SSVEP Signal For Identity Recognition Using Convolutional Neural Networks

With the development of digitalization,how to identify individuals securely has become a hot topic in society.Biometric recognition has gradually come into people's life with portability and uniqueness.Some traditional biometrics such as fingerprints,voice,human face,and iris have begun to be widely used in various fields of society.However,the above features have defects such as being easy to be stolen and forged,and it is difficult to meet the increasing security requirements.As a new type of biological characteristics,EEG signal has the characteristics of traditional biological characteristics,but also has the advantages of live detection,anti-counterfeiting,and high concealment.At the end of the last century,researchers began to use EEG signal in the field of identification.But so far,the research on EEG signal for Identity recognition is still in its infancy,and efforts are still needed in experimental paradigm design,feature extraction,and portability.This paper focuses on the application of Deep Learning methods in the identification of EEG signals.Conducted in-depth research in experimental paradigm designs,data preprocessing,EEG data organization,Convolutional Neural Network(CNN)improvement,and the persistence of EEG in identity recognition.A Steady State Visual Evoked Potential(SSVEP)signal identification system based on Deep Learning with intrusion identification function was designed and implemented.The main content and contributions to the paper are as follows.First,based on related papers on the study of signals and the mechanism of human vision formation,three different-color flicker-induced experimental paradigms were designed.This paradigm can make individual signals more comfortable while inducing a signal with a high signal-to-noise ratio.Secondly,according to the characteristics of each electrode's EEG signals from different brain regions,four different electrode selection modes were designed to effectively reduce the redundancy of EEG signals through experiments.At the same time,according to the input requirements of the convolutional neural network,the EEG signals are effectively organized to form an electrode-sampling point data matrix.Third,Deep Conv Net and Shallow Conv Net are improved.The augmented data set used dense cropping,while sliding windows training was used to reduce the amount of data while increasing the data.By adding a penalty function to modify the parameter optimization function,the network pays more attention to the stable features co-existing between adjacent samples and penalizes the differences between samples during the training phase,thereby extracting the common features of the EEG signals of the clipped samples to improve the accuracy of the algorithm.Subsequent design of multiple classifier thresholds for intruder identification.Fourth,the persistence of SSVEP signal in identity recognition is studied.The experimental results show that the accuracy of the data collected in the same batch reaches 100%.For the multiple flicker frequencies of 6-12 Hz,the frequency of the induced EEG signal is better in this system.The color showed the best trend of blue,followed by yellow,and last of red.For different batches of data collected two weeks apart,the accuracy of the multi-class deep networks designed in this paper reaches 91.85%.Except that some individuals may be affected by their own factors,the steady-state visual-evoked EEG signals of other system registrants have certain persistence.