Online Signature Verification: State of the art

online or Dynamic signature verification (DSV) is one of the most acceptable, intuitive, fast and cost effective tool for user authentication. DSV uses some dynamics like speed, pressure, directions, stroke length and pen-ups/pen-downs to verify the signer's identity. The state of the art in DSV is presented in this paper. several approaches for DSV are compared and the most influential techniques in this field are highlighted. We concentrate on the relationship between the verification approach used (the nature of the classifier) and the type of features that are used to represent the signature.


Literature Review
The signing process can be described at two levels: at the high level, the signing method is recovered from long-term memory; parameters are then specified such as size, shape, timing; finally, at a peripheral level, commands are generated for the biophysical muscles. So, the signing process is believed to be a reflex action (ballistic action), rather than a deliberate action. Ballistic handwriting is characterized by a spurt of activity, without positional feedback, whereas deliberate handwriting is characterized by a conscious attempt to produce a visual pattern with the aid of positional feedback [25]. The problem arises here is that many signers can produce their signatures both ballistically and deliberately based on the nature and importance of the task. The process of Signature verification is a two-class problem where the input signature is classified as genuine or forged (i.e. belonging to an impostor). The simplest approach to compare two signatures is to compute correlations between the test signature values and the reference signature values. Such point-topoint comparison usually does not work well since there are a large number of variations between the two signatures and the correlation can be affected by the translation, rotation or scaling of the signature . The decision is made by comparing the similarity score between the input and the enrolled signatures with a predetermined threshold. A tablet device is used to capture dynamic information of a signature. Such dynamic information is a set of sequences of sampled points over time. Each sequence depicts an action performed during the signing process. Signatures from the same person may have different trajectory lengths due to local stretching, compression, omission or additional parts, and hence they are represented by feature vectors of differing lengths. Therefore, straight forward methods, such as the Euclidian distance or autocorrelation, are not very useful in calculation of the dissimilarity value between two signatures [6]. Such point-to-point comparison usually does not work because some portions of any two genuine signatures of the same person may vary significantly. To overcome the problem, non-linearly approaches such as Dynamic Time Warping (DTW) algorithm and Hidden Markov Models (HMM) are commonly used in aligning two signatures. Plamondon and Lorette [141] categorized the various signature verification methods into two types: (i) parametric (Feature)-based approach: In which, each time a signature is presented to the system, equal number of global features derived from the acquired signature trajectories are extracted then a distance measure (ex. Euclidian distance) is used to compare reference and test signatures, and (ii) functional-based approach: In which time sequences describing local properties of the signature are used for verification (e.g., position trajectory, velocity, acceleration, pressure). In the parametric approach, a complete feature vectors are extracted from the signatures and compared to obtain similarity measures. In this approach, once a set of features has been selected, only the values of these features are stored. Parametric approach may also called signature compression because it only needs to store a set of features' values which may be useful if, for example, the reference signature needs to be stored on a card. When a test signature is presented, only the features' values are extracted and compared, not the signature. The main research issues [26] on the parametric approach are: (i) How many global features are sufficient?, (ii) Do the signatures require some transformation before computing the features, for example resizing, rotating or smoothing?, (iii) How many sample signatures will be used in computing the reference signature?, (iv) Would the reference signatures be updated regularly? If yes, how often it will be updated?, (v) Should user-dependent set of features be used rather than the same features for all?, (vi) Should user-dependent threshold be used rather than the same for all?, (vii) How is the distance between a test signature and the reference signature is computed?. When the algorithm uses a large number of features, a single feature can unduly influence the decision through deviating far from the mean value, even if the other features have values close to their means for the genuine reference set. To solve this problem, it either must have a large threshold for the norm of the distances or use some criterion similar to the Majority Classifier (MC) [12]. The Majority Classifier (MC) accepts a test signature as a genuine one if the number of feature values which pass a predetermined test is larger than half the total number of tested features. Function-based approach treats the signing process as a set of one-dimensional discrete time functions and performs the matching process based on an elastic measure such as DTW [7,9,14,15,17,18,31,36,44,45,51,54,67,71,72,79,81,84,85,86,93,96,98,99,101,104,106,112,122,124,126,129,140,142,143,146,163,167,200,221,222] or HMM [2,5,21,29,30,32,58,61,73,128,131,132,135,149,155,157,168,175,187,192,203,207,216,218,223,224,225]. DTW performs flexible matching of local features of two signatures. HMM performs stochastic matching of the two signatures using a sequence of probability distributions of the features along the signature. The major research issues [26] on the functional approach are: (i) How

FEATURES-BASED APPROACHES
Depending on the type of the capturing device, data such as x, y coordinates, pen velocity, pen pressure and pen tilt are recorded. A fixed number of global features is extracted from these data and arranged in a feature vector which describes discriminative characteristics of the signature and stored as a template. Gupta and Joyce [162] studied the effect of using six global features on the performance of the verification system. The features studied are: total time, the total pen-up time, number of velocity sign changes in the x-and y-direction, number of acceleration sign changes in the x-and ydirection. There are many different distance metrics [26] that can be used to compute the distance between a test signature and the corresponding reference signature. The most common method to find the similarity between the input feature vector and the stored template is to use a classical distance measure such as Euclidean distance, Mahalanobis distance, Canberra Distance, Euclidean Distance, City Block distance and Hamming Distance. Some of the algorithms that used features-based approaches can be found in [4,9,12,22,28,33,49,52,65,93,94,107,139,161,162,179,193,195,199,200,205,208,209,210,215]. L. Lee et al. [12] discussed a number of techniques for feature selection from a set of 42 features and 49 normalized features. Kim et al. [33] studied 75 features and selected the best combination of features for each individual using an evaluation function called the Degree of Difficulty to Forge (DDF). Ketabdar et al. [49] selected the most discriminative global features by combining a modified Fisher ratio cost function and a near-optimal feature space search algorithm. Taherzadeh et al. [161] extracted 30 features from the online signature and then categorized them into four categories: Direction-related, Pressure-related, Dynamic-related and Shape-related. Carlos Vivaracho-Pascual et al. [205] proposed a DSV system that is new in both feature extraction and classification stages. The proposed approach is based on the use of size normalized signatures, which allows similarity estimation to be performed by an easy distance calculation between vectors. They used the fractional distance, instead of the more typical Euclidean one, so as to overcome the concentration phenomenon that appears when data are high dimensional. David Aristizábal et al. [208] proposed a DSV system that extracts 42 features from seven time-series functions by applying 6 nonlinear dynamic techniques to each series. Julio Romo and Rogelio Silva [209]

Dynamic Time Warping (Elastic Matching)
After modeling each signature by a set of time based vectors, there must be a method to compare signatures based on their vector representations. Since there are variations among genuine signatures of the signer, the time based vectors representing these signatures may be of different lengths which makes it hard to use a naive element-wise comparison of signatures, such as Euclidian distance between two vectors. One of the simplest and successful approaches for online signature verification is DTW. DTW is used to measure similarity between two sequences that vary in time or speed. DTW finds the best nonlinear alignment of two vectors such that the overall distance between corresponding vector elements is minimized in least square sense. In DTW the input and reference signatures are compared by using dynamic programming (DP) matching algorithm which can manage the variability on the signature's length. DTW is used to find the best matching path, in terms of the least global cost, between the test signal and the stored template. The algorithm is a well-known and widely used for aligning vectors of different lengths. This approach is also called elastic matching or elastic distance because it is tolerant to local variations in the compared signatures. DTW finds an optimal match between two sequences that may contain stretched or compressed sections. The idea falls behind that we can predict missing information in the compared sequence, if we have long enough segments to match. DTW has been widely used in online signature verification systems to cope with variable length of signatures. The approach relies on minimization of a global cost function that contains local distances between the two signatures that are compared. Many researchers use DTW to obtain a point to point correspondence between two similar signatures. DTW is used to accommodate the timing but not the X and Y coordinates of the signature. By applying DTW to X and Y signals, it will distort the shape of signature. DTWbased techniques can be further divided into several sub-categories: Full Signal Warping [14,15,51,72,79,81,85,176,212], Extreme Points Warping [17,44,45], Shape contexts DTW [78,83,194], Continuous DTW [67,71,86,99,101,114,167,214], Mahalanobis-based DTW [45,106,140,158], PCA, NCA and MCA based DTW [18,146,226], Bidirectional backward-merging DTW [147], GA and PA-based DTW [121,143].

Hidden Markov Model (HMM)
Hidden Markov Models (HMM) were introduced in the pattern recognition field as a robust method to model the variability of discrete time random signals where time or context information is available [19]. HMMs are widely used in classification of input patterns and they have a great ability to model stroke-based sequences. These models have found to be well suited for signature modeling since they are highly adaptable to personal variability [80]. The underlying assumption of the HMM is that the signal can be well characterized as a parametric random process, and that the parameters of the stochastic process can be estimated in a precise, well-define manner [20]. In HMM, the probability of moving from one state to another depends on both the transition probabilities and the previously visited states. The left-to-right topology is the most applicable in DSV systems since it is considered well suited for signature modeling. some of the techniques that used Ergodic topology can be found in [135,168]. Juan J. Igarza et al. [105] compared between different topologies used for DSV and reported that the worst results were obtained using Ergodic or generalized models, in which transitions between all the states are allowed. Donato Impedovo and Giuseppe Pirlo [80] reported that the ring topology can be used to guarantee rotation invariance of the signatures. the ring topology is equivalent to left-to-right topology with a transition from the last state to the first state. An algorithm such as Viterbi algorithm can be combined with HMMs to find the most likely state sequence that generates the observed sequence. Other algorithms such as the Baum-Welch or the Segmental K-means can be used to update the parameters of the model to maximize the probability of the sequence generated by the model. Julian Fierrez-Aguilar et al. [19,22] proposed applying HMM to 14 time sequences derived from dynamic functions. Ly-Van et al. [30,149] proposed an HMM-based DSV approach, that fuses two complementary information levels issued from the same continuous HMM with a multivariate Gaussian mixture in each state. Yang et al. [37] extracted normalized angles along signature trajectory and model the generation of this sequence by HMMs. Dolfing et al. [38] proposed a DSV system in which each signature is modeled by a single HMM without any preprocessing. Rigoll and Kosmala [39] presented the first systematic comparison of on-line and off-line methods for signature verification using HMM on the same database. Ortega-Garcia et al. [43] proposed a function-based approach to online signature verification using an initial set of 8 time sequences. Kashi et al. [55,87] proposed a DSV system using local and global features where Local features of inclination angle and difference between adjacent angles are used to build a left-to-right HMM model. Fierrez-Aguilar et al. [59] studied the two function-based approaches for DSV. These approaches are divided into local and regional depending on the features used for matching. The authors made some modifications to the local approach using DTW that was proposed by Jain et al. [24]. It was found that the DTW system outperforms the HMM system. It was shown that combining DTW and HMM and employing user-dependent thresholds performs significantly better than individual DTW or HMM systems. M. Shafiei and H. Rabiee [90] proposed a DSV system Using Variable Length Segmentation and HMM. Juan J. Igarza et al. [105] presented a research of various HMM based techniques for DSV. Different topologies are compared in order to obtain an optimized high performance signature verification system. Daw-Tung Lin and Yu-Chia Lia [128] presented a DSV system using Hidden Semi-Markov Model (HSMM). Zhong-Hua Quan and Kun-Hong Liu [135] proposed an online signature verification system based on Ergodic HMM and a time delay Neural Networks. Muramatsu and Matsumoto [155] proposed a HMM-based approach for online signature verification using only pen position trajectories. Pascual-Gaspar and Cardenoso-Payo [175] proposed an online Signature Verification system based on HMM with user-dependent structure. Colin Sindle [131] presented a DSV system based on HMM that uses only the local features after some novel preprocessing steps. Enrique Argones et al. [168] proposed a novel approach for online signature verification based on building HMM user models by adapting an Ergodic Universal Background Model (UBM) with a fixed number of states. M. Martinez-Diaz et al. [138,197] presented an HMM-based DSV system in which the number of states and Gaussian mixtures of the HMM are specifically set for each user. It was shown that when using the user-dependent structure, the system error rate is significantly decreased. Mitsuru Kondo et al. [60] proposed a DSV approach by performing nonlinear separation using Bayesian Markov Chain Monte Carlo (MCMC). Daigo Muramatsu et al. [46] used a similar approach to that proposed in [60] except that they used four distance measure. The fourth distance measure is the data length difference.

Gaussian Mixture Models (GMM)
Gaussian Mixture Models (GMM) are considered to be a stable classifier because small variations in the sampling of the training set will produce not a large change in the parameters of the trained classifier. Some of the techniques that used GMM are Jonas Richiardi and Andrzej Drygajlo [27], Martinez-Diaz et al. [23], Andreas Humm [63], JUGURTA MONTALVÃO [156], Wan and Liang [192], Jonas Richiardi et al. [198], Riccardo Corradin [220]. GMMs are well-known flexible modeling tools can approximate any probability density function. Richiardi et al. [27] was the first one who introduced and motivated the use of GMMs for DSV. JUGURTA MONTALVÃO [156] proposed another model for DSV based on GMM which is a degenerated version of an HMM with only one state. Andreas Humm et al. [63,150,151] proposed a novel multimodal user authentication system based on a Combined Handwriting and Speech Modalities (CHASM). Mandeep Kaur et al. [137] proposed a Multimodal Biometric System Using Speech and Signature Modalities.

Structural aapproaches
Structural approaches are divided into three categories string matching, structural description graph and tree matching. String matching approaches [24,34,35,56,136,217] can be used for both verification and identification purposes. Structural description graph (SDG) approaches [54,99,101,111] are used to verify the structural organization of a questioned signature. Tree matching approaches [176] involve building a tree of peaks and valleys of a signature. In this approach, two trees are built for the test signature and the reference signature, then the distance between them is computed in terms of minimum number of operations needed to transform one tree to the other. It was found that no technique was globally superior to the other two although regional correlation was often better and much faster [80]. Anil K. w w w . i j c t o n l i n e . c o m Jain et al. [24] proposed a DSV system using string matching. Two types of local features are extracted from the x and y coordinates: spatial and temporal. Local features of each signature are represented as a string at each sampling position. A customized DTW that includes a penalty for differences in the number of strokes is used to find dissimilarity values. Gupta and Joyce [35] described a technique which uses x and y profile extrema (peaks or valleys) values to capture a signature's shape. X, Y, speed and acceleration profiles are calculated at each sampling point, then these profiles are divided into valleys and peaks. Each peak or valley in each profile is assigned a unique symbol. To compare two signature shapes, the string representations of the extrema for each are found and the two strings are then compared and distance between them is calculated. Texas Instruments [34] adopted this method and implemented it on the TMS320 Digital Signal Processors, after adding neural clustering techniques to enhance the discriminating power and arrive at a very simple and low-cost solution that can be embedded in existing pen-based systems, such as handheld computers and transaction units. Gupta and Joyce [56] presented four experiments to evaluate the performance of the technique in [35] and a number of its variations. Felicia Soedjianto et al. [136] proposed a DSV system based on dominant points of the signature. These dominant points are converted into codes which is used to determine the stroke direction. The code extracted from the input signature is then matched with the stored template. The alternation of the stroke direction will be converted into a number which is compared to a threshold. Sascha Schimke et al. [217] presented a new approach for DSV system based on event-string modeling of features derived from pen-position and pressure signals. The Levenshtein Distance which is a distance measure well known from textual pattern recognition, is used for comparison of signatures. the classification is carried out by applying a nearest neighbor classifier. Dimauro et al. [54] proposed a DSV system that finds the fundamental components of the test signature and then checks if these components appear in the sequence derived from the reference signatures using a knowledge-base that consists of the Component Reference Table and the Structural Description Graph. Kaiyue Wang et al. [111] proposed an on-line signature verification approach based on graph representation and graph matching techniques. Kai Huang and Hong Yan [101,114] used the dynamic warping of the sequences of basic handwriting strokes in the reference signatures to deduce the stability and style-variation characteristics of the signer. The technique models each signature by a weighted Structure Description Graph (SDG) of the handwriting components. SDG was first used by Dimauro et al. [54,96] to model handwriting strokes separated by pen-lifts. The stable handwriting features are emphasized. The style variations allows following all permissible paths within the SDGs. Parizeau and Plamondon [176] compared between three different signal matching algorithms for online signature verification: regional correlation, DTW, and skeletal tree matching (STM).

Statistical-Based Approaches
Most of existing online signature verification approaches involves preprocessing stages such as re-sampling, normalization and alignment of signatures in order to obtain robust signature features of equal length. Unfortunately, this process may cause unique properties of the user signatures to be lost. In the other hand, statistical analysis-based DSV systems [23,42,50,102,133,154] avoid preprocessing step and still be able to develop a robust mechanism for detecting forgeries in the claimed signature. Herbst and Liu [53] proposed two approaches for segmentation: the first uses pen-down and pen-up occurrences within the signature and the second uses equal time components of the signature. Before applying the correlation test, the algorithm rejects any signature that is larger or smaller in time than 20% of the reference signature. This resulted in almost 60% of forgeries being rejected. Once passing this initial test, the correlation test is applied after the segments are shifted if necessary by up to 20% of the reference signature's segment duration, extra pen lifts are eliminated and correlation results are weighted to penalize excessive shifting, and weighted by the reference signature length of the segment since larger segments are believed to be harder to forge. Martinez-Diaz et al. [23] proposed a score normalization approach for DSV using Universal Background Model Bayesian adaptation to generate the client model. This Universal Background Model (UBM) is trained once and can be then used for all the claiming users. The UBM is chosen to be The Gaussian Mixture Model (GMM) trained with feature vectors from a pool of Users. In the verification process, the user signature is compared to his claimed template and the resulting score is normalized by its similarity to the UBM. The similarity between signature samples is computed using statistical pattern classification techniques. Hastie and Kishon [42] proposed a statistical model for DSV that assumes that the test signature consists of a reference signature which is transformed from occasion to occasion. The model consists of a template signature for each individual, and several factors controlling variations in each rendition of this template. The estimated template contains the mean of a sample of signatures from an individual as well as the variation factors that are used to estimate several measures of variance. Winston Nelson et al. [50] discussed three methods for on-line signature verification that are based on statistical models of features that summarize different aspects of signature shape and the dynamics of signature production. Rather than using detailed comparisons of shape and speed patterns of a test signature with the model of the reference signatures, the authors developed a decision criteria based on statistical models for the set of global features that summarize various aspects of a signature's shape and dynamics. Jonas Richiardi et al. [133] proposed a middle way between static discriminative classifiers and dynamic generative models, by using a static generative model together with specific feature extraction steps. They used a specific topology of static Bayesian network to perform classification of multivariate time series. Results showed that when using appropriate features extracted from the time series, the framework of static Bayesian networks offers an elegant approach to classification of multivariate time series. Hairong Lv and Wenyuan Wang [224] proposed a DSV system based on Dynamic Bayesian Network. Loris Nanni et al [8,10] proposed a feature based DSV system based on Parzen window classifier (PWC). It used global analysis of input signatures and one-class classifiers. They reported that the fusion between Principal component analysis description and PWC drastically outperforms the state-of-the-art works using only PWC. Alisher Kholmatov and Berrin Yanikoglu [102] proposed a DSV system based on the Bayes classifier on the original data, as well as a linear classifier used in conjunction with Principal Component Analysis (PCA). Shern Cheng Yau [154]

Support Vector Machine
Support vector machines (SVM) have been widely used to solve classification and regression problems. The main advantages of SVM are having a convex objective function with efficient training algorithms and good generalization properties. SVM use kernel functions to measure the similarity of two data objects, in DSV case, time series. However, standard kernel functions such as Gaussian or polynomial kernels are not well suited for time-series processing [80]. Yenwei Lee [182]

Neural Networks
In order to learn the signature verification system to distinguish between genuine signatures and forgeries, we must provide it with samples of the two types. Since it is hard to define a class of forgers, some authors prefer to use genuine signatures that have been deformed instead. F. LECLERC and R. PLAMONDON [11] presented a comparison between neural network-based DSV systems that appear from 1989 to 1993. Many authors used the parametric approach to incorporate the usage of neural networks for the verification of online signatures. Tarig Osman [97] proposed a neural network-based online signature verification system using vector autoregressive modeling and a velocity segmentation scheme. Artur Szklarzewski and Marcin Derlatka [120] proposed an On-line Signature verification System based on the single-output multilayer perceptron (MLP). Daniel Oldham [164] proposed a DSV system based on time-delay neural network. Wan-Suck Lee [166], proposed a DSV system based on a multilayer perceptron neural network with a single hidden layer; where the inputs to the neural network are second order AR coefficients. Lina Mi [186] proposed a DSV system based on a three-layer feed-forward neural network with a Gaussian ridge activation function of neurons in hidden layer and output layer of neuro-template. P. Gentric and J. Minot use the signals x(t), y(t) and p(t) (the coordinates and pressure obtained by a special sensor) and elastic matching combined with dynamic programming to define four measures: mean pressure distortion, written shape distortion, dynamic difference and the ratio of signature durations [11]. M. MAILAH et al. [13] employed two multilayer feed-forward (MLF) neural network to train and verify the data. Tham Keit et al. [145] used time series modeling with autoregressive (AR) technique to calculate the AR coefficients from all segments of the signature. These coefficients are then used to train a multi-layer perceptron (MLP) neural network using the back propagation algorithm. Milton and Fernando [148] proposed a DSV system based on ANN. Julio Martínez-Romo et al. [171,172] presented an on-line handwritten signature verification based on optimal feature representation (OFR) and neural-network-driven fuzzy reasoning (NND-FR). Alan McCabe et al. [173] proposed a DSV system using neural networks. Several static and dynamic features are extracted and used to train the neural network. Several Network topologies are tested and their accuracy is compared. Adam Czajka and Andrzej Pacut [174] proposed a DSV system based on neural networks. first, cluster analysis is applied to segment the feature space into sub-regions of "similar" signatures to facilitate the classification and verification tasks. Then, a two-layer sigmoidal perceptron was used to approximate the classification function. The authors also applied an adaptive radial-basis network called the RCE network. finally, the two networks are compared. Experimental results showed that, the RCE network failed in verification tasks.

Transform Domain Approaches
Time domain graph shows how a signal changes over time, whereas a frequency-domain graph shows how much of the signal lies within each given frequency band over a range of frequencies. The background and advantages of transform domains are beyond the scope of this work. This section only concentrates on the techniques that used transform domain. Transform domain approaches include Fourier descriptors-based approaches [68,108,116,123,130,139], Discrete Wavelet Transform approaches [74,76,109,110,117,125,127,130,160,183,184,185], Zernike moments [57], Linear Prediction Coding [40,41], Mellin transform [100], Discrete Cosine Transform [47,88,144], Gabor [91,92], Autoregressive [97,103,153] and Linear Programming Descriptor [47]. Fourier descriptors are invariant with respect to translation, rotation and magnification of the trace. Impedovo et al. [108] used the one-dimensional spectral analysis to establish the variability of a writer's signature in the frequency domain. Mingfu Zou et al. [116] presented a novel DSV approach using local shape Analysis. First, the input signature is segmented into several segments using HMM, then each two adjacent segments are combined to form a long segment and finally, the spectral and tremor information of this segment are calculated using Fast Fourier Transform (FFT). The test signature is accepted or rejected based on the similarity between w w w . i j c t o n l i n e . c o m its spectral and its prototype. Giovanni Dimauro [68] proposed a DSV system based on Fourier analysis. Fourier descriptors are used to extract a set of discriminant features that can be used to describe the shape of the signature. The verification system depends on one-dimensional spectral analysis of the strokes that separate ‗writing units'. Kholmatov and Yanikoglu [139] presented a preliminary individuality model for online signatures using the Fourier domain representation of the signature. The normalized Fourier coefficients are used as a global features describing the signature.
Wavelets are a powerful statistical tool which can be used for a wide range of applications such as Signal processing, Data compression, Smoothing and image de-noising, Fingerprint verification, DNA analysis, protein analysis, Bloodpressure, heart-rate and ECG analyses, Finance, Internet traffic description, Industrial supervision of gear-wheel, Speech recognition and Computer graphics. DWT can be also used authenticating online signatures. Some of the techniques that used DWT for online signature verification are in [74,76,109,110,117,125,127,130,160,183,184,185]. Nakanishi et al. [109,183,185] presented a DSV system based on the DWT and the adaptive signal processing. The same author [74] extracted features of pen-position and pen-movement angle by the sub-band decomposition using the DWT. Low frequency sub-band signals are used as features for suppressing the intra-class variation in the pen-movement angle. The same authors [76] proposed reducing the computational complexity of the DWT domain on-line signature verification by utilizing the pen-movement vector as an input parameter. Sánchez-Diaz et al. [110] proposed a DSV system based on applying a modified DTW algorithm to the signature wavelet coefficients. DWT is used to analyze the temporal and spectral properties of non-stationary signals. The modification on DTW relies on using direct matching points (DMP) information to dynamically adapt the similarity measure. Chang et al. [117] proposed a DSV system based on wavelet transform of global feature. Firstly, signature data were preprocessed using normalizing, de-bouncing and Gaussian filtering. Then wavelet transform was used for feature extraction. Five functions were used to represent the signatures. Each function was compressed using the Duabechies-4 wavelet. Zhang and Wu [125] proposed a DSV system that uses wavelet transform to extract characteristic points of signatures and builds 5-dimension feature sequences, then matches these sequences of test signature and its corresponding templates and fuse five decisions by majority voting strategy to make final decision. Wang et al. [127] proposed a DSV system using wavelet packet, which performs better than classical wavelet transform in majority of cases. In the DWT, only the approximation signal at each level is decomposed, without further operations on the detail signals. In order to allow decompositions of detail signals at each level, the same process of decomposing approximation signals can be applied to them. The resulting decomposition tree will be a complete binary tree since at each node, the signal is decomposed. This type of wavelet decomposition is called wavelet packet (WP). Hong-Wei Ji and Zhong-Hua Quan [160] proposed a DSV system based on Wavelet transform and SVM.
Radhika et al. [57] proposed a DSV system based on Zernike moments. The Zernike moments are a set of complex polynomials that are orthogonal on the unit circle. Zernike moments are orthogonally stable, rotation invariance and provide image reconstruction. A lower order Zernike moments are used to analyze the shape of the acceleration plot. The low order moments represent the global shape of a pattern and the higher order represents the detail. High order Zernike moments are not used because they are more sensitive to noise and take significantly longer time to compute. The acceleration values of an online signature are represented by a weighted sum of Zernike polynomials. Quen-Zong Wu et al. [40,41] [113] proposed a DSV system in which signature representation is performed using Sigma-Lognormal modeling. Jianbin Zheng et al. [196] applied discrete Fréchet distance to on-line signature verification using only coordinate position features of extreme point.

Intelligent Approaches
A large number of features can be extracted from the signature but not all these features are used in the verification system because some of them may be unfavorable for verification of particular signatures. The most important issues in the signature verification are how to select a reduced set of features that enhance the discriminative power of a signer and how can a legitimate analysis be made. Genetic algorithm (GA) is used to provide an effective solution for the problem of feature selection. GA can determine the type and the number of features to constitute the feature set for the signer that will maximize a suitably chosen discriminating function. The intelligence in this verification model is made up of fuzzy logic and genetic algorithm. The advantages of Fuzzy logic is its ability to solve complex problems by determining outputs of a given set of inputs without using conventional mathematical models. XUHUA et al. [82] proposed a new method to select the partial curves and features of the curves of signatures for verification using GA. The features are determined by chromosomes and the crossover points are determined by the order number of strokes. W. Wijesoma [62,89] proposed a DSV system that uses the Genetic Algorithm (GA) to determine the optimal personalized features for each subject and Fuzzy Logic to transform the heuristics conceived by a human verifier to rules that govern the behavior of the DSV system. The fuzzy system takes in the total degree of match (TDOM) of shape and the TDOM of dynamics as its inputs. The output linguistic variable of the fuzzy block is the Degree Of Authenticity (DOA) of signature which is compared to a w w w . i j c t o n l i n e . c o m threshold to accept or reject the signature. Di Lecce et al. [66] proposed a multi-expert system for DSV that combines three experts by majority voting. The system consists of three experts. the first expert uses shape-based features and performs signature verification by a wholistic analysis. The second and third experts use speed-based features and performs signature verification by a regional analysis. The decisions of the three experts are combined by majority voting: if at least two decisions are "genuine" the final response is "genuine"; if at least two decisions are "false" the final response is "false"; otherwise the final response is "rejected". Ji Moon et al. [70] proposed a Multi-Confidential Levels for DSV using Data Mining Techniques. Two distance measures are computed. The feature-based approach uses the Canberra Distance (CD) to calculate the similarity between the input and the reference signature. Function-based signature verification method uses DTW to determine the similarity between reference signatures and the input signature. Jianbin and Guangxi [75] proposed a DSV system based on Evolutionary Computation (EC). Evolutionary Computation is a subfield of artificial intelligence that involves a metaheuristic or stochastic based continuous and combinatorial optimization methods which are mostly applied for black box problems. A. A. Zaher [77] proposed a robust hybrid technique for dsv using intelligent encoding of spatio temporal data. Joanna Putz-Leszczynska and Andrzej Pacut [195] proposed a DSV system based on a Global Classifier That Uses Universal Forgery Features. The idea of universal forgery features, where a global classifier is used to classify a signature as a genuine one or, as a forgery, without the actual knowledge of the signature template and its owner. This classifier is trained once on a group of historical data.

Protected Approaches
In this approaches, a set of noninvertible transformations are applied to the biometric template which can be represented by a set of sequences, in order to generate multiple transformed versions of the template. Once these transformations are made, retrieving the original data from the transformed template is computationally as hard as random guessing. Some of the techniques that used protected approaches are in [61,130,132,188,189,190,191,211]. Loris Nanni et al. [65] proposed a template protection scheme employing the BioHashing and the BioConvolving approaches. Manuel R. Freire [189,190] presented a biometric template protection approach for dynamic signature verification that uses auxiliary (helper) data that allows the matching with secure templates but do not provide information to a potential attacker. Kuan et al. [130] proposed a method for secure computation of biometric hash on dynamic signatures using BioPhasor mixing and 2^N discretization. Emanuele Maiorana et al. [61,132] proposed a protected DSV system that uses BioConvolving to guarantee security and renewability of biometric templates and HMM-based matching strategy. The security and privacy issues still represent the main problems for the deployment of biometric-based authentication systems.

Semi-Online Approaches
Instead of using tablets for capturing online signatures, there are several techniques that used hand gloves [48,64,134,219] to capture dynamics of the hand during signing, or camera to track the signature production [115, 118, 152 159, 165, 180, 181, 204]. Nidal Kamel et al. [48,134,219] used data glove as an input device and the singular value decomposition (SVD) numerical tool for modeling and matching. Another novel vision-based techniques for personal identification that uses a camera for tracking the signature is proposed by Munich and Perona [115, 118, 152 159, 165], KIRAN GURRALA [204], T. Venkatesh et al. [95] and Daigo Muramatsu et al. [180,181]. The system falls within the category of on-line systems since the visual tracker of handwriting captures the timing information in the generation of the signature. The drawback of the technique is that if signatures have extremely fast strokes, the system loses track of the pen tip, and if users were asked to sign more slowly, the consistency in the signing behavior will be lost.