Online Signature Verification System

Introduction

A signature is treated as an image carrying a certain pattern of pixels that pertains to a specific individual. Signature verification problem therefore is concerned with determining a whether a particular signature truly belongs a person or not. Signatures are a special case of handwriting in which special characters and flourishes are viable. Signature Verification is a different pattern recognition problem as because no two genuine signature of a person are precisely the same. Its difficulty also stems from the fact that skilled forgeries follow the genuine pattern unlike fingerprints or irises where fingerprint or irises from two persons vary widely. Ideally, interpersonal variation should be much more than intrapersonal variation. Therefore, it is important to identify those features, which minimize the intrapersonal variation and maximize interpersonal variations.

In case of online signatures, many dynamic features are also considered in addition to the static features. This makes it harder to forge, even if the skilled forger is able to copy the shape it is very unlikely that he can simultaneously reproduce all the dynamic features as well. Online signature verification scheme extracts features from the signatures, which characterize the signature. The feature statistics of a training set of genuine signatures are used to generate a model, which is further used for testing. Selecting a good feature set to represent the model is a very important part for a verification scheme.

There are mainly three direct approaches for the selection of features. The first one is based on point-to-point local feature relating position, velocity, acceleration, pressure etc. The second one deals with the global features like writing time, pen up time, number of breakpoints, maximum/minimum pressure and speed, pen direction at crucial points like starting and ending points. The third one deals with the shape of the signature. There are several methods of using local features in signature verification. The most commonly used strategies are matching by Dynamic Warping and by using Hidden Markov Model (HMM). Dynamic warping approaches give a flexible matching of the local features. An HMM performs stochastic matching of a model and a signature using a sequence of probability distributions of the features along the signature. We used the Dynamic warping approach to match the local features, mainly because learning techniques like HMM requires many test data, which does not suit our condition. In case of global feature verification, a number of features (global) are extracted, these features are then compared with reference signature features. The point-to-point local feature comparison is more sensitive to handwriting variations than the other approaches but is also more resource intensive. One more important point, which should be noticed while selecting the features, is the variation tolerance quality of the feature. For example the direction, speed acceleration features are rotation variant. Hence, if these are used for verification, the signature sample has to be rotated to a fixed slant before extracting these features. Similarly the displacement, area features are size dependent. The signature has to be transformed to a fixed size before extracting them.

There has been a lot of work on online signature verification using learning technique. Shafiei, Rabiee applied the HMM approach, where one segments a signature based on its perceptually important points and then computes for each segment a number of features that are scale and displacement invariant. The resulted sequence is then used for training an HMM to achieve signature verification. They got a false acceptance rate (FAR) of 4% and a false rejection rate (FRR) of 12% on there database, which includes 622 genuine signatures and 1010 forgery signatures collected from 69 individuals. Mohankrishnan, Paulik and Khalil (1993) propose a method based on an autoregressive (AR) model that treats the signature as an ordering of curve types. A database of 58 sample signatures from 16 individuals was used for testing. No skilled forgeries were available but random forgeries were used. There total error rates using threshold value that gave equal FAR and FRR for each user vary from a low of 7.92% to a high of 21.83%.

Methodology

Online verification methods can have an accuracy rate of as high as 99%. The reason behind is its use of both static and dynamic (or temporal) features, in comparison to the offline, which uses only the static features (Ramesh and Murty, 1999). The major differences between offline and online verification methods do not lie with only the feature extraction phases and accuracy rates, but also in the modes of data acquisition, preprocessing and verification/recognition phases, though the basic sequence of tasks in an online verification (or recognition) procedure is similar to that of offline. However, online signatures are much more difficult to forge than offline signatures (reflected in terms of higher accuracy rate in case of online verification methods), since online methods involve the dynamics of the signature such as the pressure applied while writing, pen tilt, the velocity with which the signature is done etc. In case of offline, the forger has to copy only the shape (Jain and Griess, 2000) of the signature. On the other hand, in case of online, the hardware used captures the dynamic features of the signature as well. It is extremely difficult to deceive the device in case of dynamic features, since the forger has not only to copy the characteristics of the person whose signature is to be forged, but also at the same time, he has to hide his own inherent style of writing the signature. There are four types of forgeries: random, simple, skilled and traced forgeries (Ammar, Fukumura and Yoshida, 1988, Drouhard, Sabourin and Godbout, 1996). In case of online signatures, the system shows almost 100% accuracy for the first two classes of forgeries and 99% in case of the latter. But again a forger can also use a compromised signature-capturing device to repeat a previously recorded signature signal. In such extreme cases, even online verification methods may suffer from repetition   attacks when the signature-capturing device is not physically secure.

Although the basic sequence of tasks in online signature verification is similar to that of offline methods, the modules differ from each other especially in the ways the data acquisition, preprocessing and feature extraction is carried out. More specifically, the sub-modules of online are much more difficult with respect to offline (Jain and Griess, 2000). Figure 1 gives a generic structure of an online signature verification system.

The online verification system can be classified into two modules- (i) Database Preparation Module and (ii) Verification Module. Database Preparation Module consists of two sub-modules and they are (a) Enroll Module and (b) Training Module while the other module, Verification module can be divided into two modules (a) Matching Module and (b) Decision Module.