic structure is similar to an encoder, which the inputs are binary values N-tuple simulated from RAM s. It was establish that pen pressure was influenced feature. They obtained a FAR for a forgery was 1.32% and FRR for genuine signature was 0.2%. Syed Khaleel Ahmed [14] designed signature verification system using the WACOM Graphire 3 pressure pad with pressure sensitive pen. The tablet captures the position co-ordinates x, y and pressure P as a function of time of the signature were extracted. A signature verification system based on the Self-Organizing Map (SOM) of Neural Network model was designed in MATLAB to verify the signature. The system performance was evaluated by using 50 epochs and 200 epochs for two different cases. The best result reported a FRR and FAR was 19.05% and 9.29% for 200 epochs. The comparison of selected performance evaluation parameters for the Tablet PC and digitalize pen methodology illustrated in Table-1with high accuracy was reported by Thanin Maneechot[15]. Table 1- Performance Evaluation Parameters obtained from the Tablet Pc and Digitize Pen Sl. Year FAR FRR Authors No. (%) (%) Anil. K. Jain al [2] 2 3 Thanin Maneechot[15] Syed Khaleel Ahmed [14] Dynamic Time Wrapping Technique Dynamic Time Wrapping (DTW) is an algorithm used for measuring similarity between two signatures, which may vary in time. The signatures are warped non-linearly in time dimension to determine an optimal match between two signatures. A. Kolmatov [2] proposed a system for online handwritten signature verification, approaching the problem as a two-class pattern recognition problem. DTW was used to establish the test signature s legitimacy. During enrollment phase, the user supplies a number of signatures. The variation in a user s signatures, are stored as measuring parameters, with a unique user identifier in the system s database. When a test signature is input to the system for verification, it is compared to each of the reference signatures of the claimed individual. The resulting minimum, maximum, and template distance values normalized by the equivalent average values stored in the user s profile. A three dimensional feature vectors was formed in classifying the test signature. The experiment was conducted using a data base from 94 people producing 619 signatures. The experiment result was 2.8% EER when tested with skilled forgeries and awarded the first place at First International Signature Verification Competition. S.A. Daramolo [13] proposed a robust automatic online signature verification system. They have used a good resolution signature capturing device i.e. graphics tablet and hence preprocessing stage is not used in the system. To establish a correspondence between feature sequences of signature samples the Dynamic Time Wrapping (DTW) algorithm were used in the training and classification stages. The proposed online signature verification system was tested using 800 genuine signatures from 200 users, 400 skilled forgeries from 200 forgers and 400 random signatures. Experiments showed that it had 0.25% error (FRR) accepting the genuine signatures. While in rejecting skilled and random forgeries (FAR) error was of 0.5% and 0.0 % respectively. DTW technique has two drawbacks viz. Heavy computational load which consumes more time, Warping forgery which makes the verification more difficult. F. Hao. [5]proposed a new warping technique which was named as Extreme Points Warping (EPW).Which involves selected important points of a signature, matching them, and then warping the segments between them. The experiment was conducted using 1000 signatures of 25 users and compared between the DTW and EPW methods. When the Euclidean distance was used, DTW showed an EER of 33.0% but EPW showed an EER of 25.4%. When the correlation coefficient was used, DTW showed an EER of 35.0% but EPW showed an EER of 27.7%. The results shows that EPW is much faster than DTW and leads to slightly better EER J. Zhang and S. kamata [9] proposed a signature verification method dividing a signature into several segments. A Modified DTW algorithm was used by considering the local properties of the signature. This can effectively remove the redundant points in the input signature. The dissimilarity scores based on the obtained segment-to-segment mapping information was calculated to evaluate whether the input signature is genuine. The Volume 3, Issue 2 March April 2014 Page 265

4 experiments reported the EER of 6.02% to 7.82% for the 60 and 40 user s database respectively. 5.3 Hidden Markov Models HMMs models have found to be well suited for signature modeling since they are highly adaptable to personal variability. HMM performs stochastic matching of the two signatures using a sequence of probability distributions of the features along the signature. Mohammad M. Shafli and Hamid R. Rabiee [11] introduced signature verification system using variable length segmentation and Hidden Markov Models. The signature was described by using the HMM as a sequence of vectors of values related to each point of signature in its trajectory. From each segment of the signature seven parameters were extracted. For each person, mean (µ) and variance (σ) of log-likelihood obtained from its HMM model divided by average segments number for his signatures in the training set and these values were saved beside the HMM model as a template for his signing process. These two values show acceptable range of loglikelihood value for genuine signature. For the verification purpose, the researchers used a database from 622 genuine signatures were collected, from a population of 69 human subjects including 12 women and 6 left hand writers. Additionally 1010 skilled forged samples were used to evaluate the performance of system. They obtained a FAR of 4% and a FRR of 12%. 5.4 Neural Networks Syed Khaleel Ahmed [14] designed signature verification system with 4 modules. They are feature extraction, reference module, sample module, and intelligent decision module. The feature selection module captures the position co-ordinates X, Y, and pressure P as a function of time of the signature were extracted from the tablet. The reference module stores the data for training. The sample module consists of data for validation. The Self-Organizing Map (SOM) of Neural Network model was used as an intelligent decision module. This system makes it easy to visualize and classify data by clustering similar data groups. The high dimensional data map into a 1 or 2 dimensions. The system performance was evaluated by using 50 epochs and 200 epochs for two different cases. The FAR and FRR were 11.43% and 20.95%for 50 epochs 9.29% and 19.05% for 200 epochs respectively. Results obtained using prototype system was encouraging. Zhan Enqi et al [16] presented online handwritten signature verification based on two levels back propagation neural network [BPNN]. The first level selects 15 statistic features as the input to BPNN. These features can reflect the signature speed and shape of signature. The second level selects wavelet features, which includes only important features extracting for matching and recognition of signature. The author used five groups of signature samples. Each group of training sample is composed of 5 genuine signature and 10 forged signatures, which include genuine, random, and skilled forged signatures. They obtained 1.07%of FAR and 6% of FRR. Nan Xu [12] proposed a method for online handwritten signature verification by using back propagation neural networks. A touch-based tablet was used to acquire the signature visual outline, the coordinates (x, y), pressure, time, length and other information in the form of vector was stored in computer. These acquired features of the signature were not sufficient to build the template of the reference signature. They considered the intrinsic properties of signature which leads to the FR errors. i.e. randomness of handwriting the signature such as the size and spacing between the words. Total of 49 features were designed using Matlab by considering the above acquired features of the signature and then calculated the corresponding characteristic values for each specimen signatures and a standard uniform vector used as a template of the reference signature. They obtained a FAR of 1.5% and a FRR of 3.5%. Table-2 summarizes the results of various Neural Networks based signature verification techniques. A high accuracy was reported by Zhan Enqi [16]. Table- 2 Performance Evaluation Parameters for Neural Networks Sl. FAR Authors Year No. (%) Syed Khallel Ahmed et al [14] FRR (%) Zhan Enqi et al [16] Nan Xu et al [12] Conclusion The signature verification system must take into an account the volatile characteristics of the signature. i.e. the signatures of the same person are similar but not identical. In addition, a person s signatures often change during their life due to age, illness and up to some extent the emotional state of the person. This makes it difficult for the researcher to achieve a better performance of the system. The signature acquiring system plays a very important role in signature verification systems. This opens another research area. The accurate segmentation of the signature is a key factor of signature verification system. Many promising techniques and algorithms have been developed but, there is still room for improvement in signature segmentation methodologies. The study of number of works did on signature verification systems shows that a large number of existing systems employ limited number of feature extraction with Volume 3, Issue 2 March April 2014 Page 266

5 the intention of reducing the algorithm size, thus making the system run faster. Among the studied papers in this review, S.A. Daramolo [10] has the best performance evaluation parameter values with FAR of 0.25% and FRR of 0.5% for skilled forgeries and completely rejecting the random forgeries. References: [1] Anil.K.Jain,FriederikeGriess, Scotto,Connel online signature verification. Pattern Recognition letters, volume 35, Issue 12, 2002 pp [2] AlisheiKolmatov, BerrinYainkoglu, Identity authentication using improved online signature verification method, Pattern Recognition letters, volume 26,Issue 15, November 2005,pp [3] Alonso-Fernandez, J. Fierrez-Aguilar, and J. Ortega- Garcia, Sensor interoperability and fusion in signature verification: A case study using tablet PC, (Lecture Notes in Computer Science 3781) in Proc. Int. Workshop Biometric Recognit. Syst. (IWBRS). Beijing, China Springer- Verlag, Oct. 2005, pp [4] Alonso-Fernandez, J. Fierrez-Aguilar, F. Del Valle, and J. Ortega- Garcia, On-line signature verification using tablet PC, in Proc. 4th Int. Symp. Image Signal Process. Anal. (ISPA 2005), pp [5] Hao, and C. W. Chan, Online Signature Verification Using a New Extreme Points Warping Technique, Pattern Recognition Letters, Volume 24 Issue 16, December [6] Shimizu, S. Kiyono, T. Motoki, and W. Gao, An electrical pen for signature verification using a twodimensional optical angle sensor, Sens. Actuators, vol. 111, pp , [7] Taguchi, K. Kiriyama, E. Tanaka, and K. Fujii, On-line recognition of handwritten signatures by feature extraction of the pen movements, IEICE Trans., vol. 71, no. 5, pp , [8] Lee, H.-S. Yoon, J. Soh, B. T. Chun, and Y. K. Chung. Using geometric extrema for segment-tosegment characteristics comparison in online signature verification. [9] J.Zhang and S.kamata, Online signature verification using, segment to segment matching. In:Int. conf. on Frontiers in handwriting recognition ICFHR,2008. [10] L. Lee, T. Berger, and E. Aviczer. Reliable on-line human signature verification systems. IEEE Trans. on Pattern Analysis and Machine Intelligence, 18(6): , [11] Mohammad M.Shafiei and Hamid R.Rabiee,a new on-line signature verification algorithm using variable length segmentation and Hidden Markov Models. Proceedings of the seventh international conference on document Analysis and recognition, [12] Nan Xu, Li Cheng, Yan Guo, Xiaogang Wu and Jiali Zhao, online handwritten signature verification by using back propagation neural networks /11/$ IEEE. [13] S.A. Daramolo and T. S.Ibiyemi, Efficient Online Signature Verification System, International Journal of Engineering and Technology IJET-IJENS, [14] Syed Khallel Ahmed, Agileswari K. Ramaswamy, anissalwa Mohd. Khairuddin and Jamaludinomar, Automatic online signature verification, A prototype using neural networks, TENCON 2009, pp [15] Thanin Maneechot and Yuttana Kitjaidure Signature Verification Using N-tuple Learning Machine Proceedings of the 2005 IEEE Engineerimg in Medicine and Biology 27 th Annual conference Shanghai,China,September 1-4,2005. [16] Zhan Enqi,Guojinxu,ZhengJianbin Ma chan&wanglinjuan, online handwritten signature verification based on two levels back propagation neural network. In: International Symposium on Intelligent Ubiquitous Computing and Education, 2009 pp AUTHORS Prathiba M.K received her Bachelor of Engineering Degree in Electronics and Communication Engineering from Mysore University in She did her Master of Technology in Industrial Electronics from Visvesvaraya Technological University, Belgaum in Presently she is pursuing Ph.D. in the area of Biometric Application under the guidance of Dr. L. Basavaraj. Her area of interest is Signal processing and Image processing. Dr. L. Basavaraj received his M.E degree in Digital Electronics from Karnatak University in 1994, and Ph.D in Electronics from University of Mysore in His researches interests include VLSI Implementation of signal Processing Applications, Biomedical Signal Processing, Offline, and on-line handwritten signature recognition. Volume 3, Issue 2 March April 2014 Page 267