Research On Online Signature Verification Algorithm

With the development of human society,intelligent technology plays an important role in human daily life.Therefore,many people are concerned about the personal information security.As an identity authentication with a long history,online signature verification has gradually attracted more and more attention.At present,there are three problems of online signature verification: first,the lack of Chinese signature samples in the open database is not conducive to the development of online signature verification and the lack of different signature modes makes it impossible to judge the impact of different collection devices and collection sections;second,the traditional DTW-based method is inefficient and is difficult to deploy in real scenario;third,the traditional description operators cannot generalize well on different signature databases due to its weak capacity.For the signature database,this paper proposed a multimodal Chinese online signature database called SCUT-MMSIG(Multimodal Signature Database at South China University of Technology),which includes three modes: mobile phone,tablet and in-air signature.SCUT-MMSIG contains 6000 online signature samples from 50 users and each user contains 20 genuine signatures and 20 forged signatures.In order to study the impact of time-shifting,the data collection was divided into two sections with an interval of one month.For online signature verification algorithm,this paper proposed a MDB-DTW algorithm(Multi-Distance Based DTW)and a two stream deep neural network.The former aims to reduce the amount of feature description operators and extract features by fusing the information between 7 different distance measurements.The experimental results further proved the robustness of the proposed MDB-DTW algorithm.The latter employs the deep learning-based methods to automatically extract both pattern and temporal features for online signature verification.The first stream takes signature verification as a text matching problem,whose input is a scoring matrix representing the similarities between two signatures.Then a convolutional neural network is utilized to capture rich matching patterns layer-by-layer.The second stream uses a bidirectional Long Short-Term Memory network with self-attention mechanism to extract temporal features,which is critical for different users.With the ablation study and model visualization,we found that the extracted pattern and temporal features are complementary to each other,which enhance the final verification performance.Finally,by analyzing the signatures of different modes and different sections in SCUT-MMSIG,we found that the limitation of the writing area and time-shifting have a great impact on the behavior habits of users,thus reducing the accuracy of the online signature verification system.In addition,experimental results on SCUT-MMSIG,MCYT-100,SUSIG,Biosecure ID and MOBISIG demonstrate that MDB-DTW can further enhance the verification performance of traditional methods and the proposed two stream deep neural network is capable of achieving state-of-the-art results compared with other methods.The SCUT-MMSIG database and the proposed methods will be released at Github for further study.