Identity authentication method for remote medical treatment by fusing multiple biological characteristics

Abstract

The invention discloses an identity authentication method for remote medical treatment by fusing multiple biological characteristics. The method comprises the following steps: (1) data acquisition: firstly, acquiring a fingerprint, a face image and a personal signature key of an expert; (2) feature extraction: respectively inputting the collected biological features into corresponding branch models, namely, respectively extracting respective features of a fingerprint branch model, a face branch model and a signature key branch model; (3) feature fusion: designing a multiple cross attention structure, weighting feature information output by different branch models through the cross attention structure, transmitting the weighted feature information to other branches, realizing fusion and complementation of the feature information, and outputting an identity recognition result; (4) identity matching: if the identity recognition result is matched with the preset expert identity, allowing the expert to carry out remote inquiry; otherwise, refusing access and giving an alarm. The method is used for ensuring the effectiveness and safety of telemedicine services, and a safe and accurate identity authentication method is provided for the field of telemedicine.

Description

Technical Field

[0001] The present invention belongs to the technical field of telemedicine, and relates to identity authentication based on multiple biometrics in a telemedicine process, and specifically is an identity authentication method for telemedicine that integrates multiple biometrics. Background Art

[0002] Telemedicine uses the Internet technology to enable experts to transcend geographical restrictions and provide high-quality medical services to patients in areas with scarce medical resources. However, medical behavior not only has legal effects, but also directly affects the life safety of patients. Therefore, it is crucial to ensure the authenticity of the expert's identity and the confidentiality of medical data during telemedicine.

[0003] At present, telemedicine identity authentication technology mainly relies on the traditional username and password verification method, which has defects such as easy theft and insufficient security. In order to improve security, some systems have introduced single biometric recognition technology, such as fingerprint recognition or facial recognition. However, traditional single biometric recognition technology has defects such as easy forgery, easy impersonation, and high misidentification rate. Therefore, a safer and more accurate identity authentication method is needed to ensure the quality and data security of telemedicine services. Summary of the invention

[0004] In view of the security and reliability issues of existing telemedicine identity authentication technologies, the present invention aims to provide a method for telemedicine identity authentication by integrating multiple biometric features. The method integrates multiple biometric features such as fingerprints, faces, and personal signature keys by constructing a multimodal model based on multiple biometric recognition technologies to improve the accuracy and security of identity authentication. At the same time, the present invention also optimizes model performance and improves the efficiency and accuracy of identity recognition by designing a three-branch model and a cross-attention structure.

[0005] The present invention is achieved by adopting the following technical solutions:

[0006] A method for identity authentication for telemedicine by integrating multiple biometric features comprises the following steps:

[0007] (1) Data collection: First, collect the expert’s fingerprint, facial image, and personal signature key;

[0008] (2) Feature extraction: The collected biometric features are input into the corresponding branch models, namely the fingerprint branch model, the facial branch model, and the signature key branch model, to extract their respective features;

[0009] (3) Feature fusion: Design a cross-attention structure, through which the feature information output by different branch models is weighted and passed to other branches, achieving the fusion and complementarity of feature information and outputting an identity recognition result;

[0010] (4) Identity matching: If the identity recognition result matches the preset expert identity, the expert is allowed to conduct remote consultation; otherwise, access is denied and an alarm is issued.

[0011] The technical solution provided by the present invention has the following advantages compared with the prior art:

[0012] 1. Improved security and reliability

[0013] (1) Fusion of multiple biometric features: This invention integrates three biometric features: fingerprint, facial recognition and personal signature key, which greatly improves the accuracy and security of identity authentication compared to single biometric recognition technology. The multiple verification mechanism reduces the risk of a single feature being forged or counterfeited.

[0014] (2) Multimodal model: By constructing a multimodal model consisting of three branches, different biological features are processed separately, so that each feature can be processed in a specially optimized manner, thereby enhancing the overall performance and robustness of the model.

[0015] 2. Technological innovation

[0016] (1) Application of deep learning algorithms: Each branch model adopts a deep learning algorithm, which can automatically learn and extract features from large amounts of data, thereby improving the efficiency and accuracy of feature extraction.

[0017] (2) Residual structure and convolutional neural network (CNN): In the fingerprint branch model, the residual structure is used to reduce the feature disappearance problem and effectively extract the texture and detail features of the fingerprint. In the facial branch model, the improved FaceNet model is used to extract features from facial images and extract the key feature points of the face.

[0018] (3) Three-branch model and multiple cross-attention structure: The present invention designs a three-branch model and a double cross-attention structure, which optimizes the model performance and improves the efficiency and accuracy of identity recognition. The cross-attention mechanism is an innovative information fusion technology that can enhance the information interaction between different biometric features.

[0019] (4) Encryption of personal signature key: In the signature key branch, the personal signature key is encrypted and its hash value is extracted as a feature. This processing method improves the security of the personal signature key and the accuracy of identification.

[0020] 3. Practicality and operability

[0021] (1) The specific embodiment of the present invention describes in detail the specific structure and implementation steps of the model, including the parameter settings of each convolutional layer and fully connected layer, which provides clear guidance for practical applications.

[0022] (2) The embodiment of the present invention uses the ImageNet dataset for pre-training and uses expert identity data for fine-tuning. This training method can ensure that the model has good performance in practical applications.

[0023] The present invention is reasonably designed and shows obvious advantages in technical solutions, model design, security improvement and practicality. It is used to ensure the effectiveness and security of telemedicine services and is expected to provide a safer and more accurate identity authentication method for the telemedicine field, which has great practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0026] Figure 1 Represents the structure diagram of the fingerprint branch model.

[0027] Figure 2 A diagram showing the principle of identity authentication by integrating multiple biometric features according to the present invention. DETAILED DESCRIPTION

[0028] In order to more clearly understand the above-mentioned objectives, features and advantages of the present invention, the scheme of the present invention will be further described below. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0029] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein; it is obvious that the embodiments in the specification are only part of the embodiments of the present invention, rather than all of the embodiments.

[0030] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0031] During the telemedicine process, when an expert logs into the system, the system first collects his fingerprint, facial image, and personal signature key. These biometric features are then input into the multimodal model, which outputs an identity recognition result. If the recognition result matches the preset expert identity, the expert is allowed to conduct a remote consultation; otherwise, access is denied and an alarm is issued. The implementation process is as follows:

[0032] A method for identity authentication for telemedicine by integrating multiple biometric features comprises the following steps:

[0033] (1) Data collection: The system first collects the expert’s fingerprint, facial image, and personal signature key.

[0034] (2) Feature extraction: The collected biometric features are input into the corresponding branch models, namely the fingerprint branch model, the facial branch model and the signature key branch model, and their respective features are extracted.

[0035] (3) Feature fusion: Based on the three-branch model, a cross-attention structure is designed. Through the cross-attention structure, the feature information output by different branch models is weighted and passed to other branches to achieve the fusion and complementarity of feature information and output an identity recognition result.

[0036] (4) Identity matching: If the identity recognition result matches the preset expert identity, the expert is allowed to conduct remote consultation; otherwise, access is denied and an alarm is issued.

[0037] The present invention first constructs a multimodal model, which includes three branch models, which are used to process three biometric features: fingerprint, face and personal signature key. Each branch model uses a deep learning algorithm to extract effective information of its own biometric feature.

[0038] First, fingerprint branch model: the improved backbone network of the fingerprint recognition model FingerNet is used to extract features of fingerprint images, extract the texture and detail features of fingerprints, and the specific model structure is as follows: Figure 1 As shown in the figure. This model is an improvement on the fingerprint recognition model FingerNet. It uses a residual structure to reduce the problem of feature disappearance and effectively extract the texture and detail features of the fingerprint. The specific structure of the model is as follows:

[0039] Input layer: receives fingerprint image;

[0040] Feature extraction layer: extracts the features of the fingerprint image through a series of convolutional layers (conv) and pooling layers (pool). Specifically include:

[0041] The first convolutional layer (conv1):

[0042] Input size: fingerprint image.

[0043] Output size: the size after convolution and pooling with a 7×7×64 convolution kernel and a stride of 2.

[0044] The second convolutional layer (conv2):

[0045] Input size: Output of the first pooling layer.

[0046] Output size: the size after convolution with a 3×3×64 convolution kernel and a stride of 2.

[0047] Residual connection: The input of the second convolutional layer is directly connected to the output of the second convolutional layer, forming a residual connection.

[0048] The third convolutional layer (conv3):

[0049] Input size: Output of the second convolutional layer (including residual connections).

[0050] Output size: the size after convolution with a 3×3×64 convolution kernel.

[0051] Residual connection: The input of the third convolutional layer is directly connected to the output of the third convolutional layer.

[0052] The fourth convolutional layer (conv4):

[0053] Input size: Output of the third convolutional layer (including residual connections).

[0054] Output size: the size after convolution with a 3×3×64 convolution kernel.

[0055] Residual connection: The input of the fourth convolutional layer is directly connected to the output of the fourth convolutional layer.

[0056] The fifth convolutional layer (conv5):

[0057] Input size: Output of the fourth convolutional layer (including residual connections).

[0058] Output size: the size after convolution with a 3×3×128 convolution kernel.

[0059] Residual connection: The input of the fifth convolutional layer is directly connected to the output of the fifth convolutional layer.

[0060] The sixth convolutional layer (conv6):

[0061] Input size: Output of the fifth convolutional layer (including residual connections).

[0062] Output size: the size after convolution with a 3×3×128 convolution kernel and a stride of 2.

[0063] Residual connection: The input of the sixth convolutional layer is directly connected to the output of the sixth convolutional layer.

[0064] The seventh convolutional layer (conv7):

[0065] Input size: Output of the sixth convolutional layer (including residual connections).

[0066] Output size: the size after convolution with a 3×3×128 convolution kernel.

[0067] Residual connection: The input of the seventh convolutional layer is directly connected to the output of the seventh convolutional layer.

[0068] The eighth convolutional layer (conv8):

[0069] Input size: Output of the seventh convolutional layer (including residual connections).

[0070] Output size: the size after convolution with a 3×3×128 convolution kernel and a stride of 2.

[0071] Residual connection: The input of the eighth convolutional layer is directly connected to the output of the eighth convolutional layer.

[0072] The ninth convolutional layer (conv9):

[0073] Input size: Output of the eighth convolutional layer (including residual connections).

[0074] Output size: the size after convolution with a 3×3×256 convolution kernel.

[0075] Residual connection: The input of the ninth convolutional layer is directly connected to the output of the ninth convolutional layer.

[0076] The tenth convolutional layer (conv10):

[0077] Input size: Output of the ninth convolutional layer (including residual connections).

[0078] Output size: the size after convolution with a 3×3×256 convolution kernel.

[0079] Residual connection: The input of the tenth convolutional layer is directly connected to the output of the tenth convolutional layer.

[0080] The eleventh convolutional layer (conv11):

[0081] Input size: Output of the tenth convolutional layer (including residual connections).

[0082] Output size: the size after convolution with a 3×3×256 convolution kernel.

[0083] Residual connection: The input of the eleventh convolutional layer is directly connected to the output of the eleventh convolutional layer.

[0084] The twelfth convolutional layer (conv12):

[0085] Input size: Output of the eleventh convolutional layer (including residual connections).

[0086] Output size: the size after convolution with a 3×3×256 convolution kernel.

[0087] Residual connection: The input of the twelfth convolutional layer is directly connected to the output of the twelfth convolutional layer.

[0088] The thirteenth convolutional layer (conv13):

[0089] Input size: Output of the 12th convolutional layer (including residual connections).

[0090] Output size: the size after convolution with a 3×3×512 convolution kernel.

[0091] Residual connection: The input of the thirteenth convolutional layer is directly connected to the output of the thirteenth convolutional layer.

[0092] The fourteenth convolutional layer (conv14):

[0093] Input size: The output of the 13th convolutional layer (including residual connections).

[0094] Output size: the size after convolution with a 3×3×512 convolution kernel and a stride of 2.

[0095] Residual connection: The input of the fourteenth convolutional layer is directly connected to the output of the fourteenth convolutional layer.

[0096] The fifteenth convolutional layer (conv15):

[0097] Input size: The output of the 14th convolutional layer (including residual connections).

[0098] Output size: the size after convolution with a 3×3×512 convolution kernel and a stride of 2.

[0099] Residual connection: The input of the fifteenth convolutional layer is directly connected to the output of the fifteenth convolutional layer.

[0100] Fully connected layer:

[0101] Fully connected layer (FC layer):

[0102] Input size: The output of the fifteenth convolutional layer (including residual connections).

[0103] Output size: 128 neurons.

[0104] Function: Map the features extracted by the convolutional layer to the final output space, and finally output the fingerprint features for the subsequent identity recognition process.

[0105] Second, the facial branch model: the improved backbone network of the face recognition algorithm FaceNet is used to build the facial branch, perform feature extraction on the facial image, and extract the key feature points of the face. The specific model structure is shown in Table 1. The specific structure of the model is as follows:

[0106] Input layer: receives face images with a size of 220×220×3.

[0107] Convolutional layer (conv1): The first convolutional layer uses a 7×7×3 kernel and a stride of 2, with an output size of 110×110×64.

[0108] Pooling layer (pool1): The first pooling layer uses a 3×3×64 convolution kernel and a stride of 2, with an output size of 55×55×64.

[0109] Normalization layer (rnorm1): Normalizes the features.

[0110] Convolutional layer (conv2a, conv2): The second convolutional layer uses 1×1×64 and 3×3×64 convolution kernels with an output size of 55×55×192.

[0111] Normalization layer (rnorm2): Normalizes the features.

[0112] Pooling layer (pool2): The second pooling layer uses a convolution kernel of 3×3×192 and a stride of 2, with an output size of 28×28×192.

[0113] Convolutional layer (conv3a, conv3): The third convolutional layer uses 1×1×192 and 3×3×192 convolution kernels with an output size of 28×28×384.

[0114] Pooling layer (pool3): The third pooling layer uses a convolution kernel of 3×3×384 and a stride of 2, with an output size of 14×14×384.

[0115] Convolutional layer (conv4a): Input size: the output of the third pooling layer, i.e. 14×14×384. Output size: 14×14×384. Convolution kernel: 1×1×384, stride 1. Function: Adjust the number of channels through a 1×1 convolution kernel without changing the spatial dimension.

[0116] Convolutional layer (Cony4): Input size: the output of conv4a layer, i.e. 14×14×384. Output size: 14×14×256. Convolution kernel: 3×3×384, stride 1. Function: Feature extraction through 3×3 convolution kernel while reducing the number of channels.

[0117] Convolution layer (cony5a): Input size: the output of Cony4 layer, i.e. 14×14×256. Output size: 14×14×256. Convolution kernel: 1×1×256, step size is 1. Function: adjust the number of channels through 1×1 convolution kernel without changing the spatial dimension.

[0118] Convolutional layer (conv5): Input size: the output of cony5a layer, i.e. 14×14×256. Output size: 14×14×256. Convolution kernel: 3×3×256, stride 1. Function: Feature extraction through 3×3 convolution kernel, keeping the number of channels unchanged.

[0119] Convolutional layer (conv6a): Input size: the output of conv5 layer, i.e. 14×14×256. Output size: 14×14×256. Convolution kernel: 1×1×256, stride 1. Function: Adjust the number of channels through 1×1 convolution kernel without changing the spatial dimension.

[0120] Convolutional layer (Cony6): Input size: the output of conv6a layer, i.e. 14×14×256. Output size: 14×14×256. Convolution kernel: 3×3×256, stride 1. Function: Feature extraction through 3×3 convolution kernel, keeping the number of channels unchanged.

[0121] Pooling layer (pool4): Input size: the output of Cony6 layer, i.e. 14×14×256. Output size: 7×7×256. Convolution kernel: 3×3×256, stride 2. Function: Downsample through the pooling layer to reduce the spatial dimension while keeping the number of channels unchanged.

[0122] Feature fusion and fully connected layer:

[0123] Concatenation layer (concat): Input size: the output of pool4 layer, that is, 7×7×256. Output size: 7×7×256. Function: Concatenate feature maps from different sources to increase feature diversity.

[0124] Fully connected layer (fcl): Input size: the output of the concat layer, i.e. 7×7×256. Output size: 1×32×128. Activation function: ma×out, parameter p=2. Function: Perform nonlinear transformation of features through ma×out activation function to reduce feature dimension.

[0125] Fully connected layer (fc2): Input size: the output of the fcl layer, i.e. 1×32×128. Output size: 1×32×128. Activation function: ma×out, parameter p=2. Function: Further perform nonlinear transformation of features, keeping the feature dimension unchanged.

[0126] Fully connected layer (fc3): Input size: the output of the fc2 layer, i.e. 1×32×128. Output size: 1×1×128. Function: Reduce the feature dimension to 1×1×128 to prepare for the final classification or other tasks.

[0127] Loss layer (L2): Input size: output of fc3 layer, i.e. 1×1×128. Output size: 1×1×128.

[0128] Table 1. Structure diagram of facial branch model

[0129]

[0130] Third, the signature key branching model: encrypt the personal signature key provided by the expert, extract its hash value as a feature, input the hash value of the key into the multi-layer perceptron MLP model to obtain the key feature. The structure of the multi-layer perceptron MLP model is as follows:

[0131] Input layer: Input size: 1×1×1024. Function: Receive a 1024-bit personal digital key.

[0132] Fully connected layer 1 (fc1): Input size: 1×1×1024. Output size: 1×1×512. Activation function: ReLU (Rectified Linear Unit). Function: Perform nonlinear transformation through the ReLU activation function, reduce the feature dimension, and increase the nonlinear expression ability of the network.

[0133] Fully connected layer 2 (fc2): Input size: 1×1×512 (output from fc1). Output size: 1×1×256. Activation function: ReLU. Function: Further perform nonlinear transformation to continue reducing feature dimensions.

[0134] Fully connected layer 3 (fc3): Input size: 1×1×256 (output from fc2). Output size: 1×1×128. Activation function: tanh. Function: Reduce the feature dimension to 128 bits in preparation for the final output.

[0135] Output layer: Output size: 1×1×128. Function: Provides 128-bit output, which can be used for various tasks such as classification, feature representation or others.

[0136] Fourth, cross-attention structure design: Based on the three-branch model, the present invention designs a cross-attention structure to enhance the information interaction and fusion between different branches. Specifically, a cross-attention structure is designed between the face and personal key branches and between the fingerprint and personal key branches. The input tensor of the personal key branch is mapped to K = V = X2W K , the input tensors of the face branch and fingerprint branch are mapped to Q = X1W Q , the calculation of cross attention is shown in formula (1). Taking the cross attention between the face and personal key branches as an example, the input tensor mapping of the personal key branch is K = V = X2W K , the input tensor mapping of the face branch is Q = X1W Q , the feature information K and Q of the two branches are correlated, and the correlation weight is added to the tensor V of the personal key branch, and the cross attention weight of the two branches is obtained. This achieves the fusion and complementarity of feature information. This design can effectively improve the robustness and generalization ability of the model.

[0137]

[0138] Fifth, in the process of the present invention, three biometric features of the expert's fingerprint, facial image and personal signature key are first collected through the data collection step. Subsequently, these features are respectively sent to the corresponding branch models for feature extraction: fingerprint branch model, facial branch model and signature key branch model. Each branch model uses deep learning technology to extract the key information of its own biometric features. After feature extraction, the output of the facial branch model (FaceNet model) and the output of the signature key branch model (MLP model) are calculated by formula (1) through the designed cross-attention structure to obtain the Attention Score Matrix, and then the output of the facial branch model and the output of the signature key branch model are concat spliced ​​with the Attention Score Matrix, and the output result is recorded as A. Similarly, the output of the fingerprint branch model (FingerNet model) and the output of the signature key branch model (MLP model) are calculated by formula (1) to obtain the Attention Score Matrix, and then the output of the fingerprint branch model and the output of the signature key branch model are concat spliced ​​with the Attention Score Matrix, and the output result is recorded as B. Then, the spliced ​​result A is multiplied and fused with the output of the fingerprint branch model to output the feature vector a. Similarly, the spliced ​​result B is multiplied and fused with the output of the face branch model to output the feature vector b. Finally, the output feature vector a and feature vector b are additively fused, and the output is the result of multiple biometrics, that is, the identity recognition result. This is to determine whether the current user matches the preset expert identity, thereby allowing or denying the expert to conduct remote consultations and issuing an alarm in the event of a mismatch. The entire system aims to improve the accuracy and security of telemedicine identity authentication through this multimodal and multi-branch approach.

[0139] Sixth, model training and optimization

[0140] The fingerprint branch model uses the ImageNet dataset for model pre-training; the facial branch model uses the trained model parameters provided by FaceNet; then a large amount of expert identity data containing multiple biometric features is used to fine-tune the model and optimize the model parameters.

[0141] In summary, the present invention improves the accuracy and security of identity recognition by integrating multiple biometric features for telemedicine identity authentication, and provides a strong guarantee for the effectiveness of telemedicine services and the life safety of patients.

[0142] The above is only a specific implementation of the present invention, which enables those skilled in the art to understand or implement the present invention. Although detailed descriptions are given with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments, and they should all be covered by the protection scope of the claims.

Claims

1. A method for identity authentication in telemedicine by integrating multiple biometric features, characterized in that: The steps include: (1) Data collection: First, collect the expert’s fingerprint, facial image, and personal signature key; (2) Feature extraction: The collected biometric features are input into the corresponding branch models, namely the fingerprint branch model, the facial branch model, and the signature key branch model, to extract their respective features; (3) Feature fusion: Design a cross-attention structure, through which the feature information output by different branch models is weighted and passed to other branches, achieving the fusion and complementarity of feature information and outputting an identity recognition result; (4) Identity matching: If the identity recognition result matches the preset expert identity, the expert is allowed to conduct remote consultation; otherwise, access is denied and an alarm is issued.

2. According to claim 1, a method for identity authentication for remote medical treatment by integrating multiple biometric features, characterized in that: In step (2), the fingerprint branch model uses the improved backbone network of the fingerprint recognition model FingerNet to extract features of the fingerprint image and extract the texture and detail features of the fingerprint. The specific model structure is as follows: Input layer: receives fingerprint image; The feature extraction layer specifically includes: The first convolutional layer: Input size: fingerprint image; Output size: size after convolution and pooling with a 7×7×64 convolution kernel and a stride of 2; Second convolutional layer: Input size: output of the first pooling layer; Output size: size after convolution with 3×3×64 convolution kernel and stride 2; Residual connection: The input of the second convolutional layer is directly connected to the output of the second convolutional layer to form a residual connection; The third convolutional layer: input size: the output of the second convolutional layer; output size: the size after convolution with a 3×3×64 convolution kernel; Residual connection: The input of the third convolutional layer is directly connected to the output of the third convolutional layer; The fourth convolutional layer: input size: the output of the third convolutional layer; output size: the size after convolution with a 3×3×64 convolution kernel; Residual connection: The input of the fourth convolutional layer is directly connected to the output of the fourth convolutional layer; The fifth convolutional layer: input size: the output of the fourth convolutional layer; output size: the size after convolution with a 3×3×128 convolution kernel; Residual connection: The input of the fifth convolutional layer is directly connected to the output of the fifth convolutional layer; The sixth convolutional layer: Input size: the output of the fifth convolutional layer; Output size: the size after convolution with a 3×3×128 convolution kernel and a stride of 2; Residual connection: The input of the sixth convolutional layer is directly connected to the output of the sixth convolutional layer; The seventh convolutional layer: input size: the output of the sixth convolutional layer; output size: the size after convolution with a 3×3×128 convolution kernel; Residual connection: The input of the seventh convolutional layer is directly connected to the output of the seventh convolutional layer; The eighth convolutional layer: input size: the output of the seventh convolutional layer; output size: the size after convolution with a 3×3×128 convolution kernel and a stride of 2; Residual connection: The input of the eighth convolutional layer is directly connected to the output of the eighth convolutional layer; Ninth convolutional layer: Input size: output of the eighth convolutional layer; Output size: size after convolution with 3×3×256 convolution kernel; Residual connection: The input of the ninth convolutional layer is directly connected to the output of the ninth convolutional layer; The tenth convolutional layer: Input size: the output of the ninth convolutional layer; Output size: the size after convolution with a 3×3×256 convolution kernel; Residual connection: The input of the tenth convolutional layer is directly connected to the output of the tenth convolutional layer; Eleventh convolutional layer: Input size: output of the tenth convolutional layer; Output size: size after convolution with 3×3×256 convolution kernel; Residual connection: The input of the eleventh convolutional layer is directly connected to the output of the eleventh convolutional layer; The twelfth convolutional layer: input size: the output of the eleventh convolutional layer; output size: the size after convolution with a 3×3×256 convolution kernel; Residual connection: The input of the twelfth convolutional layer is directly connected to the output of the twelfth convolutional layer; Thirteenth convolutional layer: Input size: output of the twelfth convolutional layer; Output size: size after convolution with 3×3×512 convolution kernel; Residual connection: The input of the thirteenth convolutional layer is directly connected to the output of the thirteenth convolutional layer; The fourteenth convolutional layer: Input size: the output of the thirteenth convolutional layer; Output size: the size after convolution with a 3×3×512 convolution kernel and a stride of 2; Residual connection: The input of the fourteenth convolutional layer is directly connected to the output of the fourteenth convolutional layer; The fifteenth convolutional layer: Input size: the output of the fourteenth convolutional layer; Output size: the size after convolution with a 3×3×512 convolution kernel and a stride of 2; Residual connection: The input of the fifteenth convolutional layer is directly connected to the output of the fifteenth convolutional layer; Fully connected layer: Fully connected layer: Input size: output of the fifteenth convolutional layer; Output size: 128 neurons.

3. The identity authentication method for remote medical treatment by integrating multiple biometric features according to claim 1, characterized in that: In step (2), the facial branch model uses the improved backbone network of the face recognition algorithm FaceNet to build the facial branch, extract features from the facial image, and extract the key feature points of the face. The specific model structure is as follows: Input layer: receives facial images with a size of 220×220×3; Convolutional layer conv1: The first convolutional layer uses a 7×7×3 convolution kernel and a stride of 2, with an output size of 110×110×64; Pooling layer pool1: The first pooling layer uses a 3×3×64 convolution kernel and a stride of 2, with an output size of 55×55×64; Normalization layer rnorm1: normalizes the features; Convolutional layers conv2a and conv2: The second convolutional layer uses 1×1×64 and 3×3×64 convolution kernels with an output size of 55×55×192; Normalization layer rnorm2: normalizes the features; Pooling layer pool2: The second pooling layer uses a convolution kernel of 3×3×192 and a stride of 2, with an output size of 28×28×192; Convolutional layers conv3a and conv3: The third convolutional layer uses convolution kernels of 1×1×192 and 3×3×192, with an output size of 28×28×384; Pooling layer pool3: The third pooling layer uses a convolution kernel of 3×3×384 and a stride of 2, with an output size of 14×14×384; Convolutional layer conv4a: Input size: the output of the third pooling layer, i.e. 14×14×384; Output size: 14×14×384; Convolution kernel: 1×1×384, step size is 1; Convolutional layer Cony4: Input size: the output of conv4a layer, i.e. 14×14×384; Output size: 14×14×256; Convolution kernel: 3×3×384, step size is 1; Convolutional layer cony5a: Input size: the output of Cony4 layer, that is, 14×14×256; Output size: 14×14×256; Convolution kernel: 1×1×256, step size is 1; Convolutional layer conv5: Input size: output of cony5a layer, i.e. 14×14×256; Output size: 14×14×256; Convolution kernel: 3×3×256, step size is 1; Convolutional layer conv6a: Input size: the output of conv5 layer, that is, 14×14×256; Output size: 14×14×256; Convolution kernel: 1×1×256, step size is 1; Convolutional layer Cony6: Input size: the output of conv6a layer, i.e. 14×14×256; Output size: 14×14×256; Convolution kernel: 3×3×256, step size is 1; Pooling layer pool4: Input size: the output of Cony6 layer, that is, 14×14×256; Output size: 7×7×256; Convolution kernel: 3×3×256, step size is 2; Feature fusion and fully connected layer: Concatenation layer concat: Input size: output of pool4 layer, i.e. 7×7×256; Output size: 7×7×256; Fully connected layer fcl: input size: the output of the concat layer, i.e. 7×7×256; output size: 1×32×128; activation function: ma×out, parameter p=2; Fully connected layer fc2: Input size: output of fcl layer, i.e. 1×32×128; Output size: 1×32×128; Activation function: ma×out, parameter p=2; Fully connected layer fc3: Input size: output of fc2 layer, i.e. 1×32×128; Output size: 1×1×128; Loss layer L2: Input size: output of fc3 layer, i.e. 1×1×128; Output size: 1×1×128.

4. The method for identity authentication by integrating multiple biometric features for remote medical treatment according to claim 1, characterized in that: In step (2), the signature key branch model encrypts the personal signature key provided by the expert, extracts its hash value as a feature, and inputs the hash value of the key into the multi-layer perceptron MLP model to obtain the key feature; The multi-layer perceptron MLP model structure is as follows: Input layer: input size: 1×1×1024; Fully connected layer fc1: input size: 1×1×1024; output size: 1×1×512; activation function: ReLU; Fully connected layer fc2: input size: 1×1×512; output size: 1×1×256; activation function: ReLU; Fully connected layer fc3: input size: 1×1×256; output size: 1×1×128; activation function: tanh; Output layer: Output size: 1×1×128.

5. The method for identity authentication by integrating multiple biometric features for remote medical treatment according to claim 1, characterized in that: In step (3), a cross-attention structure is designed between the face and personal key branches and the fingerprint and personal key branches respectively; The cross attention structure is: the input tensor mapping of one branch is K = V = X2W K , the input tensor mapping of the other branch is Q = X1W Q ,but:

6. The method for identity authentication for remote medical treatment by integrating multiple biometric features according to claim 5, characterized in that: In step (3), after feature extraction, the fingerprint branch model, the face branch model and the signature key branch model use a cross-attention structure to calculate the output of the face branch model and the output of the signature key branch model by formula (1) to obtain the Attention Score Matrix, and then concatenate the output of the face branch model and the output of the signature key branch model with the Attention Score Matrix, and the result is recorded as A. Similarly, the output of the fingerprint branch model and the output of the signature key branch model are calculated by formula (1) to obtain the Attention Score Matrix, and then concatenate the output of the fingerprint branch model and the output of the signature key branch model with the Attention Score Matrix, and the result is recorded as B; then, the concatenated result A is multiplied and fused with the output of the fingerprint branch model to output the feature vector a. Similarly, the concatenated result B is multiplied and fused with the output of the face branch model to output the feature vector b; finally, the output feature vector a and feature vector b are additively fused, and the output obtained is the result of multiple biometrics, that is, the identity recognition result.

Images