Console login identity information acquisition and verification method based on deep learning

Abstract

The invention relates to the technical field of information security, and discloses a console login identity information acquisition and verification method based on deep learning, which comprises the following steps: S1, acquiring biological characteristics and behavior characteristics of a user, extracting a feature vector of the biological characteristics by adopting a deep learning model, extracting a feature vector of the behavior characteristics by adopting a time sequence model, and acquiring a user login identity information database; the feature vectors of the biological features and the feature vectors of the behavior features are spliced to obtain identity feature vectors, and the feature vectors of the biological features and the feature vectors of the behavior features are spliced to construct the identity feature vectors; s2, carrying out dimensionality reduction on the identity feature vector by adopting a topology preserving projection method, and constructing a low-dimensional manifold embedding representation; and S3, adding a manifold regularization item in the loss calculation process of the deep learning model. According to the method, deep learning and a time sequence model are combined to extract identity features, the effect of quickly and accurately collecting biological and behavior information is achieved, and compared with a single feature recognition scheme in the prior art, the problems of high false recognition rate and low recognition efficiency are solved.

Description

Technical Field

[0001] The present invention relates to the field of information security technology, and particularly to a method for collecting and verifying console login identity information based on deep learning. Background Art

[0002] In the modern field of information security, identity authentication is a key link to ensure system security. Traditional console login verification methods mainly rely on the combination of username and password, which are vulnerable to threats such as password leakage and brute-force attacks, and have low security. To improve security, in recent years, biometric recognition technologies have gradually been applied to identity authentication, such as face recognition and fingerprint recognition. However, a single biometric recognition scheme has problems of high false recognition rate and low recognition efficiency in practical applications. Especially in complex environments, the recognition accuracy cannot be guaranteed.

[0003] In addition, existing feature dimensionality reduction methods mostly adopt simple linear dimensionality reduction techniques, such as principal component analysis. When dealing with high-dimensional data, it is easy to cause loss of feature structure, which in turn affects the generalization ability of the model. The deficiency of feature dimensionality reduction makes the separability of identity features decrease in the high-dimensional feature space, affecting the accuracy of identity authentication.

[0004] In terms of abnormal login detection, traditional methods usually rely on a single measurement means, such as simple threshold judgment. When facing complex attack patterns, they show insensitivity, resulting in a high false positive rate and being unable to effectively identify abnormal login behaviors.

[0005] In the prior art, a single biometric recognition method performs poorly in complex environments. Due to the influence of factors such as environmental light and angle changes, the recognition accuracy and efficiency are limited, resulting in a high false recognition rate and being unable to meet the requirements of high-security scenarios.

[0006] Traditional linear dimensionality reduction techniques are prone to losing important feature structure information when dealing with high-dimensional identity features. The loss of feature structure directly affects the generalization ability of the model, making the model unable to adapt to the changing distribution of identity features in practical applications and reducing the reliability of identity authentication.

[0007] Existing abnormal login detection methods mostly rely on simple threshold judgment and lack sensitivity to complex attack patterns. A single measurement means is prone to misjudgment when facing diverse attacks, and is unable to effectively distinguish normal and abnormal login behaviors, affecting the security of the system.

[0008] Therefore, the present invention proposes a method for collecting and verifying console login identity information based on deep learning to solve the above-mentioned problems. Summary of the Invention

[0009] In view of the deficiencies of the prior art, the present invention provides a method for collecting and verifying console login identity information based on deep learning to solve the problems raised in the above background art.

[0010] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for collecting and verifying console login identity information based on deep learning, including: S1, collect the biometric and behavioral characteristics of the user, extract the feature vectors of the biometric characteristics using a deep learning model, extract the feature vectors of the behavioral characteristics using a time series model, and splice them to obtain an identity feature vector. Moreover, splice the feature vectors of the biometric characteristics and the feature vectors of the behavioral characteristics to construct an identity feature vector; S2, use the topology-preserving projection method to reduce the dimension of the identity feature vector and construct a low-dimensional manifold embedding representation; S3, add a manifold regularization term during the loss calculation process of the deep learning model; S4, calculate the Mahalanobis distance of the identity features for identity verification. If the calculated Mahalanobis distance exceeds the set threshold, it is determined that the current login behavior is abnormal and secondary verification is triggered; S5, calculate the probability density value of the identity features. If it is lower than the set threshold or the Mahalanobis distance exceeds the set threshold, reject the current login request, otherwise allow normal login to finally complete the decision-making process of identity verification.

[0011] Preferably, in the above S1, the collection of the biometric and behavioral characteristics of the user further includes: Sub-step 1.1, collect the user's face, fingerprint, voiceprint data, and keyboard input and mouse trajectory data. Among them, the face, fingerprint, and voiceprint data are biometric inputs, and the keyboard input and mouse trajectory data are behavioral inputs; Sub-step 1.2, use a deep learning model to extract the feature vectors of the biometric inputs to obtain biometric feature vectors. The deep learning model is defined as: , where, is the biometric input, is the weight matrix, is the bias vector, is the activation function; Sub-step 1.3, use a time series model to extract the feature vectors of the behavioral inputs to obtain behavioral feature vectors. The time series model is defined as: , where, is the behavioral input, is the weight matrix, is the bias vector, is an activation function; Then, the biometric feature vector and the behavioral feature vector are synthesized into an identity feature vector through feature concatenation.

[0012] Preferably, in step S2, the topological preservation projection method is used to reduce the dimension of the identity feature vector, and constructing the low-dimensional manifold embedding representation further includes: Sub-step 2.1: Input the identity feature vector into the topological preservation projection method, where the identity feature vector is the biometric feature vector and the behavioral feature vector concatenated in step S1; Sub-step 2.2: In the topological preservation projection, construct an adjacency matrix A to represent the similarity between samples, and the adjacency matrix is calculated by the following formula: , where, and are identity feature vectors, is the adjacency matrix; Sub-step 2.3: According to the adjacency matrix A, calculate the projection matrix W so that the distance between adjacent samples remains unchanged in the projected low-dimensional space, and the projection matrix is calculated through the following optimization objective: , where W is the projection matrix, and are identity feature vectors, is the adjacency matrix; Through the optimization objective, similar identity samples are kept close in the low-dimensional space, thereby constructing the low-dimensional manifold embedding representation.

[0013] Preferably, in step S3, adding a manifold regularization term in the process of calculating the loss of the deep learning model further includes: Sub-step 3.1: Calculate the cross-entropy loss based on the identity feature vector generated in step S1, and the cross-entropy loss is defined as: , where, is the true class probability of the sample, is the predicted probability output by the model; Sub-step 3.2: Calculate the manifold regularization loss using the adjacency matrix constructed in step S2, and the manifold regularization loss is defined as: , where, is the low-dimensional representation obtained after mapping the identity feature vector , is the low-dimensional representation obtained after mapping the identity feature vector , is the adjacency matrix, is the regularization weight; In sub-step 3.3, combining sub-step 3.1 and sub-step 3.2, the cross-entropy loss and the manifold regularization loss are added to obtain the total loss: , wherein, is the overall loss of the model, ensuring classification accuracy and maintaining the consistency of the low-dimensional manifold structure during the identity feature extraction process.

[0014] Preferably, in S4, calculating the Mahalanobis distance of the identity feature for identity verification further includes: Sub-step 4.1, according to the low-dimensional manifold embedding representation generated in step S2, calculate the identity feature vector of the logged-in user and the mean of the identity features of normal users The difference is defined as: , wherein, is the difference between the identity feature vector of the logged-in user and the mean of the identity features of normal users, is the identity feature vector of the logged-in user, is the mean of the identity features of normal users; Sub-step 4.2, calculate the covariance matrix of the difference of the identity feature vectors , and the matrix is used to quantify the variance and covariance of the identity features. The covariance matrix is defined as: , wherein, is the identity feature vector, is the number of samples, is the covariance matrix of the samples; Sub-step 4.3, calculate the Mahalanobis distance to measure the distance between the identity feature of the logged-in user and the mean of normal users. The Mahalanobis distance is defined as: , wherein, is the Mahalanobis distance, is the identity feature vector of the logged-in user, is the mean of the identity features of normal users, is the inverse matrix of the covariance matrix, is the transpose operation; is used to calculate the similarity between identity features. If the Mahalanobis distance exceeds the set threshold, it is determined that the login behavior is abnormal and secondary verification is triggered.

[0015] Preferably, in S5, calculating the probability density value of the identity feature further includes: Sub-step 5.1: Embed the identity feature vector of the logged-in user in the low-dimensional manifold embedding representation generated in step S2 Input it into the Gaussian probability density function to calculate the probability density value. The formula is: , where is the probability density value of the identity feature vector of the logged-in user, d is the dimension of the low-dimensional manifold space, is the covariance matrix of the determinant, is the covariance matrix of the identity feature vector, is the identity feature vector of the logged-in user, is the mean of the identity features of normal users; Sub-step 5.2: Compare the probability density value calculated in sub-step 5.1 with the set probability density threshold . represents the critical value for judging the probability density of identity features; Sub-step 5.3: Combine the comparison result of the probability density value in sub-step 5.2 with the Mahalanobis distance calculated in step S4 and compare it with the set Mahalanobis distance threshold for judgment. The decision formula is: , where is the Mahalanobis distance, represents the Mahalanobis distance critical value for judging the similarity of identity features, which is used to make an identity verification decision after integrating the probability density and Mahalanobis distance results.

[0016] Preferably, the extraction of the biometric features includes: extracting face features using the ResNet-50 network, extracting fingerprint features using the MobileNet network, and extracting voiceprint features using the LSTM network; The extraction of the behavioral features includes: extracting keyboard input pattern features using Keystroke-Dynamics and extracting mouse trajectory features using dynamic time warping.

[0017] Preferably, the projection matrix of the topological preservation projection method is obtained by calculating the optimization objective, and the above optimization objective is to minimize the sum of the squared Euclidean distances after projection constrained by the adjacency matrix between all samples.

[0018] A terminal device includes a sensor module for collecting user biometric and behavioral features and a processing module for executing the console login identity information collection and verification method.

[0019] A storage medium stores an instruction program for controlling a terminal device to collect and verify identity information according to a console login identity information collection and verification method.

[0020] The present invention provides a console login identity information collection and verification method based on deep learning, having the following beneficial effects: 1. The present invention jointly extracts identity features by using deep learning and a time series model, achieving the effect of quickly and accurately collecting biological and behavioral information. Compared with the single feature recognition scheme in the prior art, it solves the problems of high misrecognition rate and low recognition efficiency.

[0021] 2. The present invention introduces topological preserving projection and manifold regularization to achieve the effect of optimizing the low-dimensional embedding representation. Compared with the simple dimensionality reduction method in the prior art, it solves the defects of feature structure loss and weak model generalization ability.

[0022] 3. The present invention designs a verification mechanism based on the joint decision of Mahalanobis distance and probability density, achieving the effect of improving the accuracy of identity verification. Compared with the single metric means in the prior art, it solves the problems of insensitive abnormal login detection and high false positive rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0024] To enable those skilled in the art to understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0025] The present invention will be described in detail below with reference to the accompanying drawings: Embodiment

[0026] Please refer to the attached Figure 1 , the embodiment of the present invention provides a console login identity information collection and verification method based on deep learning, including: S1. Collect the biological features and behavioral features of the user, extract the feature vector of the biological features by using a deep learning model, extract the feature vector of the behavioral features by using a time series model, and splice them to obtain an identity feature vector, and splice the feature vector of the biological features and the feature vector of the behavioral features to construct an identity feature vector; The extraction of biological features includes: extracting face features by using a ResNet-50 network, extracting fingerprint features by using a MobileNet network, and extracting voiceprint features by using an LSTM network; The extraction of behavioral features includes: extracting keyboard input pattern features using Keystroke - Dynamics, and extracting mouse trajectory features using Dynamic Time Warping; Sub - step 1.1: Collect user face, fingerprint, voiceprint data and keyboard input, mouse trajectory data. Among them, face, fingerprint, and voiceprint data are biometric inputs, and keyboard input and mouse trajectory data are behavioral feature inputs; Sub - step 1.2: Use a deep - learning model to extract features from biometric inputs to obtain biometric feature vectors. The deep - learning model is defined as: , where, is the biometric input, is the weight matrix, is the bias vector, is the activation function; Sub - step 1.3: Use a time - series model to extract features from behavioral feature inputs to obtain behavioral feature vectors. The time - series model is defined as: , where, is the behavioral feature input, is the weight matrix, is the bias vector, is the activation function; Then, the biometric feature vector and the behavioral feature vector are synthesized into an identity feature vector through feature concatenation; S2: Use the topological - preservation projection method to reduce the dimension of the identity feature vector and construct a low - dimensional manifold embedding representation; The projection matrix of the topological - preservation projection method is obtained by calculating the optimization objective. The above - mentioned optimization objective is to minimize the sum of the squared Euclidean distances after projection subject to the adjacency matrix constraint of all samples; Sub - step 2.1: Input the identity feature vector into the topological - preservation projection method. Among them, the identity feature vector is the biometric feature vector and the behavioral feature vector concatenated in step S1; Sub - step 2.2: In the topological - preservation projection, construct an adjacency matrix A to represent the similarity between samples. The adjacency matrix is calculated by the following formula: , where, and are identity feature vectors, is the adjacency matrix; Sub - step 2.3: According to the adjacency matrix A, calculate the projection matrix W so that in the projected low - dimensional space, the distance between adjacent samples remains unchanged. The projection matrix is obtained by calculating the following optimization objective: , where W is the projection matrix, and is the identity feature vector, is the adjacency matrix; By optimizing the objective, similar identity samples are kept close in the low-dimensional space, and then a low-dimensional manifold embedding representation is constructed; S3. During the loss calculation process of the deep learning model, add a manifold regularization term; Sub-step 3.1. Calculate the cross-entropy loss based on the identity feature vector generated in step S1. The cross-entropy loss is defined as: , where is the true class probability of the sample, is the predicted probability output by the model; Sub-step 3.2. Calculate the manifold regularization loss using the adjacency matrix constructed in step S2. The manifold regularization loss is defined as: , where is the identity feature vector after being mapped to the low-dimensional representation, is the identity feature vector after being mapped to the low-dimensional representation, is the adjacency matrix, is the regularization weight; Sub-step 3.3. Combine sub-step 3.1 and sub-step 3.2 to add the cross-entropy loss and the manifold regularization loss to obtain the total loss: , where is the overall loss of the model, ensuring classification accuracy and maintaining the consistency of the low-dimensional manifold structure during the identity feature extraction process; S4. Calculate the Mahalanobis distance of the identity feature for identity verification. If the calculated Mahalanobis distance exceeds the set threshold, it is determined that the current login behavior is abnormal and secondary verification is triggered; Sub-step 4.1. According to the low-dimensional manifold embedding representation generated in step S2, calculate the identity feature vector of the logged-in user and the mean of the normal user identity features. The difference is defined as: , where is the difference between the identity feature vector of the logged-in user and the mean of the normal user identity features, is the identity feature vector of the logged-in user, is the mean of the normal user identity features; Sub-step 4.2: Calculate the covariance matrix of the identity feature vector differences , where the matrix is used to quantify the variance and covariance of the identity features, and the covariance matrix is defined as: , where, is the identity feature vector, is the number of samples, is the covariance matrix of the samples; Sub-step 4.3: Calculate the Mahalanobis distance to measure the distance between the identity features of the logged-in user and the mean of the normal users. The Mahalanobis distance is defined as: , where, is the Mahalanobis distance, is the identity feature vector of the logged-in user, is the mean of the normal user identity features, is the inverse matrix of the covariance matrix, is the transpose operation; is used to calculate the similarity between identity features. If the Mahalanobis distance exceeds the set threshold, it is determined that the login behavior is abnormal and secondary verification is triggered; S5: Calculate the probability density value of the identity features. If it is lower than the set threshold or the Mahalanobis distance exceeds the set threshold, reject the current login request; otherwise, allow normal login to finally complete the decision-making process of identity verification; Sub-step 5.1: Input the identity feature vector of the logged-in user in the low-dimensional manifold embedding representation generated in step S2 into the Gaussian probability density function to calculate the probability density value. The formula is: , where, is the probability density value of the identity feature vector of the logged-in user, d is the dimension of the low-dimensional manifold space, is the covariance matrix of the determinant, is the covariance matrix of the identity feature vector, is the identity feature vector of the logged-in user, is the mean of the normal user identity features; Sub-step 5.2: Compare the probability density value calculated in sub-step 5.1 with the set probability density threshold , represents the critical value for judging the probability density of the identity features; Sub-step 5.3: Combine the probability density value comparison result in sub-step 5.2 with the Mahalanobis distance calculated in step S4 with the set Mahalanobis distance threshold to make a judgment. The decision formula is: , where is the Mahalanobis distance, represents the Mahalanobis distance critical value for judging the similarity of identity features, which is used to make an identity verification decision after comprehensively considering the probability density and Mahalanobis distance results.

[0027] In step S1, by combining a deep learning model and a time series model, the biometric and behavioral characteristics of the user are comprehensively collected. Advanced network structures such as ResNet-50, MobileNet, and LSTM are used to extract face, fingerprint, and voiceprint features in sequence to ensure the high-precision extraction of biometric features. At the same time, using Keystroke-Dynamics and dynamic time warping techniques, the keyboard input pattern and mouse trajectory features of the user are extracted to enrich the diversity of behavioral characteristics. Through feature splicing, the biometric and behavioral characteristics are synthesized into an identity feature vector to enhance the expression ability of the identity features, providing a basis for subsequent dimensionality reduction and verification.

[0028] In step S2, a topology-preserving projection method is used to reduce the dimensionality of the identity feature vector and construct a low-dimensional manifold embedding representation. By minimizing the sum of the squared Euclidean distances after projection constrained by the adjacency matrix, it is ensured that similar identity samples maintain a neighboring relationship in the low-dimensional space. This method can effectively reduce the feature dimension, reduce the computational complexity, and retain the topological structure of the identity features, improving the generalization ability and robustness of the model, and laying a foundation for subsequent regularization and verification steps.

[0029] In step S3, during the loss calculation process of the deep learning model, a manifold regularization term is added. By calculating the cross-entropy loss and the manifold regularization loss and adding the two to obtain the total loss, it is ensured that the model can maintain both classification accuracy and the consistency of the low-dimensional manifold structure during the training process. The introduction of manifold regularization effectively prevents the overfitting phenomenon of the model, improves the adaptability of the model in different environments, and enhances the reliability of identity verification.

[0030] In step S4, identity verification is performed by calculating the Mahalanobis distance of the identity features. Using the low-dimensional manifold embedding representation, the difference between the identity features of the logged-in user and the mean of normal users is calculated, and the variance and covariance of the features are quantified through the covariance matrix. The calculation of the Mahalanobis distance provides an effective similarity measurement method, which can accurately identify abnormal login behaviors, trigger secondary verification, and improve the security and protection capabilities of the system.

[0031] In step S5, the final identity verification decision is made by calculating the probability density value of the identity feature and combining the result of the Mahalanobis distance. The Gaussian probability density function is used to evaluate the probability density of the identity feature of the logged-in user and compare it with the set threshold. By comprehensively judging the probability density and the Mahalanobis distance, the accuracy and sensitivity of the identity verification are ensured, the false positive rate is reduced, and the user experience and the security of the system are improved.

[0032] In summary, through the combination of deep learning and time series models, the present invention comprehensively collects and extracts the biological and behavioral characteristics of users, uses topological preservation projection and manifold regularization techniques to optimize the feature dimension reduction and model training processes, and finally improves the accuracy and security of identity verification through the joint decision of the Mahalanobis distance and probability density. It effectively solves the problems of high false recognition rate, loss of feature structure, and insensitivity to anomaly detection in the prior art, and provides an efficient and reliable solution for the acquisition and verification of console login identity information.

[0033] A terminal device includes a sensor module for collecting the biological and behavioral characteristics of a user and a processing module for executing the method for collecting and verifying console login identity information.

[0034] A storage medium stores an instruction program for controlling the terminal device to collect and verify identity information according to the method for collecting and verifying console login identity information.

[0035] The terminal device integrates a sensor module for collecting the biological and behavioral characteristics of a user, and can obtain the identity information of the user in real time and efficiently. Through the sensor module, the device can accurately capture biological characteristics such as face, fingerprint, voiceprint, etc., and behavioral characteristics such as keyboard input mode and mouse trajectory, providing rich data support for identity verification. In addition, the built-in processing module of the device can execute the method for collecting and verifying console login identity information to ensure the efficiency and accuracy of data processing.

[0036] The storage medium stores an instruction program for controlling the terminal device to collect and verify identity information, ensuring the flexibility and scalability of the system. Through the storage medium, the device can quickly load and update the identity verification algorithm to adapt to different application scenarios and security requirements. The use of the storage medium simplifies the maintenance and upgrade process of the device, and improves the stability and reliability of the system, ensuring the continuous effectiveness of identity verification.

[0037] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for collecting and verifying console login identity information based on deep learning, characterized in that, Including: S1, collect the biometric and behavioral characteristics of the user, extract the feature vector of the biometric using a deep learning model, extract the feature vector of the behavior using a time series model, and splice them to obtain an identity feature vector. Specifically, splice the feature vector of the biometric and the feature vector of the behavior to construct an identity feature vector; S2, use the topology-preserving projection method to reduce the dimension of the identity feature vector and construct a low-dimensional manifold embedding representation; S3, add a manifold regularization term during the loss calculation process of the deep learning model; S4, calculate the Mahalanobis distance of the identity feature for identity verification. If the calculated Mahalanobis distance exceeds the set threshold, it is determined that the current login behavior is abnormal and secondary verification is triggered; S5, calculate the probability density value of the identity feature. If it is lower than the set threshold or the Mahalanobis distance exceeds the set threshold, reject the current login request, otherwise allow normal login to finally complete the decision-making process of identity verification.

2. The method for collecting and verifying console login identity information based on deep learning according to claim 1, characterized in that, In the above S1, the collection of the biometric and behavioral characteristics of the user further includes: Sub-step 1.1, collect the user's face, fingerprint, voiceprint data, and keyboard input and mouse trajectory data. Among them, the face, fingerprint, and voiceprint data are biometric inputs, and the keyboard input and mouse trajectory data are behavioral inputs; Sub-step 1.2, use a deep learning model to extract the feature of the biometric input to obtain a biometric feature vector. The deep learning model is defined as: ， Among them, is a biometric input, is a weight matrix, is a bias vector, is an activation function; Sub-step 1.3, use a time series model to extract the feature of the behavioral input to obtain a behavioral feature vector. The time series model is defined as: ， Among them, is the input of behavioral characteristics, is the weight matrix, is the bias vector, is the activation function; Then, synthesize the biometric feature vector and the behavioral feature vector into an identity feature vector through feature splicing.

3. A method for collecting and verifying console login identity information based on deep learning according to claim 1, characterized in that, In the above S2, using the topology-preserving projection method to reduce the dimension of the identity feature vector and construct a low-dimensional manifold embedding representation further includes: Sub-step 2.1, input the identity feature vector into the topology-preserving projection method. Here, the identity feature vector is the biometric feature vector and the behavioral feature vector spliced in step S1; Sub-step 2.2, in the topology-preserving projection, construct an adjacency matrix A to represent the similarity between samples. The adjacency matrix is calculated by the following formula: ， Among them, and are identity feature vectors, is an adjacency matrix; Sub-step 2.3, according to the adjacency matrix A, calculate the projection matrix W so that the distance between adjacent samples remains unchanged in the projected low-dimensional space. The projection matrix is calculated through the following optimization objective: ， where, W is the projection matrix, and is the identity feature vector, is the adjacency matrix; Through the optimization objective, make similar identity samples remain adjacent in the low-dimensional space, and then construct a low-dimensional manifold embedding representation.

4. A method for collecting and verifying console login identity information based on deep learning according to claim 1, characterized in that In the above S3, adding a manifold regularization term during the loss calculation process of the deep learning model further includes: Sub-step 3.1, calculate the cross-entropy loss based on the identity feature vector generated in step S1. The cross-entropy loss is defined as: ， Among them, is the true class probability of the sample, is the predicted probability output by the model; Sub-step 3.2, calculate the manifold regularization loss using the adjacency matrix constructed in step S2. The manifold regularization loss is defined as: ， Among them, is the low-dimensional representation obtained after mapping of the identity feature vector , is the low-dimensional representation obtained after mapping of the identity feature vector , is the adjacency matrix, is the regularization weight; Sub-step 3.3, combine sub-step 3.1 and sub-step 3.2 to add the cross-entropy loss and the manifold regularization loss to obtain the total loss: ， Among them, is the overall loss of the model, ensuring classification accuracy and maintaining the consistency of the low-dimensional manifold structure during the identity feature extraction process.

5. A method for collecting and verifying console login identity information based on deep learning according to claim 1, characterized in that In the above S4, calculating the Mahalanobis distance of the identity feature for identity verification further includes: Sub-step 4.1: Calculate the identity feature vector of the logged-in user based on the low-dimensional manifold embedding representation generated in step S2 and the mean of the identity features of normal users The difference is defined as: ， Among them, is the difference between the identity feature vector of the logged-in user and the mean of the identity features of normal users, is the identity feature vector of the logged-in user, is the mean of the identity features of normal users; Sub-step 4.2, calculate the covariance matrix of the identity feature vector differences , the matrix is used to quantify the variance and covariance of the identity features, and the covariance matrix is defined as: ， Among them, is the identity feature vector, is the number of samples, is the covariance matrix of the samples; Sub-step 4.3, calculate the Mahalanobis distance to measure the distance between the identity characteristics of the logged-in user and the mean of normal users. The Mahalanobis distance is defined as: ， Among them, is the Mahalanobis distance, is the identity feature vector of the logged-in user, is the mean of the identity features of normal users, is the inverse matrix of the covariance matrix, is the transpose operation; Used to calculate the similarity between identity features. If the Mahalanobis distance exceeds the set threshold, it is determined that the login behavior is abnormal and secondary verification is triggered.

6. The method for collecting and verifying console login identity information based on deep learning according to claim 1, wherein, In step S5, further comprising calculating the probability density value of the identity feature: Sub-step 5.1: Embed and represent the identity feature vector of the logged-in user in the low-dimensional manifold generated in step S2 Input it into the Gaussian probability density function to calculate the probability density value. The formula is: ， Among them, is the probability density value of the identity feature vector of the logged-in user, d is the dimension of the low-dimensional manifold space, is the covariance matrix of the determinant, is the covariance matrix of the identity feature vector, is the identity feature vector of the logged-in user, is the mean value of the identity features of normal users; Sub-step 5.2, the probability density value calculated in sub-step 5.1 is compared with the set probability density threshold for comparison, which represents the critical value for judging the probability density of identity features; Sub-step 5.3, combine the probability density value comparison result in sub-step 5.2 with the Mahalanobis distance calculated in step S4 and the set Mahalanobis distance threshold to make a judgment. The decision formula is: ， Among them, is the Mahalanobis distance, represents the Mahalanobis distance threshold for judging the similarity of identity features and is used to make an identity verification decision after integrating the probability density and the Mahalanobis distance result.

7. A method for collecting and verifying console login identity information based on deep learning according to claim 2, characterized in that The extraction of the biometric features includes: using a ResNet-50 network to extract face features, using a MobileNet network to extract fingerprint features, and using an LSTM network to extract voiceprint features; The extraction of the behavior features includes: using Keystroke-Dynamics to extract keyboard input pattern features, and using dynamic time warping to extract mouse trajectory features.

8. A method for collecting and verifying console login identity information based on deep learning according to claim 3, characterized in that, The projection matrix of the topological preserving projection method is obtained by calculating the optimization objective, and the above optimization objective is to minimize the sum of the squared Euclidean distances after projection constrained by the adjacency matrix between all samples.

9. A terminal device, characterized in that, Comprising a sensor module for collecting user biometric and behavior features and a processing module for executing the console login identity information collection and verification method according to claim 1.

10. A storage medium, characterized in that, Stored with an instruction program for controlling the terminal device to collect and verify identity information according to the console login identity information collection and verification method of claim 1.

Images