Research On Living Authentication Technology Based On Dynamic Expression And Broad Learning

Living authentication is based on biometric identity recognition.With the development of artificial intelligence,it is gradually replacing traditional identity authentication.It has been widely used in finance,security,military and other fields.However,there are endless attacks against the living authentication,which requires that the living authentication can effectively defend the corresponding attack means.For this reason,we expect that living authentication not only has better detection and authentication ability for living organisms,but also should have real-time online processing,less influence by lighting or photographic equipment,simple and friendly user interface and less user interaction.Therefore,safety,timeliness and customer physical examination are three important performance parameters of living certification.However,there are some defects in the current practical application,such as long authentication time,strong light dependence,poor user experience,expensive equipment and so on.On the other hand,with the development of facial expression research.Dynamic expression has attracted more attention because of its spontaneity and unforgeability.In addition,with the increasing of myopia,the proportion of people wearing glasses is increasing year by year.In simple blink recognition,the reflection of glasses can easily lead to the error of eye contour detection,resulting in recognition error.The dynamic facial expression is the change of facial features,which is the change of micro expression,normal expression and blink.Therefore,this thesis proposes to improve customer experience by using the judgment of facial expression changes to authenticate in vivo.At the same time,in order to prevent the authentication time from being too long,we introduce the breadth learning system and attention mechanism,and combine their advantages,and propose a live authentication technology based on dynamic expression breadth learning.The main research contents are as follows:1)In order to make feature extraction of face image more efficient,a feature extraction method based on broad learning system and attention mechanism is proposed in this thesis.Attention mechanism is added to the feature layer of the breadth learning system,which makes feature extraction combine the rapidity of the breadth learning system and the principle of effective information prominence of the attention mechanism.2)This thesis constructs a live authentication data set with dynamic expression.We selected the most effective and practical micro expression data set CASME II database and added attack data to it,including the images of holding photos and video playback.At the same time,we mixed the self collection data set,which was collected through the i Phone XS mobile phone.Fifteen 10s dynamic expression video segments with 5 targets,30 video segments with handheld i Pad for video playback,and 300 randomly selected images from the video frame images of each video segment to print and handheld photography.3)This thesis presents an living authentication algorithm model based on dynamic expression broad learning.The algorithm model extracts features through broad learning system combined with attention mechanism,and uses cyclic neural network mechanism to process time series,and recognizes dynamic expression to carry outliving authentication.4)Finally,the feasibility and superiority of the algorithm model based on dynamic expression width learning are verified by experiments.The specific method is to expand the popular data set and mix the self collected data set for experiment and analysis,and compared with CNN-RNN model.In this thesis,the living authentication model based on dynamic expression width learning has an average processing time of 0.124s and 0.2S for each sequence on the extended database of CASME II database and its hybrid database with self collection database,0.13%and 0.2%for photo attack,53.73%and 62.73%for video playback attack,respectively;the order of CNN-RNN model is compared The average processing time of columns was 0.482s and 0.523s respectively,and the HTER values were similar.As can be seen from the experimental results that the sequence processing time of this model is shorter than that of CNN-RNN,and there is no need to act according to the system instructions in the process of living authentication,so it can meet the certain security and timeliness of living authentication,and has the characteristics of non interactive and rapid verification.At the same time,the model has a good defense ability to the photo attack,but the model has a poor defense to the video playback attack.Finally,the results show that the new method can be effectively used in the live authentication,and has certain advantages.