The Application Of Bayes Ying Yang Machine Based Gaussian Mixture Models In On-line Handwritten Signature Verification

With the rise of e-commerce and remote-access banking, there has been a lot of interest in authentication solutions for securing transactions. Knowledge-based or physical-based access tokens suffer from various shortcomings as they can be forgotten, stolen, or duplicated. Biometrics-based access tokens, on the other hand, promise easier interactions and higher levels of security to the end-user.The on-line signature verification methods usually take into account dynamic characteristics such as pressure, tilts, position velocity, etc. These signals are acquired when the signature is being made. Due to this, and the higher quantity of information available on on-line systems, generally these automatic verification systems get higher reliability than off-line ones.Signatures are particularly useful for identification because each person’s signature is highly unique, especially if the dynamic properties of the signatures are considered in addition to the static shape of the signature. Even if skilled forgers can accurately reproduce the shape of signatures, it is unlikely that they can simultaneously reproduce the dynamic properties as well.A lot of research concerning on-line and off-line signature verification has taken place. Numerous methods such as Dynamic Time Warping(DTW), hidden Markov models, fuzzy logic inference, neural networks have been used for on-line signature verification. Currently the statistical method hidden Markov models(HMMs) exhibit state-of-the-art performance. Though the Gaussian mixture models(GMMs) has been proven useful in speaker verification, few research apply it to the handwritten signature verification.In the work, we used a Bayes Ying Yang (BYY) machine based GMMs algorithm to build signature models for the users. This theory is further shown to function as a general theory for supervised and unsupervised learning, too, from which new theories for supervised classification and regression are obtained, which shows that the existing approaches for multilayer net, mixture-of-experts, and radial basis function nets are unified as special cases, with not only new insights and new learning algorithm but also new selection criteria for the number of hidden units and experts. Besides, this theory is further shown to function as a general theory for learning on time series also, not only with the hidden Markov model and the linear state space based Kalman filter as special case, but also with several temporal learning models and algorithms obtained. The BYY theory is based on the Chinese acient philosophy that "every entity in the universe includes the interaction of Ying and Yang". Baed on this, the harmony function is proposed in the BYY theory. By iteratively optimizing this function can push the two parts represented by Ying and Yang to a best match. So when the two domains in the GMMs, represental domain and observation domain, are represented by the Ying and Yang in BYY, then the harmony function can be used for parameter learning. Based on this thougt, this work creatively apply the BYY harmony based GMMs in the on-line handwritten signature verification. As the harmony function in BYY can automatically choose the optimal component number during the GMMs parameters learning, comparing with the traditional Expectation Maximization(EM), the BYY based GMMs can achieve a higher model accuracy. And the following experiment as well as the comparison results also prove that in the application of on-line handwritten signature verification, the BYY based GMMs is qualified to get a better recognition rate.