Biometric Authentication and Correlation Analysis Based on CNN-SRU Hybrid Neural Network Model

With the continuous development of computer technology, many institutions in society have higher requirements for the efficiency and reliability of identification systems. In sectors with a high-security level, the use of traditional key and smart card system has been replaced by the identification system of biometric technology. The use of fingerprint and face recognition in biometric technology is a biometric technology that does not constitute an infringement on the human body and is convenient and reliable. The biometric technology has been continuously improved, and the existing biometric technologies are based on unimodal biometric features. The unimodal biometric technology has its own limitations such as proposing single information and checking data affected by the environment, which makes it difficult for the technology to play its advantages in practical applications. In this paper, we use CNN-SRU deep learning to preprocess a large amount of complex data in the perceptual layer. The data collected in the perceptual layer are first transmitted to CNN convolutional neural network for simple classification and analysis and then arrives at the LSTM session to update again and optimize the screening to improve the biometric performance. The results show that the CNN-LSTM, CNN-GRU, and CNN algorithms show a decreasing trend in accuracy under the three error evaluation criteria of RMSE, MAE, and ME, from 0.35 to 0.07, 0.58 to 0.19, and 0.38 to 0.15, respectively. The recognition rate of multifeature fusion can reach 95.2%; the recognition efficiency of the multibiometric authentication system and accuracy rate has been significantly improved. It provides a strong guarantee for the regional standardization, high integration, generalization, and modularization of multibiometric identification system application products.


Introduction
Te rapid development of computer information technology has put forward higher requirements for the update and processing of information technology [1,2]. In the feld of system security protection, relying solely on hardware devices has been difcult to meet customer needs. Te traditional key-based authentication method becomes more and more unreliable and stable due to the large amount of data, the rapid spread, and the simple replication. Terefore, according to customer needs, we provide an overall solution and provide a multibiometric authentication platform to meet their actual needs. Biometric technology is a technology for automatic identifcation by collecting the physiological characteristics and biological characteristics of the human body [3]. Human biometrics are relatively stable, and there are no problems such as loss and forgetting. Compared with the traditional cryptographic key authentication method [4][5][6][7], it has more efective system security. Te biometric identifcation method uses the human body marking method, which is easy to carry with you and can realize real-time collection, identifcation, and judgment of biometric information. Generally speaking, human biometrics have seven characteristics, including universality, uniqueness, permanence, collectability, acceptability, safety, and performance requirements [8][9][10]; biometrics do not change with time and the environment. Large changes can be quantitatively collected and analyzed, the collected feature information can meet the needs of users, and the features are not easy to be imitated or forged [11,12]. Te bottom line is that everyone is biometric and diferent.
Biometric technology has the characteristics of good anticounterfeiting performance, high security, and reliability and has become the most widely used security and identity authentication technology [13,14]. Biometric technology relies heavily on physiological or behavioral features, and common face recognition [15,16], language control [17], and fngerprint recognition [16,18] can be used as biometrics. At present, the application of the above-mentioned biometric technology has made great progress, and the related scientifc research at home and abroad has become more and more extensive. We are conducting a visual analysis of biometrics related to Chinese and English documents in the CNKI database, as shown in Figure 1. We found that the number of studies on biometrics gradually increased from 2000, and after 2005, the number of articles published each year was more than 1000, and in recent years, it reached more than 2000; meanwhile, in the study of the distribution of keywords and topics co-occurring in the articles, we could fnd that the keywords with the highest frequency were mainly biological characteristics, as shown in Figure 1. Face recognition, fngerprint recognition, algorithm research, recognition technology, etc., show that the algorithm research has gradually deepened and achieved extraordinary results in the research of biometrics. Te study [19] focused their work on the application of machine learning to biometric speech, and they focused on the same concept of examining the recognition accuracy of the machine learning algorithm REPTree on a selected speech biometric dataset that was deployed and evaluated with the mining tool WEKA. Te goal was to achieve a percentage greater than or equal to 95 in order to accurately classify the given sample data. In terms of processing other biometric information, a systematic study and comparison in ear biometrics has been conducted by [20]. Tey proposed another idea to perceive the ear in an online irregular orientation. Tis work can touch on the ftting of inwardly curved ears and the extraction of highlights from the internal drills of the ear edges. Here, the watchful edge recognition computation is used to locate the edges of the ear. Te ear images are processed by the shape follower computation and then the rough images with edges are found as the yield of the frame. Finally, each of the three component vectors is decontrasted and compared with an alternate library of ear pictures and tracked for a specifc match. In [21], they mainly used algorithms to enhance the security of face recognition techniques. In their paper, they proposed a fast and robust fuzzy C-mean clustering (FRFCM) algorithm and face recognition optically selective encryption scheme for biometric medical images. In the proposed scheme, a new selection method for obtaining regions of interest (ROI) based on the FRFCM algorithm is proposed. Te security and robustness of the proposed cryptosystem is also verifed by numerical simulations. Te study [22] systematically analyzes the key points and difculties of biometric techniques in spectral imaging. Te reliability of conventional face recognition systems working in the visible range will be afected by light variations, pose variations, and spoofng attacks. Tere are no large databases available for benchmark evaluation. Existing databases do not capture the same test subjects on all possible frequency bands for which experimental evaluations have been carried out. Also, this has been limited so far due to the small number of test objects and their images in the existing database. Deep learning based methods require a large number of parameters for training. Tis leads to overftting due to the small number of such samples in the existing databases. Te study [12] was conducted to remotely determine current human biological parameters by algorithmically processing infrared images of human faces. Te problems that hinder the widespread use of remote biometric algorithms in practice are highlighted. Te urgency of creating a metrological database containing video recordings of face images in the infrared range to assess the efciency of biometric algorithms, timesynchronized records of human biometric parameters, and information on the complexity of the test tasks performed was confrmed. Te structure of the laboratory system used to acquire complex data of the tested person is considered. A typical data structure for a single test cycle is given as shown in Figures 1 and 2.
Trough the analysis and summary of the above-given literature research work, it can be seen that the development and application optimization of deep learning in biometrics has carried out systematic excellence. However, the research work on multibiometrics technology is still at an immature stage, and the evaluation standards are not uniform. Secondly, multibiometrics authentication also increases the burden on users. Not perfect in multibiometric fusion. Terefore, this paper studies the multibiometric recognition system through the fusion of facial features, speech recognition, and fngerprint recognition. In addition, we introduce three algorithms of CNN, LSTM, and GRU to extract and fuse biometric parameters. Optimization studies are conducted. In view of the problems such as the continuous increase of biometric data, the untimely update of features, and the resulting system lag, we optimize the algorithm to improve the system update rate and improve the security and stability of the biometric identifcation system. In addition, we analyze and study the collected 2000 data, and verify the robustness and accuracy of the algorithm through simulation. Biometric identifcation technology, referred to as biometrics, is the identifcation and authentication of personal identity by acquiring biometric characteristics unique to human beings. We classify these inherent biometric characteristics into two major categories, namely, physiological characteristics and behavioral characteristics. Physiological characteristics are those that are innate and would not have led to changes in the absence of special factors. Typical examples are fngerprints, iris, DNA, etc. Te other behavioral traits are mainly acquired habits. Typically, they are the voice of the person speaking, the font of the signature, etc., [23,24].
All biometrics include the following processes: acquisition, decoding, comparison, and matching. Te basic process is shown in Figure 3.
Biometric identifcation technology is mainly through the detection of the physical characteristics that the human body has always had so as to carry out identity confrmation, not all the characteristics of the human body can be used as the collection point for identifcation, and the characteristics that need to meet the following conditions can be used as the identifcation target: Most of the human body's biological information has the characteristic of uniqueness, uniqueness mainly for the comparison between individual people, absolutely unique. In the case of fngerprints, the texture details are not the same between each individual. (2) Stability Te biological information of a human individual always remains the same from birth to death, but of course, if the feature is changed or damaged by human or external factors, it will change. For this reason, it can be used to identify individuals.
(3) Identifability Te biometric characteristics of human beings are very diferent from each other, and by using certain computer algorithms, it is possible to distinguish this information and identify the key elements of the characteristics, making it possible to use the identifcation technology in a wide area. (4) Capturability With the development of computer technology and related hardware technology, the image quality and pixels of the feature image information collected by the instrument are also improved, making the image a convenient carrier for recognition.

The Feasibility of CNN Network in Biometric Authentication
With the rapid development of science and technology, more and more information is obtained and authorized remotely through online, and the required resources can be legally accessed and the right of application is a necessary condition for the security of the internal data of the communication system. In addition, in the network environment, how to identify and verify the reliability of biometric information, safely realize remote authentication, and ensure the privacy of biometrics from illegal use has become the main problem to be solved in this paper. Terefore, this paper proposes a 2D CNN network to track the lines in electronic data and implement vectorization to realize the recognition processing. 2D CNN can not only extract features from a large number of graphs, data reconstruction, and other preprocessing methods. At the same time, each computing layer itself has the same weight for sharing links, which greatly reduces the computing cost and computing time and avoids too many parameters afecting the clarity of text/graphics. In addition, LSTM can easily and accurately capture the main information of the sequence from pictures to speech because of its linear structure. Strive to propose a CNN-LSTM network based on deep learning biometric authentication to provide an efective solution quickly, conveniently, and safely. Figure 4 presents a generalized framework for digital image source forensics under the CNN model theory. In the image preprocessing, the image to be detected is frst cut into image blocks (P k in Figure 4(a) indicates the k th image block), and then the image fngerprint characterizing the source of the shot is extracted using CNN (image feature extraction in Figure 4(b)), and the detection result Y k of each image block is output (Y k in Figure 4(c) indicates the feature extractor predicts the label for the k th image block), and the majority voting algorithm is used to fuse the detection results of the k th image block and output the image level prediction results, i.e., device model multiclassifcation identifcation.  Computational Intelligence and Neuroscience

Learning Algorithm of CNN Network.
Te training process of the CNN network is mainly divided into forward propagation and back propagation [25]. First, by inputting data to the convolution layer, the feature extraction is performed on the convolution operation based on the flter and the convolution kernel (Kernel), the feature map is obtained, and the bias term is added to it, and then the activation function (ReLu, Tanh, sigmoid, and softmax) to calculate the output of the convolutional layer; the pooling layer samples the data processed by the receiving convolutional layer, and then converts the feature map into a vector by stretching and sums the weighted biases. Finally, the class probability output is obtained through the activation function, and the calculation is repeated until the loss function is the minimum value [26].
(1) Calculation of the convolutional layer. Te convolution formula is shown in equation (1), the input x is weighted w, and the bias b is added, and fnally the total sum obtained is output through the nonlinear activation function f.
Among them, D and F are the number of flters and the size of the convolution kernel, respectively. (2) Calculation of pooling layer and full connection. Te pooling layer retains the main efective information by locally sampling the feature map and reduces the infuence of unnecessary data on the calculation result. In addition, as long as there is a relative relationship with the main information of the feature map, no matter whether the image is scaled, distorted, translated, etc., the accuracy of the result cannot be afected. After the data is passed through the pooling layer, a large number of parameters can be fltered out, which can improve the training accuracy and reduce the error. Te average sampling method sampled in this paper samples the feature map, and in the traversed region, the average value is selected as the new feature of the region. Due to the characteristics of CNN itself, under the action of the fully connected layer, the output of the convolutional layer is weighted w ofset summation, and then output through the activation function f, as shown in the following equation:  ...

Computational Intelligence and Neuroscience
(3) Calculation of softmax output layer. Te activation function nonlinearizes the total number of weighted bias sums to solve the multiclass problem, and its calculation formula is as follows: In the formula, n is the number of inputs. (4) Back propagation (1)-(3) are forward propagation, and the error is calculated by the loss function for reverse propagation. Te process is divided into three main parts: (a) Calculate the network error from the predicted and actual results: Among them, y predict is the prediction data, y real is the experimental data. (b) Te calculation error is passed in the reverse direction. Te specifc propagation formula of CNN is as follows: Fully connected layer.
Among them, l is the current layer and σ(a i,l ) is the activation function. (c) Te fnal goal of reverse transfer is to update the weight w and bias b, and the specifc calculation is as follows. Weight update in the fully connected layer: Fully connected layer bias update.
Weight update in the fully connected layer.
Bias update in fully connected layer: Among them, α i,l− 1 is the output of the ith neuron in the l -1th layer.
In addition, this paper chooses the cross entropy loss function as the loss function, and the formula is as follows: Among them, y c (i) is the predicted value, y(i) is the real value, and N is the number of samples.

LSTM Network.
A single-channelLSTM-based method for analyzing factors related to youth physical activity behavior mainly includes: LSTM neural networks that use memory units to avoid gradient disappearance and gradient explosion during backpropagation and can learn long-term dependencies and make full use of historical information. Te LSTM was improved and extended in 2013 by [27,28], making it widely used in natural language processing, speech recognition, and other felds.
As shown in Figure 5, the LSTM unit has a memory unit c for saving historical information. Te updating and utilization of the history information is controlled by three gates: input gate i, forget gate f, and output gate o. Te updating process of the LSTM unit at time t is as follows: where x i is the input data of the memory unit, σ is the logistic sigmoid function, the symbol ⊙ is the dot product operation between vectors, and W i , W f, W c , W o , U i , U f , U c , U o , V o is the weight matrix. i t , o t , f t , c t are the values of the input gate, output gate, forgetting gate, and memory cell at time t, respectively, c t are the values of the candidate memory states of the memory cell, and h t are the outputs of the LSTM cell at time t.
For the analysis of the factors associated with youth physical activity behavior, we frst used the random undersampling method to obtain a balanced sample of the factors associated with each youth physical activity behavior and then used a single-channel LSTM neural network as the classifcation method. Figure 6 shows the framework of the single-channel LSTM neural network classifer, which has only one LSTM layer. Te frst dashed box shows the internal structure of the single-channel LSTM model, and the second dashed box shows the process of unbalanced samples. Te input to the LSTM model is a word vector representation of the training samples, which has good semantic features and is a common way to represent word features. Te input feature vectors are passed through the LSTM layer to obtain high-dimensional vectors, which can learn deeper features that can better describe the samples. Te fully-connected Computational Intelligence and Neuroscience layer is similar to the hidden layer of a traditional multilayer perceptron, receiving all the outputs from the previous layer, weighting and summing these output vectors, and propagating the weighted outputs through the excitation function to the dropout layer. In this experiment, the layer uses ReLu as the excitation function, which reduces the interdependence between parameters and is closer to the biological activation model, and the excitation function is shown in (14).
where x is the output vector and the ReLu function sets all values less than 0 to 0, with the ability to bootstrap moderate sparsity. Te dropout layer randomly leaves some hidden layer nodes in the network inactive during training and prediction, reducing the number of features and efectively preventing the network from overftting. Te dropout layer appears as a hidden layer in the LST M neural network model, as shown in the following equation: where D denotes the dropout operator and p is an adjustable superparameter (the ratio of retained hidden layer cells). Finally, the output of the single-channel LST M model is used to classify the samples by the Softmax output layer. We choose the category with the highest posterior probability as the prediction label, as shown in the following equation: where x is the upper layer output vector, i is the label prediction, W, U, V are the coefcient matrices in the LSTM update method, and label pred is the predicted label with the highest posterior probability.

Normalization Processing.
Since the magnitude of data input by diferent channels is very diferent, it will afect the increase of training error. Terefore, we frst need to normalize the data to normalize all inputs to the same interval.
In this paper, we choose max-min normalization for normalization, and the calculation method is as follows: Among them, x max is the maximum value in the data set, x min is the minimum value in the data set. After normalization, the training error will not increase due to the order of magnitude between the data.

Determination of Model Parameters.
Te main infuencing factors of biometric vectorization are physiological characteristics and behavioral characteristics. Physiological features include fngerprints, facial images, and    Computational Intelligence and Neuroscience irises which are identifed according to each individual's unique biological features; behavioral features are gait, voice, handwriting, etc., which are also a method of identifying the appraiser. For example, when identifying a person's handwriting, the outline of the font is clear and distinct, and the efect of the network calculation on displaying the vector diagram is more obvious; the more the number of words, the more data the network can obtain, which makes the network output results have a certain robustness. Specifc steps are as follows: (1) Te two categories of parameters, physiological characteristics, and behavioral characteristics, are used as the input nodes of the prediction model, and predictions are used as the output nodes of the prediction model. (2) Te hidden layer plays a key role in the network architecture. Te number of flters in the convolutional layer in this model 2D CNN is 2, the size of the convolution kernel is 16, the stride is 1, and the padding is 2. Te LSTM layer nodes are 4, and the fully connected layer nodes are 3. (3) In order to adjust the appropriate learning rate parameters and avoid overftting or underftting, it is necessary to continuously test and adjust. Te Adam algorithm model is used, and the learning rate parameter is fnally selected as 0.0014 and the decay rate is 0.08 [29]. (4) After data preprocessing, the data can be used as an input layer node to output the results.

Several Multibiometric Systems.
We can use diferent feature points, sensors, and feature extraction quantities and methods, multibiometric systems can have the following combinations: (1) Unimodal biometric, a combination of multiple sensors, where the biometric features of the same target are acquired by multiple sensors, so that the sample data can be acquired twice. By acquiring face feature data in this way, this data is combined in the data layer and the matching layer, which can effectively improve the recognition rate of the face recognition system. (2) Te combination of unimodal biometric features, multiple classifers, and one sensor is diferent from unimodal biometric features in that biometric data is acquired by one sensor and these sample data are processed by multiple classifers. Te respective defned features are generated. Matching at the logical layer can improve the recognition rate.
(3) Unimodal biometric features, a combination of multiple categories, in the case of iris and fngerprint features, extracts two or more biometric information from the target person. Tis combination does not require multiple sensors to acquire information and is not particularly demanding for multiple feature extraction and matching models. (4) Combination of multiple biometric features, the combination is to use multiple biometric features for recognition and obtain diferent feature data by diferent sensors. Due to the relative independence of the biometric features and thus the accuracy of the recognition system is greatly improved.

Comparative Analysis of Recognition Accuracy of Biometric Results.
In order to verify that the CNN-LSTM network has better performance, this paper compares and analyzes CNN, CNN-GRU, and CNN-LSTM. In order to further analyze the performance of these three networks, we choose root mean square error as the evaluation index: root mean square error (RMSE): measure the error between the observed value and the actual value which is calculated as follows: Specifcally, compared with the actual results, the RMSE of the CNN-LSTM algorithm is smaller than that of the CNN-GRU algorithm and CNN algorithm, which are 0.07, 0.17, and 0.35, respectively, and the CNN-LSTM algorithm is improved by 59% and 80%, as shown in Figure 7. Using the CNN-LSTM recognition technology with fast acquisition speed and high efciency of feature comparison for efective recognition has great potential value for future biometric optimization, as shown in Figure 7.
In addition, the error of the CNN-LSTM algorithm is smaller than that of the CNN-GRU algorithm and CNN algorithm, which are 0.19, 0.31, and 0.58, respectively, and the CNN-LSTM algorithm improves 38% and 57% compared with the actual results of physiological features, as shown in Figure 8.
Finally, compared with the actual results, the CNN-LSTM algorithm has smaller errors than the CNN-GRU algorithm and CNN algorithm in terms of physiological and behavioral features of speech, which are 0.15, 0.22, and 0.38, respectively, and the improvement of the CNN-LSTM algorithm is 32% and 60%, respectively, Computational Intelligence and Neuroscience illustrating the excellent performance of CNN-LSTM deep learning in biometric recognition as shown in Figure 9.
In the biometric system, only the sound feature is used in long-term applications, and the biometric recognition accuracy can reach up to 67.7%, the fngerprint recognition accuracy can reach up to 82.1%, and the face recognition accuracy can reach up to 84.5%. Te multifeature fusion recognition can reach up to 95.2%. Tis shows that the algorithm can better match the actual needs of biometric identifcation. It can greatly improve the recognition accuracy and improve the recognition efciency as shown in Figure 10 and Table 1.

Conclusion
In this paper, we give a basic description of the concept of multicommunication framework and neural network algorithm and introduce the structure and calculation process of SRU algorithm for wireless communication receiving module. And, we compared the three algorithms of SRU, GRU, and LSTM, and the hidden layer, number of nodes, number of iterations, learning rate, and activation function in the network structure are all the same. the result shows the following theories:    , which have a great improvement, which shows that the algorithm can better match the actual needs of biometrics. It provides a strong guarantee for the regional standardization, high integration, generalization, and modularization of multibiometrics system application products.
At present, the system has been applied in a large number of customers, but after the actual use, the system still reveals some problems, and recognition technology still needs to be improved; these will be the focus of later work. With the biometric identifcation technology application in the feld is deepening and the scope is expanding, many enterprises are investing more and more in this technology. Te updated speed of biometric identifcation technology will also be accelerated. For the application software of biometrics technology, if it needs to have a lasting vitality, it needs to keep up with the advanced technology. Due to the development period and resources, this system still has many unsatisfactory points and needs to be improved later.

Data Availability
Te experimental data used to support the fndings of this study are available from the corresponding author upon request.