Kutsman V. Dynamic signature identification based on spiking neural network

The dissertation solves the scientific problem of dynamic signature identification methods and means using spiking neural networks and dynamic signature parameters, which are robust to intrapersonal and sensitive to interpersonal signature variabilities. It is shown that dynamic identification of signatures, in particular, has the following unsolved problems: first, the instability of signature reproduction by a person is a source of intrapersonal variability of its dynamic parameters, and secondly, imperfect methods of dynamic signature parameters classification. The dynamic signature identification method based on spiking neural network is developed It has a number of advantages over known methods, in particular: it does not require prior conversion of dynamic parameters into a vector of static features, it can identify predictive signatures, used neural network has a simplified learning procedure and does not require the entire network retraining when adding new signatures. The structure and architecture of the spiking neural network, focused on application in the process of dynamic signature identification, has been developed. The choice of dynamic parameters of the signature, which are resistant to geometric and temporal variability of signatures, is substantiated. The stability of dynamic signature parameters to intrapersonal variability, as well as the sensitivity of dynamic signature parameters to interpersonal variability has been studied. Specialized software has been developed to assess the accuracy of the proposed method of dynamic signature identification. Experimental studies of the developed method, conducted using the DeepSignDB signature database, showed that the proposed system is better in accuracy than the reference both in testing on master forgeries and when testing on random forgeries.