A method for protecting administrator face recognition data export based on deep learning

Abstract

The present invention discloses a method for protecting administrator face recognition data export based on deep learning, comprising the following steps: S1, starting multimodal biometric verification; S2, dynamically adjusting verification strength and complexity; S3, detecting whether there is abnormal behavior or potential attack; S4, generating potential attack behavior samples, and combining imitation learning technology to train and optimize the abnormal detection module; S5, when the system detects abnormal behavior or potential attack, starting the enhanced active deception module, through inductive data operation and hierarchical virtualization technology, confusing the attacker to enter the virtual environment; S6, in the virtual environment, the system generates and optimizes the induction strategy through the proximal policy optimization algorithm combined with the meta-policy gradient algorithm; S7, the background system records all operation logs, and triggers multi-level security policy adjustment according to abnormal situations. The present invention combines deep learning and multimodal biometric verification, introduces an enhanced active deception module, and enhances the security of data export.

Description

Technical Field

[0001] The present invention relates to the field of information security technology, and in particular to a method for protecting administrator face recognition data export based on deep learning. Background Art

[0002] With the development of informatization and digital transformation, enterprises and organizations have accumulated a large amount of sensitive business data and user privacy data. These data are not only crucial to the business operations of enterprises, but also have great appeal to potential threats from hackers, malicious attackers and insiders. Especially in the data export scenario, since data export involves the migration and flow of data, there is a greater risk of data leakage, and once the data is illegally exported, it may cause serious economic losses and damage to brand reputation. Therefore, how to ensure the security and legitimacy of data during the data export process has become an important issue in the current information security field.

[0003] Existing data export protection technologies usually rely on a single authentication method or permission management strategy. For example, common identity authentication methods include biometric recognition technologies such as password verification, fingerprint recognition, and face recognition. Although these verification methods have a certain degree of security, they are relatively vulnerable to complex attack methods such as replay attacks and forged biometrics. In addition, many systems use static permission management, that is, once the administrator passes the identity authentication, he can access all sensitive data with permissions. The system lacks the ability to dynamically adjust the verification strength according to the sensitivity of the data, which makes the protection ability insufficient when facing potential attacks or unauthorized operations.

[0004] For the detection of abnormal operation behavior, most existing technologies use rule-based detection systems to determine whether the operation is normal through preset behavior rules. However, this method has limited effect when facing complex and changeable attack methods. Attackers can evade rule detection by imitating normal operation modes, further increasing the risk of the system. In addition, existing anomaly detection technologies often only stay at the level of post-event alarm and blocking operations, lacking active induction and confusion mechanisms for attackers' behavior. Once attackers break through the initial protection, they may access real data, causing irreversible losses.

[0005] The introduction of technologies such as Generative Adversarial Networks (GANs) has enhanced the capabilities of anomaly detection systems to a certain extent. In particular, GANs have shown good potential in generating forged samples and conducting adversarial training. However, current applications still have limitations. For example, forged samples generated by GANs are difficult to capture complex attack behavior patterns to a certain extent, especially when combined with operation behaviors under different environmental conditions. Therefore, how to further utilize the characteristics of generative adversarial networks and combine the attacker's historical behavior patterns to make the generated forged samples more realistic and effective is a major problem in existing technologies.

[0006] In the existing technology, the system generally lacks an active defense mechanism after detecting an attack, and most of the protection measures are passive blocking or triggering alarms. This passive protection strategy can easily alert attackers and may cause attackers to take more covert methods to bypass system detection, thereby posing a greater threat to data security. Even in some systems with strong security protection, they mainly rely on means such as operation record auditing and multiple verifications. However, when attackers intentionally evade detection, the system often cannot effectively prevent the further deepening of the attack.

[0007] In addition, the flexibility and adaptability of permission management is another major flaw of the existing system. In actual application scenarios, the sensitivity of data and the permission requirements of administrators may change dynamically depending on the environment and tasks. Most existing permission management systems are static and pre-set, lacking the ability to adaptively adjust administrator permissions and data sensitivity, resulting in a mismatch between verification strength and data sensitivity in some cases, which may cause unnecessary verification redundancy or data leakage risks caused by excessive permissions. The existing dynamic permission management mechanism still relies on relatively simple conditional judgments, which makes it difficult to adapt to complex operating environments and potential threats in real time.

[0008] Therefore, how to provide an administrator face recognition data export protection method based on deep learning is a problem that technical personnel in this field urgently need to solve. Summary of the invention

[0009] One purpose of the present invention is to propose a method for protecting administrator face recognition data export based on deep learning. The present invention combines multimodal biometric verification, generative adversarial networks and active induction defense strategies to provide comprehensive protection against potential security risks in the data export process. The system uses deep learning algorithms to perform identity authentication and dynamic authority management to ensure that the verification strength matches the data sensitivity. At the same time, it confuses potential attackers by generating false data and operation interfaces, and combines reinforcement learning to dynamically optimize the induction strategy. This method has the advantages of accurate identity authentication, flexible authority management, efficient anomaly detection, and comprehensive active protection, which significantly improves the security and robustness of the data export process.

[0010] According to an embodiment of the present invention, a method for protecting administrator face recognition data export based on deep learning includes the following steps:

[0011] S1. When the administrator requests data export, the system starts multimodal biometric verification and uses a multi-task learning neural network to verify the administrator's identity;

[0012] S2. After identity verification is passed, the system dynamically adjusts the verification strength and complexity based on the administrator's permissions and the sensitivity of the requested data;

[0013] S3, the system monitors the administrator's operation behavior, uses the Transformer model to analyze the operation sequence, and detects whether there is abnormal behavior or potential attack;

[0014] S4. Generate potential attack behavior samples through generative adversarial networks, and combine imitation learning technology to learn historical attacker behavior patterns, train and optimize the anomaly detection module;

[0015] S5. When the system detects abnormal behavior or potential attacks, it starts the enhanced active deception module to generate multi-level false data and false operation interfaces, and confuses attackers into entering the virtual environment through inductive data operations and hierarchical virtualization technology;

[0016] S6. In a virtual environment, the system generates and optimizes the induction strategy through the proximal policy optimization algorithm combined with the meta-policy gradient algorithm, dynamically adjusts the induction method according to the attacker's behavior pattern, and records all the attacker's operations;

[0017] S7. The backend system records all operation logs and triggers multi-level security policy adjustments based on abnormal situations, including increasing verification strength, locking accounts or restricting data export permissions, and retaining complete operation records.

[0018] Optionally, the S1 specifically includes:

[0019] S11. When the system starts, it receives the administrator's identity authentication request and collects the administrator's multimodal biometric data in real time, including facial images, fingerprints, irises, and voice data;

[0020] S12, inputting the collected multimodal biometric data into a multi-task learning neural network model, wherein the multi-task learning neural network includes a feature extraction layer and a task branching layer;

[0021] S13, the feature extraction layer uses a convolutional neural network to extract features from each biological feature and generate a feature vector;

[0022] S14, the task branch layer performs classification tasks according to the feature vectors of biological features, the face feature vectors are classified and judged using the fully connected layer, the fingerprint feature vectors are matched using the support vector machine model, the iris feature vectors are compared after adjusting the weights through the attention mechanism, and the voice feature vectors are processed through the long short-term memory network to generate matching results;

[0023] S15. The verification results of the biometric features are comprehensively judged by a weighted fusion method. The weights are dynamically adjusted according to the current environmental factors. The environmental factors include light intensity, sound background and fingerprint clarity:

[0024] R=w 1 ·r 1 +w 2 ·r 2 +…+w n ·r n ;

[0025] Among them, R represents the final authentication result, w i represents the weight of the i-th biometric feature, r i represents the classification result of the i-th biometric feature;

[0026] S16. When the final identity verification result R is greater than the threshold T set by the system r When the administrator identity verification is passed; if R≤T r , the verification fails, the administrator's operation request is rejected, and the security alert mechanism is triggered.

[0027] Optionally, the S2 specifically includes:

[0028] S21, the system receives the permission information of the administrator after the identity authentication is passed, and the permission information includes the administrator's role, permission level, historical operation records and risk assessment of the current operation. The administrator permission matrix is ​​P a ={p ij}, where p ij Indicates that the administrator is in the i-th system operation module at time t j The permission level at the time;

[0029] S22, the system reads the data sensitivity information requested by the administrator to be exported. The data sensitivity changes dynamically according to the data category, encryption level and business importance. The requested data sensitivity matrix is ​​D s ={d kl}, where d kl Indicates that the k-th data at time t l sensitivity;

[0030] S23, the system according to the administrator authority matrix P aand the request data sensitivity matrix D s The multi-dimensional mapping relationship between them is used to calculate the comprehensive adjustment coefficient C:

[0031]

[0032] Among them, γ ik (t j ,t l ) represents the dynamic weight adjustment coefficient between administrator privileges and data sensitivity, p ij represents the administrator's permission level in the i-th system module at time t j The permission value when d kl represents the sensitivity of the kth category of data requested by the administrator at time t l The data sensitivity value at that time, α represents the nonlinear adjustment coefficient of the permission value, β represents the nonlinear adjustment coefficient of the sensitivity value, λ represents the time attenuation coefficient, n represents the number of system permission modules, and m represents the number of data types;

[0033] S24. Verify the strength according to the comprehensive adjustment coefficient C:

[0034]

[0035] Among them, V(C) represents the strength verification function, V 1 Indicates low-intensity verification, V 2 Indicates medium strength verification, V 3 Indicates high-intensity verification, V 4 Indicates the highest strength verification, T 1 , T 2 and T 3 Indicates the system safety threshold;

[0036] S25. The system dynamically adjusts the complexity of the verification process according to the dynamic changes of the operating environment over time, and finally determines the data export scope of the administrator;

[0037] S26, when the comprehensive adjustment coefficient C is lower than the set minimum threshold T 3 When the system rejects the administrator's request to export data, it triggers an alarm and records the abnormal operation behavior.

[0038] Optionally, the S3 specifically includes:

[0039] S31, the system receives and processes the operation sequence log generated by the administrator during the data export process, the operation sequence log includes the operation time, data access frequency, data type and operation device information, the operation sequence is O={o 1 ,o 2 ,…,o N}, where oi represents the administrator’s i-th operation record, and N represents the total number of operations;

[0040] S32, pre-process the operation sequence and extract the time interval of each operation:

[0041] Δt i =t i+1 -t i ;

[0042] Among them, Δt i represents the time interval between the i-th operation and the next operation, t i represents the timestamp of the administrator’s i-th operation, t i+1 Indicates the timestamp of the administrator's i+1th operation;

[0043] S33. The system inputs the preprocessed operation sequence into the Transformer-based anomaly detection model. The Transformer model uses a multi-head self-attention mechanism to capture the temporal dependencies in the operation sequence and calculate the attention weights between operations:

[0044]

[0045] in, represents the attention weight between the i-th operation and the j-th operation under the g-th attention head, represents the query vector of the i-th operation under the g-th attention head, represents the key vector of the jth operation under the gth attention head, d q represents the dimension of the vector, η ij Indicates dynamic adjustment of bias term, and softmax indicates normalization function;

[0046] S34. Aggregate the operation dependencies under all attention heads into an operation feature vector. The model merges the outputs of multiple heads through a weighted multi-head attention mechanism to generate the final operation feature vector:

[0047]

[0048] Among them, Y i represents the final aggregated feature vector of the ith operation, G represents the number of attention heads, and V g represents the learnable weight of each attention head, Represents the value vector under the g-th attention head;

[0049] S35. According to Y i and Δt i , the system generates an abnormal behavior score for the operation:

[0050]

[0051] Among them, S op represents the abnormal behavior score of the entire operation sequence, B ij represents the attention weight between operations, κ i represents the adaptive adjustment factor of the ith operation, ζ represents the time decay coefficient, t i and t j Indicates the timestamp of the operation;

[0052] S36, when abnormal behavior score S op Exceeding the set threshold T op When the score is lower than the threshold, the system determines that the operation is abnormal, triggers the protection mechanism and records the abnormal operation log; if the score is lower than the threshold, the operation is considered normal and the system continues to allow data export operations.

[0053] Optionally, the S4 specifically includes:

[0054] S41. The system extracts normal operation and attack behavior features based on the administrator's operation log and historical attack behavior data. The normal behavior feature set is N = {n 1 ,n 2 ,…,n p}, where n i represents the i-th normal behavior feature, and the attack behavior feature set is A = {a 1 ,a 2 ,…,a q}, where a j represents the characteristics of the j-th attack behavior;

[0055] S42, the system inputs the attack behavior features into the generative adversarial network, which consists of a generator G and a discriminator D. The generator generates a random noise z 0 Generate potential attack behavior samples G(z):

[0056] G(z)=G(θ g ,z 0 );

[0057] Among them, θ g represents the learnable parameters of the generator, z 0 A random noise vector representing the input to the generator;

[0058] S43, the imitation learning part uses the historical attack behavior dataset P mimic The characteristic operation mode M in 1 ,m 2 ,…,m r} to train the generator’s output G(z) to be close to the real attack pattern:

[0059]

[0060] Among them, L G represents the loss function of the generator, E represents the mathematical expectation, P z represents the random noise distribution, D(G(z)) represents the output probability of the discriminator for the generated sample, λ g represents the adjustment coefficient of imitation learning, ∥G(z)-m∥ 2 represents the difference between the generated sample and the imitated attack behavior, and m represents the historical attack behavior pattern;

[0061] S44, discriminator D is used to distinguish between real attack behaviors and fake attack behaviors, and the optimization goal is to maximize the recognition probability of real attack samples:

[0062]

[0063] Among them, L D represents the loss function of the discriminator, P data represents the distribution of real attack behaviors, and x represents real attack samples;

[0064] S45. Generate and optimize potential attack behavior samples through generative adversarial networks. The system continuously trains the anomaly detection module. The anomaly detection module improves its ability to detect complex and hidden attack behaviors by learning the generated attack samples. When a potential attack behavior is detected, the protection mechanism is triggered to record the attack operation and update the model.

[0065] S46. Based on the continuously updated operation log, the system continuously optimizes the generation capability of the generator by combining generative adversarial networks with imitation learning technology.

[0066] Optionally, the S5 specifically includes:

[0067] S51. When the system detects abnormal behavior or potential attack, the enhanced active deception module is activated to generate multi-level false data and false operation interfaces. The false data feature set is F = {f 1 ,f 2 ,…,f m}, where f i Represents the characteristics of the i-th type of false data:

[0068] f i =g(d i ,θ f );

[0069] Among them, d i represents the real data feature of the i-th category, θ f represents the transformation parameters for generating false data, g(·) represents the data transformation function;

[0070] S52, the system generates a false operation interface through virtualization technology to induce the attacker to operate. The set of virtual operation interfaces is V = {v 1 ,v 2 ,…,v k}, where v j Represents the jth virtual operation interface:

[0071] v j =h(u j ,θ v );

[0072] Among them, u j represents the operation behavior of the jth attacker, θ v represents the transformation parameter for generating the fake interface, h(·) represents the mapping function between the attacker's operation and the fake interface;

[0073] S53, the system dynamically generates induced operations through the induced data operation mechanism, and the induced operation set is I = {i 1 ,i 2 ,…,i e}, where i l Indicates the first induced operation, and calculates the ineffectiveness score of the induced operation:

[0074]

[0075] Among them, S i represents the ineffectiveness score of the induced operation, e represents the total number of induced operations, α l represents the weight of the first induction operation, It indicates the result of the first fake operation performed by the attacker in the fake environment. Indicates the first real operation result, μ l Represents the adaptive adjustment coefficient, Δt l Indicates the time interval between the first operation and the last operation, γ l represents the time attenuation coefficient;

[0076] S54, the system scores S according to the ineffectiveness of the induced operation i Dynamically adjust the complexity of fake data and fake interface, and dynamically optimize the fake environment as the score increases:

[0077]

[0078] Among them, C f represents the complexity of the false environment, m represents the total number of false data types, ρ i represents the weight of the i-th type of false data, δ irepresents the adaptive complexity adjustment coefficient, and T represents the scoring threshold set by the system;

[0079] S55, the system dynamically adjusts the characteristics of the fake environment according to the attacker's behavior and operation records. By recording the operation log, the system compares the attacker's fake operation records and the corresponding real operation records Calculate the matching score M r :

[0080]

[0081] Among them, θ q Represents the weight of the qth operation.

[0082] Optionally, the S6 specifically includes:

[0083] S61. The system monitors the attacker's operation behavior in a virtual environment and collects behavior data in real time. The system initializes the induction strategy according to the operation sequence and generates the initial strategy using the proximal strategy optimization algorithm.

[0084] S62. The system dynamically optimizes the initial strategy through the meta-policy gradient algorithm. The meta-policy gradient algorithm adjusts the induced strategy in real time according to the attacker's operation mode:

[0085]

[0086] Among them, L Meta (θ) represents the loss function of the meta-policy gradient algorithm, θ k represents the parameters of the k-th strategy, represents the kth strategy, Represents the parameter θ k The gradient of, E represents the mathematical expectation, R t represents the reward function at time step t, and T represents the total length of the time step;

[0087] S63. The system continuously optimizes the inducement strategy based on the feedback from the attacker’s operation behavior:

[0088]

[0089] Among them, F t represents the feedback value at time step t, r i represents the reward value of the attacker's ith operation in the fake environment, λ 3 represents the time interval sensitivity coefficient, γ f represents the discount factor, t i -t i-1 Indicates the time difference between two operations;

[0090] S64. Based on the feedback function and the results of meta-strategy gradient optimization, the system dynamically adjusts the induction strategy:

[0091]

[0092] Among them, θ′ represents the updated policy parameter, θ represents the policy parameter before the update, and τ represents the learning rate;

[0093] S65. The system dynamically adjusts the induced operation in the fake environment according to the optimized strategy parameters, and generates new induced interface and data according to the attacker's operation mode.

[0094] The beneficial effects of the present invention are:

[0095] First, the present invention greatly improves the security and accuracy of identity authentication through multimodal biometric verification. Traditional identity authentication usually relies on a single biometric feature and is vulnerable to counterfeit attacks or restrictions in specific environments. The present invention uses a multi-task learning neural network to simultaneously verify multiple biometric features of the administrator, such as face, fingerprint, iris, and voice, and dynamically adjusts the weight of each biometric feature according to changes in the current environment. In this way, the system can still ensure efficient and accurate identity authentication under adverse conditions such as insufficient light, sound interference, or blurred fingerprints, enhancing the robustness and flexibility of the verification process.

[0096] Secondly, the dynamic permission management mechanism of the present invention further ensures the security of the data export process. Compared with the static, pre-set permission management mode in the prior art, the present invention calculates the verification strength and complexity in real time through a deep learning algorithm based on the administrator's permission information and the sensitivity of the requested data, ensuring that the administrator's operating permissions match the sensitivity level of the data. This method of dynamically adjusting the verification strength can effectively prevent administrators from unauthorized access to highly sensitive data, thereby reducing the risk of data leakage. At the same time, the system can flexibly adjust the permission verification process according to changes in the operating environment, so that security and operational convenience are well balanced.

[0097] In terms of anomaly detection, the present invention significantly improves the ability to detect potential attack behaviors by introducing generative adversarial networks (GANs) and imitation learning technology. Traditional anomaly detection often relies on preset rules or simple behavior analysis models, which are difficult to deal with complex and hidden attack methods. The present invention generates forged samples of potential attack behaviors through generative adversarial networks, and further optimizes the training process of the anomaly detection module by imitating the operation mode of historical attackers. As a result, the system can more accurately identify those attack behaviors that attempt to imitate normal operations, ensuring that any potential threats can be identified and processed in a timely manner during the data export process.

[0098] In addition, a notable feature of the present invention is the introduction of an enhanced active deception module. When the system detects that an attacker has potential abnormal behavior, it does not simply block or alarm. The system will generate multi-level false data and false operation interfaces to induce the attacker to enter the virtual environment. This active deception mechanism generates false data and operation processes, causing the attacker to mistakenly believe that he has obtained valuable information, thereby avoiding further attempts to access real data. By dynamically adjusting the complexity of the virtual environment, the system can keep the attacker confused in the false data, and by recording his behavior in the virtual environment, it provides an important reference for the subsequent optimization of protection strategies. This mechanism not only enhances the security of the data, but also avoids the problem of alerting the attacker in traditional protection measures.

[0099] Furthermore, the proximal policy optimization algorithm of the present invention is combined with the meta-policy gradient algorithm to dynamically generate and optimize the induction strategy in a false environment, so that the system can make real-time adjustments according to the attacker's behavior pattern, thereby improving the flexibility of defense. The system can continuously learn and optimize through feedback on the attacker's operational behavior, thereby ensuring the continued effectiveness of the induction strategy. This feature makes the system highly adaptable when facing unknown or complex attacks, and can take the most appropriate protective measures according to different attack behaviors.

[0100] Finally, the multi-level security policy adjustment mechanism of the present invention can effectively respond to security threats at different levels. The system triggers the corresponding security policy based on the administrator's operation log and the detection results of abnormal behavior, from simple operation records to improving verification strength, until locking accounts or limiting data export permissions, responding to different levels of threats step by step, ensuring that the system can effectively protect data security under any circumstances. This multi-level response mechanism is relatively rare in the prior art and effectively fills the gap in the security management system in dynamically responding to complex attack behaviors. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0102] Figure 1 This is a flowchart of a method for exporting and protecting administrator face recognition data based on deep learning proposed by the present invention;

[0103] Figure 2 A schematic diagram of the structure of a multimodal biometric verification method for administrator face recognition data export protection method based on deep learning proposed by the present invention;

[0104] Figure 3This is a schematic diagram of the generation of false data and false operation interface and active induction strategy for the deep learning-based administrator face recognition data export protection method proposed in the present invention. DETAILED DESCRIPTION

[0105] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0106] refer to Figure 1-3 , a method for protecting administrator face recognition data export based on deep learning, comprising the following steps:

[0107] S1. When the administrator requests data export, the system starts multimodal biometric verification and uses a multi-task learning neural network to verify the administrator's identity;

[0108] S2. After identity verification is passed, the system dynamically adjusts the verification strength and complexity based on the administrator's permissions and the sensitivity of the requested data;

[0109] S3, the system monitors the administrator's operation behavior, uses the Transformer model to analyze the operation sequence, and detects whether there is abnormal behavior or potential attack;

[0110] S4. Generate potential attack behavior samples through generative adversarial networks, and combine imitation learning technology to learn historical attacker behavior patterns, train and optimize the anomaly detection module;

[0111] S5. When the system detects abnormal behavior or potential attacks, it starts the enhanced active deception module to generate multi-level false data and false operation interfaces, and confuses attackers into entering the virtual environment through inductive data operations and hierarchical virtualization technology;

[0112] S6. In a virtual environment, the system generates and optimizes the induction strategy through the proximal policy optimization algorithm combined with the meta-policy gradient algorithm, dynamically adjusts the induction method according to the attacker's behavior pattern, and records all the attacker's operations;

[0113] S7. The backend system records all operation logs and triggers multi-level security policy adjustments based on abnormal situations, including increasing verification strength, locking accounts or restricting data export permissions, and retaining complete operation records.

[0114] In this implementation, S1 specifically includes:

[0115] S11. When the system starts, it receives the administrator's identity authentication request and collects the administrator's multimodal biometric data in real time, including facial images, fingerprints, irises, and voice data;

[0116] S12, inputting the collected multimodal biometric data into a multi-task learning neural network model, wherein the multi-task learning neural network includes a feature extraction layer and a task branching layer;

[0117] S13, the feature extraction layer uses a convolutional neural network to extract features from each biological feature and generate a feature vector;

[0118] S14, the task branch layer performs classification tasks according to the feature vectors of biological features, the face feature vectors are classified and judged using the fully connected layer, the fingerprint feature vectors are matched using the support vector machine model, the iris feature vectors are compared after adjusting the weights through the attention mechanism, and the voice feature vectors are processed through the long short-term memory network to generate matching results;

[0119] S15. The verification results of the biometric features are comprehensively judged by a weighted fusion method. The weights are dynamically adjusted according to the current environmental factors. The environmental factors include light intensity, sound background and fingerprint clarity:

[0120] R=w 1 ·r 1 +w 2 ·r 2 +…+w n ·r n ;

[0121] Among them, R represents the final authentication result, w i represents the weight of the i-th biometric feature, r i represents the classification result of the i-th biometric feature;

[0122] S16. When the final identity verification result R is greater than the threshold T set by the system r When the administrator identity verification is passed; if R≤T r , the verification fails, the administrator's operation request is rejected, and the security alert mechanism is triggered.

[0123] In this implementation, S2 specifically includes:

[0124] S21, the system receives the permission information of the administrator after the identity authentication is passed, and the permission information includes the administrator's role, permission level, historical operation records and risk assessment of the current operation. The administrator permission matrix is ​​P a ={p ij}, where p ij Indicates that the administrator is in the i-th system operation module at time t j The permission level at the time;

[0125] S22, the system reads the data sensitivity information requested by the administrator to be exported. The data sensitivity changes dynamically according to the data category, encryption level and business importance. The requested data sensitivity matrix is ​​D s ={d kl}, where d kl Indicates that the k-th data at time t l sensitivity;

[0126] S23, the system according to the administrator authority matrix P a and the request data sensitivity matrix D s The multi-dimensional mapping relationship between them is used to calculate the comprehensive adjustment coefficient C:

[0127]

[0128] Among them, γ ik (t j ,t l ) represents the dynamic weight adjustment coefficient between administrator privileges and data sensitivity, p ij represents the administrator's permission level in the i-th system module at time t j The permission value when d kl represents the sensitivity of the kth category of data requested by the administrator at time t l The data sensitivity value at that time, α represents the nonlinear adjustment coefficient of the permission value, β represents the nonlinear adjustment coefficient of the sensitivity value, λ represents the time attenuation coefficient, n represents the number of system permission modules, and m represents the number of data types;

[0129] S24. Verify the strength according to the comprehensive adjustment coefficient C:

[0130]

[0131] Among them, V(C) represents the strength verification function, V 1 Indicates low-intensity verification, V 2 Indicates medium strength verification, V 3 Indicates high-intensity verification, V 4 Indicates the highest strength verification, T 1 , T 2 and T 3 Indicates the system safety threshold;

[0132] S25. The system dynamically adjusts the complexity of the verification process according to the dynamic changes of the operating environment over time, and finally determines the data export scope of the administrator;

[0133] S26, when the comprehensive adjustment coefficient C is lower than the set minimum threshold T 3 When the system rejects the administrator's request to export data, it triggers an alarm and records the abnormal operation behavior.

[0134] In this implementation, S3 specifically includes:

[0135] S31, the system receives and processes the operation sequence log generated by the administrator during the data export process, the operation sequence log includes the operation time, data access frequency, data type and operation device information, the operation sequence is O={o 1 ,o 2 ,…,o N}, where o i represents the administrator’s i-th operation record, and N represents the total number of operations;

[0136] S32, pre-process the operation sequence and extract the time interval of each operation:

[0137] Δt i =t i+1 -t i ;

[0138] Among them, Δt i represents the time interval between the i-th operation and the next operation, t i represents the timestamp of the administrator’s i-th operation, t i+1 Indicates the timestamp of the administrator's i+1th operation;

[0139] S33. The system inputs the preprocessed operation sequence into the Transformer-based anomaly detection model. The Transformer model uses a multi-head self-attention mechanism to capture the temporal dependencies in the operation sequence and calculate the attention weights between operations:

[0140]

[0141] in, represents the attention weight between the i-th operation and the j-th operation under the g-th attention head, represents the query vector of the i-th operation under the g-th attention head, represents the key vector of the jth operation under the gth attention head, d q represents the dimension of the vector, η ij Indicates dynamic adjustment of bias term, and softmax indicates normalization function;

[0142] S34. Aggregate the operation dependencies under all attention heads into an operation feature vector. The model merges the outputs of multiple heads through a weighted multi-head attention mechanism to generate the final operation feature vector:

[0143]

[0144] Among them, Yi represents the final aggregated feature vector of the ith operation, G represents the number of attention heads, and V g represents the learnable weight of each attention head, Represents the value vector under the g-th attention head;

[0145] S35. According to Y i and Δt i , the system generates an abnormal behavior score for the operation:

[0146]

[0147] Among them, S op represents the abnormal behavior score of the entire operation sequence, B ij represents the attention weight between operations, κ i represents the adaptive adjustment factor of the ith operation, ζ represents the time decay coefficient, t i and t j Indicates the timestamp of the operation;

[0148] S36, when abnormal behavior score S op Exceeding the set threshold T op When the score is lower than the threshold, the system determines that the operation is abnormal, triggers the protection mechanism and records the abnormal operation log; if the score is lower than the threshold, the operation is considered normal and the system continues to allow data export operations.

[0149] In this implementation, S4 specifically includes:

[0150] S41. The system extracts normal operation and attack behavior features based on the administrator's operation log and historical attack behavior data. The normal behavior feature set is N = {n 1 ,n 2 ,…,n p}, where n i represents the i-th normal behavior feature, and the attack behavior feature set is A = {a 1 ,a 2 ,…,a q}, where a j represents the characteristics of the j-th attack behavior;

[0151] S42, the system inputs the attack behavior features into the generative adversarial network, which consists of a generator G and a discriminator D. The generator generates a random noise z 0 Generate potential attack behavior samples G(z):

[0152] G(z)=G(θ g ,z 0 );

[0153] Among them, θ grepresents the learnable parameters of the generator, z 0 A random noise vector representing the input to the generator;

[0154] S43, the imitation learning part uses the historical attack behavior dataset P mimic The characteristic operation mode M in 1 ,m 2 ,…,m r} to train the generator’s output G(z) to be close to the real attack pattern:

[0155]

[0156] Among them, L G represents the loss function of the generator, E represents the mathematical expectation, P z represents the random noise distribution, D(G(z)) represents the output probability of the discriminator for the generated sample, λ g represents the adjustment coefficient of imitation learning, ∥G(z)-m∥ 2 represents the difference between the generated sample and the imitated attack behavior, and m represents the historical attack behavior pattern;

[0157] S44, discriminator D is used to distinguish between real attack behaviors and fake attack behaviors, and the optimization goal is to maximize the recognition probability of real attack samples:

[0158]

[0159] Among them, L D represents the loss function of the discriminator, P data represents the distribution of real attack behaviors, and x represents real attack samples;

[0160] S45. Generate and optimize potential attack behavior samples through generative adversarial networks. The system continuously trains the anomaly detection module. The anomaly detection module improves its ability to detect complex and hidden attack behaviors by learning the generated attack samples. When a potential attack behavior is detected, the protection mechanism is triggered to record the attack operation and update the model.

[0161] S46. Based on the continuously updated operation log, the system continuously optimizes the generation capability of the generator by combining generative adversarial networks with imitation learning technology.

[0162] Optionally, the S5 specifically includes:

[0163] S51. When the system detects abnormal behavior or potential attack, the enhanced active deception module is activated to generate multi-level false data and false operation interfaces. The false data feature set is F = {f 1 ,f 2 ,…,f m}, where fi Represents the characteristics of the i-th type of false data:

[0164] f i =g(d i ,θ f );

[0165] Among them, d i represents the real data feature of the i-th category, θ f represents the transformation parameters for generating false data, g(·) represents the data transformation function;

[0166] S52, the system generates a false operation interface through virtualization technology to induce the attacker to operate. The set of virtual operation interfaces is V = {v 1 ,v 2 ,…,v k}, where v j Represents the jth virtual operation interface:

[0167] v j =h(u j ,θ v );

[0168] Among them, u j represents the operation behavior of the jth attacker, θ v represents the transformation parameter for generating the fake interface, h(·) represents the mapping function between the attacker's operation and the fake interface;

[0169] S53, the system dynamically generates induced operations through the induced data operation mechanism, and the induced operation set is I = {i 1 ,i 2 ,…,i e}, where i l Indicates the first induced operation, and calculates the ineffectiveness score of the induced operation:

[0170]

[0171] Among them, S i represents the ineffectiveness score of the induced operation, e represents the total number of induced operations, α l represents the weight of the first induction operation, It indicates the result of the first fake operation performed by the attacker in the fake environment. Indicates the first real operation result, μ l Represents the adaptive adjustment coefficient, Δt l Indicates the time interval between the first operation and the last operation, γ l represents the time attenuation coefficient;

[0172] S54, the system scores S according to the ineffectiveness of the induced operation iDynamically adjust the complexity of fake data and fake interface, and dynamically optimize the fake environment as the score increases:

[0173]

[0174] Among them, C f represents the complexity of the false environment, m represents the total number of false data types, ρ i represents the weight of the i-th type of false data, δ i represents the adaptive complexity adjustment coefficient, and T represents the scoring threshold set by the system;

[0175] S55, the system dynamically adjusts the characteristics of the fake environment according to the attacker's behavior and operation records. By recording the operation log, the system compares the attacker's fake operation records and the corresponding real operation records Calculate the matching score M r :

[0176]

[0177] Among them, θ q Represents the weight of the qth operation.

[0178] In this implementation manner, S6 specifically includes:

[0179] S61. The system monitors the attacker's operation behavior in a virtual environment and collects behavior data in real time. The system initializes the induction strategy according to the operation sequence and generates the initial strategy using the proximal strategy optimization algorithm.

[0180] S62. The system dynamically optimizes the initial strategy through the meta-policy gradient algorithm. The meta-policy gradient algorithm adjusts the induced strategy in real time according to the attacker's operation mode:

[0181]

[0182] Among them, L Meta (θ) represents the loss function of the meta-policy gradient algorithm, θ k represents the parameters of the k-th strategy, represents the kth strategy, Represents the parameter θ k The gradient of, E represents the mathematical expectation, R t represents the reward function at time step t, and T represents the total length of the time step;

[0183] S63. The system continuously optimizes the inducement strategy based on the feedback from the attacker’s operation behavior:

[0184]

[0185] Among them, Ft represents the feedback value at time step t, r i represents the reward value of the attacker's ith operation in the fake environment, λ 3 represents the time interval sensitivity coefficient, γ f represents the discount factor, t i -t i-1 Indicates the time difference between two operations;

[0186] S64. Based on the feedback function and the results of meta-strategy gradient optimization, the system dynamically adjusts the induction strategy:

[0187]

[0188] Among them, θ′ represents the updated policy parameter, θ represents the policy parameter before the update, and τ represents the learning rate;

[0189] S65. The system dynamically adjusts the induced operation in the fake environment according to the optimized strategy parameters, and generates new induced interface and data according to the attacker's operation mode.

[0190] Embodiment 1:

[0191] In order to verify the feasibility of the present invention in implementation, the present invention is applied to a large financial institution. The large financial institution has repeatedly faced the risk of internal administrators exporting highly sensitive data without sufficient verification or incorrect permission control. In particular, in the process of permission management and data export, the existing security protection system often cannot effectively prevent potential unauthorized operations by internal personnel or external malicious attacks.

[0192] In this financial institution, the administrator is responsible for exporting sensitive data from various departments of the company and uploading it to the headquarters server for analysis on a regular basis. This operation involves multiple key steps, including identity authentication, data permission control, and the export of sensitive data. However, the traditional single authentication mechanism and static permission management method cannot adapt to the dynamically changing business needs and security threats in the financial environment. Therefore, the financial institution decided to introduce the deep learning-based administrator face recognition data export protection method of the present invention, aiming to solve the security vulnerabilities in current identity authentication, dynamic permission management, and abnormal operation detection.

[0193] After applying the invention, the system first performs multimodal biometric verification on the administrator's identity. Whenever the administrator requests to export data, the system collects the administrator's facial image, fingerprint, iris and voice features through the camera, and inputs these features into the multi-task learning neural network for verification in real time. In a typical business operation scenario, an administrator needs to export the transaction data of a VIP customer. During the operation, the system extracts features from the facial image, fingerprint and iris, and classifies them through a convolutional neural network. At the same time, based on environmental changes (such as insufficient light or sound interference), the system dynamically adjusts the weights of different modal biometric features. For example, when the light is poor, fingerprint and iris features will have a larger weight, and when the facial features are interfered with, the weight of the voice features will increase. This multimodal fusion verification method effectively prevents attacks by forged biometrics.

[0194] By analyzing the test data of the financial institution, it was found that the system processed 632 administrators' data export requests in one week. Among them, 45 verification requests failed under single-modal verification conditions due to fuzzy or unrecognizable biometric features. However, through the multimodal weighted fusion verification technology of the present invention, the system successfully passed 38 verifications in complex environments, with a verification success rate of 84.4%. Before the introduction of the present invention, the pass rate of traditional single biometric verification methods in similar scenarios was only 65.8%. This data fully demonstrates the effectiveness of multimodal verification, especially in harsh environmental conditions, which significantly improves the robustness and accuracy of verification.

[0195] When the administrator passes the verification, the system will dynamically adjust the verification strength according to his or her authority and the sensitivity of the requested exported data. In the authority management system of the financial institution, the administrator's authority level and the sensitivity of the exported data will be mapped to the authority matrix and the data sensitivity matrix. For example, when exporting low-sensitivity data (such as ordinary customer data), the system only requires double verification of face and fingerprint, and when it comes to exporting highly sensitive data (such as VIP customer transaction data or financial statements), the system will require the administrator to pass multiple verifications, including comprehensive verification of face, fingerprint, iris and voice. Through the reinforcement learning mechanism, the system can also automatically adjust the verification strength according to the risk level of the operation. For example, in a VIP customer transaction data export operation, the system detected that the administrator's operation log was deviated from the historical record. The system proactively increased the verification strength from double verification to multiple verification to ensure the security of the data export process.

[0196] In a one-month data test, there were 165 requests for exporting highly sensitive data. The system dynamically adjusted the verification strength according to the sensitivity of the data, and automatically upgraded the verification strength in 37 operations, avoiding potential unauthorized operations. Traditional static permission management systems are almost unable to dynamically adjust the complexity of verification in similar environments, which can easily lead to overly broad or narrow permission management, thereby increasing security risks. The system's dynamic permission adjustment mechanism makes the export of highly sensitive data safer and performs well in sensitive data management within financial institutions.

[0197] In addition, the system generates potential attack samples in the background by combining the generative adversarial network (GAN) technology with the imitation learning technology, and trains the anomaly detection module with the data of historical attack behaviors. When the system detects potential abnormal behavior, it automatically generates false data and false operation interfaces to confuse the attacker. In a simulated network attack, the attacker attempted to export highly sensitive data by disguising himself as an internal administrator. The system identified abnormal behavior by analyzing his operation behavior, and automatically generated false transaction data and a false export interface to induce the attacker to enter the virtual environment. The attacker mistakenly believed that he had successfully obtained the real data, but all operations in the virtual environment were recorded and used as the basis for the subsequent optimization of the protection strategy. After testing, the system successfully induced the attacker to enter the false environment in 15 simulated attack scenarios and recorded all operations. The introduction of this active defense strategy effectively prevents the risk of external malicious attackers obtaining real data.

[0198] Through data analysis after the implementation of this invention, the data export protection of financial institutions has been significantly improved. During the one-month application period, the system processed a total of 1,021 administrator data export requests and successfully avoided 11 potential data leakage risks. Through multiple measures such as active induction, anomaly detection, and dynamic permission adjustment, the system minimizes security threats in the data export process, significantly improving the security and reliability of financial institutions in data management.

[0199] Table 1 Comparison of data export security between the traditional system and the system of the present invention

[0200] Comparison Items Traditional systems System of the present invention Total verification requests 632 632 Single-mode verification success rate 65.8％ - Multimodal verification success rate - 84.4％ Number of privilege escalations 0 37 Number of potential data breaches 11 0 Number of simulated attacks successfully blocked 5 15

[0201] In the above Table 1, we can clearly see that the system of the present invention has significantly improved over the traditional system in multiple key indicators. First, the total number of verification requests in the table remains consistent, and both systems processed 632 administrator data export requests under the same verification environment. However, in terms of verification success rate, the system of the present invention is significantly better than the traditional system.

[0202] For traditional systems, the verification process relies on single-modal verification technology, and its success rate is only 65.8%. This verification method is prone to verification failure when encountering unfavorable environments (such as insufficient light, unclear fingerprints, etc.). In the system of the present invention, multimodal biometric verification is adopted, and multiple biometric features such as face, fingerprint, iris, voice, etc. are weighted and fused, and the weight of each feature is dynamically adjusted through a deep learning algorithm, which successfully improves the verification success rate to 84.4%. This shows that the identity authentication of the present invention in complex environments is more reliable, and the verification robustness has been significantly enhanced.

[0203] In terms of the number of times the authority is upgraded, the traditional system does not have the ability to dynamically adjust the verification strength according to the sensitivity of the data, so the verification strength was not upgraded in all 632 operations. In contrast, the system of the present invention uses a reinforcement learning mechanism to automatically upgrade the verification strength in 37 high-risk data export requests by evaluating the sensitivity of the data and the operational risk in real time. Through this mechanism, the system can effectively prevent administrators from exporting highly sensitive data when operating beyond their authority, thus ensuring the security of the data.

[0204] Another key difference is the number of potential data leaks. The traditional system detected 11 potential data leaks in a month, while the system of the present invention detected 0. This result is due to the dynamic permission management and abnormal operation detection mechanism introduced by the present invention. The system can not only adjust the verification strength in real time, but also detect abnormal behavior and proactively prevent potential data leaks.

[0205] The number of successful blocks of simulated attacks is also a significantly improved indicator. The traditional system only successfully blocked 5 attacks when facing simulated attacks, while the system of the present invention successfully blocked 15 simulated attacks by generating adversarial networks (GANs) and actively inducing defense strategies. The present invention induces attackers to enter a virtual environment and records their operating behaviors by generating false data and false operation interfaces, thereby effectively preventing attackers from obtaining real data. This mechanism shows that the present invention can more actively defend and confuse attackers when facing complex external attacks, greatly improving the security of the system.

[0206] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for protecting administrator face recognition data export based on deep learning, characterized in that: The steps include: S1. When the administrator requests data export, the system based on the deep learning administrator face recognition data export protection method starts multimodal biometric verification and uses a multi-task learning neural network to verify the administrator's identity; S2. After identity authentication is passed, the system dynamically adjusts the strength and complexity of the authentication based on the administrator's authority and the sensitivity of the requested data; S3, the system monitors the administrator's operation behavior, uses the Transformer model to analyze the operation sequence, and detects whether there is abnormal behavior or potential attack; S4. Generate potential attack behavior samples through generative adversarial networks, and combine imitation learning technology to learn historical attacker behavior patterns, train and optimize the anomaly detection module; S5. When the system detects abnormal behavior or potential attack, the enhanced active deception module is activated to generate multi-level false data and false operation interfaces, and confuse the attacker into entering the virtual environment through inductive data operation and hierarchical virtualization technology; S6. In a virtual environment, the system generates and optimizes the induction strategy by combining the proximal policy optimization algorithm with the meta-policy gradient algorithm, dynamically adjusts the induction method according to the attacker's behavior pattern, and records all the attacker's operation behaviors; S7. The system records all operation logs and triggers multi-level security policy adjustments based on abnormal situations, including increasing verification strength, locking accounts or limiting data export permissions, and retaining complete operation records; The S5 specifically includes: S51, when the system detects abnormal behavior or potential attack, the enhanced active deception module is activated to generate multi-level false data and false operation interface; S52, the system generates a false operation interface through virtualization technology to induce the attacker to perform operations; S53, the system dynamically generates an induced operation through an induced data operation mechanism, and calculates an invalidity score of the induced operation; S54, the system scores S according to the ineffectiveness of the induced operation i Dynamically adjust the complexity of fake data and fake interfaces, and dynamically optimize the fake environment as the score increases; S55. The system dynamically adjusts the characteristics of the fake environment according to the attacker's behavior and operation records. By recording the operation logs, the system calculates a matching score by comparing the attacker's fake operation records with the corresponding real operation records.

2. According to the deep learning-based administrator face recognition data export protection method of claim 1, it is characterized in that: The S1 specifically includes: S11, when the system is started, receiving an identity authentication request from an administrator, and collecting multimodal biometric data of the administrator in real time, including facial images, fingerprints, irises, and voice data; S12, inputting the collected multimodal biometric data into a multi-task learning neural network model, wherein the multi-task learning neural network includes a feature extraction layer and a task branching layer; S13, the feature extraction layer uses a convolutional neural network to extract features from each biological feature and generate a feature vector; S14, the task branch layer performs classification tasks according to the feature vectors of biological features, the face feature vectors are classified and judged using the fully connected layer, the fingerprint feature vectors are matched using the support vector machine model, the iris feature vectors are compared after adjusting the weights through the attention mechanism, and the voice feature vectors are processed through the long short-term memory network to generate matching results; S15. The verification results of the biometric features are comprehensively judged by a weighted fusion method. The weights are dynamically adjusted according to the current environmental factors. The environmental factors include light intensity, sound background and fingerprint clarity: R=w1·r1+w2·r2+…+w n ·r n ; Among them, R represents the final authentication result, w i represents the weight of the i-th biometric feature, r i represents the classification result of the i-th biometric feature; S16. When the final identity verification result R is greater than the threshold value T set by the system r When the administrator identity verification is passed; if R≤T r , the verification fails, the administrator's operation request is rejected, and the security alert mechanism is triggered.

3. According to the deep learning-based administrator face recognition data export protection method of claim 1, it is characterized in that: The S2 specifically includes: S21, the system receives the permission information of the administrator after the identity authentication is passed, and the permission information includes the administrator's role, permission level, historical operation records and risk assessment of the current operation. The administrator permission matrix is ​​P a ={p ij }, where p ij Indicates that the administrator is in the i-th system operation module at time t j The permission level at the time; S22, the system reads the data sensitivity information requested by the administrator to be exported. The data sensitivity changes dynamically according to the data category, encryption level and business importance. The requested data sensitivity matrix is ​​D s ={d kl }, where d kl Indicates that the k-th data at time t l sensitivity; S23, the system according to the administrator authority matrix P a and the request data sensitivity matrix D s The multi-dimensional mapping relationship between them is used to calculate the comprehensive adjustment coefficient C: Among them, γ ik (t j ,t l ) represents the dynamic weight adjustment coefficient between administrator privileges and data sensitivity, p ij Indicates that the administrator is in the i-th system operation module at time t j The permission level at that time, d kl Indicates that the k-th data at time t l α represents the nonlinear adjustment coefficient of the authority value, β represents the nonlinear adjustment coefficient of the sensitivity value, λ represents the time attenuation coefficient, n represents the number of the system authority modules, and m represents the number of data types; S24. Verify the strength according to the comprehensive adjustment coefficient C: Wherein, V(C) represents the strength verification function, V1 represents low strength verification, V2 represents medium strength verification, V3 represents high strength verification, V4 represents the highest strength verification, and T1, T2 and T3 represent the system safety thresholds; S25, the system dynamically adjusts the complexity of the verification process according to the dynamic change of the operating environment over time function, and finally determines the data export range of the administrator; S26. When the comprehensive adjustment coefficient C is lower than the set minimum threshold value T3, the system rejects the administrator's data export request, triggers an alarm, and records abnormal operation behavior.

4. According to a method for protecting administrator face recognition data export based on deep learning according to claim 1, it is characterized in that: The S3 specifically includes: S31, the system receives and processes the operation sequence log generated by the administrator during the data export process, the operation sequence log includes the operation time, data access frequency, data type and operation device information, the operation sequence is O = {o1, o2, ..., o N }, where o i represents the administrator’s i-th operation record, and N represents the total number of operations; S32, pre-process the operation sequence and extract the time interval of each operation: Δt i =t i+1 -t i ; Among them, Δt i represents the time interval between the i-th operation and the next operation, t i represents the timestamp of the administrator’s i-th operation, t i+1 Indicates the timestamp of the administrator's i+1th operation; S33, the system inputs the preprocessed operation sequence into the Transformer-based anomaly detection module. The Transformer model uses a multi-head self-attention mechanism to capture the temporal dependencies in the operation sequence and calculate the attention weights between operations: in, represents the attention weight between the i-th operation and the j-th operation under the g-th attention head, represents the query vector of the i-th operation under the g-th attention head, represents the key vector of the jth operation under the gth attention head, d q represents the dimension of the vector, η ij Indicates dynamic adjustment of bias term, and softmax indicates normalization function; S34. Aggregate the operation dependencies under all attention heads into an operation feature vector. The Transformer model merges the outputs of multiple heads through a weighted multi-head attention mechanism to generate the final operation feature vector: Among them, Y i represents the final aggregated feature vector of the ith operation, G represents the number of attention heads, and V g represents the learnable weight of each attention head, Represents the value vector under the g-th attention head; S35. According to Y i and Δt i , the system generates an operational abnormal behavior score: Among them, S op represents the abnormal behavior score of the entire operation sequence, B ij represents the attention weight between operations, κ i represents the adaptive adjustment factor of the ith operation, ζ represents the time decay coefficient, t i and t j Indicates the timestamp of the operation; S36, when abnormal behavior score S op Exceeding the set threshold T op When the score is lower than the threshold, the system determines that the operation is abnormal, triggers the protection mechanism and records the abnormal operation log; if the score is lower than the threshold, the operation is considered normal and the system continues to allow the data export operation.

5. According to a method for protecting administrator face recognition data export based on deep learning according to claim 1, it is characterized in that: The S4 specifically includes: S41, the system extracts normal operation and attack behavior features based on the administrator's operation log and historical attack behavior data, and the normal behavior feature set is N = {n1, n2, ..., n p }, where n i represents the i-th normal behavior feature, and the attack behavior feature set is A = {a1, a2, …, a q }, where a j represents the characteristics of the j-th attack behavior; S42, the system inputs the attack behavior features into a generative adversarial network, which is composed of a generator G and a discriminator D. The generator generates a potential attack behavior sample G(z) according to random noise z0: G(z)=G(θ g ,z0); Among them, θ g represents the learnable parameters of the generator, and z0 represents the random noise vector input to the generator; S43, the imitation learning part uses the historical attack behavior dataset P mimic The characteristic operation mode M in r } to train the generator’s output G(z) to be close to the real attack pattern: Among them, L G represents the loss function of the generator, E represents the mathematical expectation, P z represents the random noise distribution, D(G(z)) represents the output probability of the discriminator for the generated sample, λ g represents the adjustment coefficient of imitation learning, ||G(z)-m|| 2 represents the difference between the generated sample and the imitated attack behavior, and m represents the historical attack behavior pattern; S44, discriminator D is used to distinguish between real attack behaviors and fake attack behaviors, and the optimization goal is to maximize the recognition probability of real attack samples: Among them, L D represents the loss function of the discriminator, P data represents the distribution of real attack behaviors, and x represents real attack samples; S45. Generate and optimize potential attack behavior samples through a generative adversarial network. The system continuously trains an anomaly detection module. The anomaly detection module improves the detection capability of complex and hidden attack behaviors by learning the generated attack samples. When a potential attack behavior is detected, a protection mechanism is triggered to record the attack operation and update the Transformer model. S46. The system continuously optimizes the generation capability of the generator based on the continuously updated operation log by combining the generative adversarial network with the imitation learning technology.

6. According to a method for protecting administrator face recognition data export based on deep learning according to claim 1, it is characterized in that: The S51 specifically includes a false data feature set F={f1, f2, ..., f m }, where f i Represents the characteristics of the i-th type of false data: f i =g(d i ,i f ); Among them, d i represents the real data feature of the i-th category, θ f represents the transformation parameters for generating false data, g(·) represents the data transformation function; The S52 specifically includes a virtual operation interface set V={v1, v2, ..., v k }, where v j Represents the jth virtual operation interface: v j =h(u j ,i v ); Among them, u j represents the operation behavior of the jth attacker, θ v represents the transformation parameter for generating the fake interface, h(·) represents the mapping function between the attacker's operation and the fake interface; The S53 specifically includes the induction operation set I={i1,i2,…,i e }, where i l Represents the lth induced operation, and calculates the ineffectiveness score of the induced operation: Among them, S i represents the ineffectiveness score of the induced operation, e represents the total number of induced operations, α l represents the weight of the lth induction operation, It represents the result of the attacker's first fake operation in the fake environment. represents the result of the lth real operation, μ l Represents the adaptive adjustment coefficient, Δt l represents the time interval between the lth operation and the last operation, γ l represents the time attenuation coefficient; The S54 specifically includes dynamically optimizing the false environment: Among them, C f represents the complexity of the false environment, m represents the total number of false data types, ρ i represents the weight of the i-th type of false data, δ i represents the adaptive complexity adjustment coefficient, and T represents the scoring threshold set by the system; S55 specifically includes calculating the matching score M r : Among them, θ q Represents the weight of the qth operation.

7. According to a method for protecting administrator face recognition data export based on deep learning according to claim 1, it is characterized in that: The S6 specifically includes: S61, the system monitors the attacker's operation behavior in the virtual environment and collects behavior data in real time. The system initializes the induction strategy according to the operation sequence and generates the initial strategy using the proximal strategy optimization algorithm; S62, the system dynamically optimizes the initial strategy through a meta-policy gradient algorithm, and the meta-policy gradient algorithm adjusts the induced strategy in real time according to the attacker's operation mode: Among them, L Meta (θ) represents the loss function of the meta-policy gradient algorithm, θ k represents the parameters of the k-th strategy, represents the kth strategy, Represents the parameter θ k The gradient of, E represents the mathematical expectation, R t represents the reward function at time step t, and T represents the total length of the time step; S63. The system continuously optimizes the induction strategy based on the feedback of the attacker's operation behavior: Among them, F t represents the feedback value at time step t, r i represents the reward value of the attacker's ith operation in the fake environment, λ3 represents the time interval sensitivity coefficient, and γ f represents the discount factor, t i -t i-1 Indicates the time difference between two operations; S64. Based on the feedback function and the result of the meta-strategy gradient optimization, the system dynamically adjusts the induction strategy: Among them, θ′ represents the updated policy parameter, θ represents the policy parameter before the update, and τ represents the learning rate; S65. The system dynamically adjusts the inducement operation in the fake environment according to the optimized policy parameters, and generates a new inducement interface and data according to the attacker's operation mode.

Images