A Study Of The Authentication Based On Online Signatures

Biometrics based authentication techniques, including voice and fingerprint identification, face recognition, retina scan, and signature verification, are becoming increasingly popular. One reason for this popularity is the explosive expanding of the need for positive personal identification in law enforcement, information security operations, and commercial transactions. A more important reason is the inherent advantages of biometrics based methods over the classical handled tokens (card, ID, passport, etc.) and knowledge based (password, PIN, etc.) authentication methods. Among these biometrics methods, signature verification is particularly important because it is one of the oldest means of identity validation and has been accepted widely while other methods unavoidably have the stigma of being associated with criminal investigation.Generally speaking, signature verification can be divided into two groups: online and offline. Offline signatures are captured once the writing process is over, thus only static images are available. And the online signatures are acquired during the writing process with a special instrument, such as a digital tablet. As a result, there is always dynamic information in the case of online signatures, such as velocity, acceleration and pressure etc., as well as the static shapes. It is believed that the dynamics of signature is more difficult to imitate than the static shape of signature. So, online signature verification can usually achieve better performance than the offline instance. Many researchers have devoted to the research of online signature verification, and numerous methods have been proposed. However, the performance of the verification, especially the accuracy of recognition is far from being satisfactory.The works of this dissertation are focused on the methods and the application of online signatures verification. The main works of this dissertation can be described as follows:(1).A program based on the Wintab interface to acquire online signatures is designed. And a database of Chinese signatures is built. The Wintab interface is the standardized programming interface to digitizing tablets, three dimensional position sensors, and other pointing devices by a group of leading digitizer manufacturers and application developers. Most of the digital tablets on the market support the features of this specification. This thesis discusses the programming of a Windows application which acquires online signatures using a Wacom tablet. And then 880 signatures which belong to 22 subjects and include 440 genuine signatures and 440 forgeries are acquired from the students in our lab.(2).A string match based algorithm for alignment of signatures' crucial points is proposed. At first, the extrema, start and end points of the storks are extracted. And then 16 types are assigned to these crucial points, according to the local geometric shapes. An approximate string match method is employed to find out the alignment between the strings which are composed of the types of the crucial points. And by comparing the edit distances between these strings, this method can detect the random forgeries quickly. Since there are distinct differences between a random forgery and a genuine signature in geometric shape and which can be represented by the types' string of the crucial points.(3).An improved time warping algorithm which combines the matching of crucial points and the DTW of the local segments is proposed. The DTW is a successful algorithm for the comparison of time series in many domains, but it often explains the variation on the extent of time series by warping the time aspect. This results unintuitive alignments where a single point on time series maps into a large subsection of another time series. This problem is circumvented in this thesis by resampling each pair of aligned segments based on the curve length. That is, the signatures are segmented by the crucial points at first, and then the crucial points are aligned, and each pair of aligned segments are reampled so that they contain the same number of sample points. However, this strategy demands that the matching of the crucial point must be accurate enough.(4).An order one predictive HMM based approach is proposed for online signature verification. For continuous HMM, the probability density functions (PDF) of observations are often represented as the Gaussian models. But such models fail to describe the dynamics of observations, so in this thesis a weak measure of dynamics is included by assigning a Gaussian model for the order one predictive errors of observations. The experimental results validates that the recognition of the states by the order one predictive HMM is much more accurate than the standard HMM. Moreover, the HMM is not only a verifier, but also is employed to segment and align signatures. Thus the local comparison can be performed based on the aligned segments; and the HMM based verification and the local comparison are combined for verification.(5).A new approach based on HMM/ANN hybrid is presented for online signature verification. The hybrid HMM/ANN model is constructed by using a type of time delayed Neural Networks as local probability estimators for an HMM. And a posterior probability of the model is worked out by the Viterbi algorithm, given an observation sequence. The proposed HMM/ANN hybrid has a strong discriminant ability i.e, from a local sense, the ANN can be regarded as an efficient classifier, and from a global sense, the posterior probability is consistent with that of a Bayes classifier. However, the topologies of the HMM employed for online signature verification are usually left-right, so that the corresponding ANN include a lot of redundant connections. We circumvent this issue by employing a group of ANN, but not a single ANN to predict the probability of each state on condition of the current observation and the previous state, and each ANN is corresponding to a state. So many redundant parameters of the model are cut down, and the training of the model may be more effective.(6).A cancelable biometrics system which employs a type of transformed signatures for identity verification is proposed. For biometrics system, a critical fact is that if the storage of biometrics data is compromised, the legitimate user has no means to switch to another new one. So the cancelable biometrics method was proposed recently and has been implemented on the face and fingerprint recognition. The cancelable biometrics method used a deliberately transformed biometrics data for authentication. As a result, the real biometrics data is hidden to anybody other than the owner. And more importantly, each user can hold many biometrics. For the system proposed in this thesis, each user needs to provide several signatures and a personal identity number (PIN) for registration. The PIN is used as the parameters for transformation of signatures, and then the transformed signatures are stored as reference samples. For verification, each user needs to provide a signature and a PIN. The signature is transformed with the PIN and then the dynamic time warping (DTW) algorithm is used to verify the transformed signatures. Compared with the experimental results of original signatures, we found that performance of verification based on the transformed signatures is rather promising and competing.