A zero-trust dynamic access control method based on GBDS user credibility evaluation

Abstract

The present invention discloses a zero-trust dynamic access control method based on GBDS user credibility assessment, adopts the authority-task-role model, assigns authority to tasks, binds tasks to roles, and then assigns roles to users. When the task is completed, the corresponding authority will be withdrawn, realizing the principle of least privilege and separation of duties. At the same time, the trust value of the user is calculated continuously in units of time. At the same time, a user credibility assessment method based on the gradient boosting decision flow GBDS is proposed, which will conduct a rapid risk assessment of abnormal behavior for users whose trust value is lower than the minimum threshold. The main concept of the method is to merge similar nodes after each split iteration, which can produce a deeper level. The present invention can continuously monitor the user's access behavior, change the entity access rights in real time, adjust the policy in real time, and perform dynamic access authorization. By applying this dynamic access control method, it can ensure that our system is more secure.

Description

Technical Field

[0001] The present invention belongs to the technical field of access control, and in particular relates to a zero-trust dynamic access control method based on GBDS user credibility evaluation. Background Art

[0002] With the continuous deepening of information technology, in order to improve the informatization and business service capabilities of the power grid, emerging technologies such as cloud computing, big data, and artificial intelligence are introduced. In the application process, the key data of some power grids are exposed, which brings new challenges to the information security of the power grid. However, on the one hand, the security of such protection measures is seriously insufficient and there is a risk of being breached. On the other hand, the existing static access control methods and defense strategies are static, solidified, inflexible, and have extensive authority control. A major result of these trends is the shift from the traditional perimeter-based security model to the application of the so-called zero trust network (ZTN) security model, which treats the enterprise intranet to the same extent, that is, lack of trust, as the Internet. The key element of the ZTN approach is to treat the internal network as an untrusted network to the same extent as the Internet. The internal network is divided into multiple segments or zones, each of which contains different functions and information. Each zone will have a different trust level, indicating the importance of the assets within the zone. In order to access the assets, the trust level assigned to the subject must be equal to or greater than the minimum trust level of the zone. According to the overall access control strategy, the communication between the partitions is restricted by the firewall. Access control is also dynamic and transaction-based, making decisions for each access request and updating the rules on the inter-zone firewall as needed for each transaction. Summary of the invention

[0003] The purpose of the present invention is to provide a zero-trust dynamic access control method based on GBDS user credibility assessment, which fully considers the multi-dimensional attributes of the subject, object and environment for security assessment, measures potential security risks and trust levels based on multi-source perceived data, standardizes security quantification standards, and forms dynamic access control strategies. In the access control of network power services based on this model, the access agent continuously monitors the user's access behavior, changes the entity access rights in real time, can adjust the strategy in real time according to the user's behavior, and perform dynamic adaptive access authorization. If sensitive operations are found, the user's access rights can be terminated in time to avoid further losses. Establish a zero-trust security architecture suitable for power services in the multi-cloud and big data era.

[0004] The present invention evaluates the security of communication based on the risk of the user and the importance of the resources in the target terminal. When a user connects to the organizational network from the outside, the system dynamically changes the access control according to the evaluation results. For example, the communication of high-risk users and the communication with the server storing important resources are restricted. At the same time, considering that abnormal behavior does not necessarily bring security risks, in order to prevent system misjudgment and over-protection, a user credibility evaluation method based on Gradient Boosting Decision Stream (GBDS) is proposed. By applying this dynamic access control method, we can ensure that our system is more secure.

[0005] The zero-trust dynamic access control method based on GBDS user credibility evaluation includes:

[0006] Step 1: The user sends an access request. If the user is a new user, he / she submits personal identity information for registration. The administrator reviews the user's identity. If the registration is successful, the user is assigned an initial trust value. If the registration fails, access is denied.

[0007] Step 2: The user logs in to the system. If the password is correct, the identity authentication is passed; if the password is incorrect, access is denied;

[0008] Step 3: The system obtains a data set of user behavior, including event logs, malicious activities, environment, and information security-related learning status triggered by the user, and calculates the user's risk value based on this data;

[0009] Step 4: The event log includes the content of the event and its severity. Based on the severity, the system calculates the event risk caused by the user and obtains the risk value for each accident level;

[0010] Step 5: The system classifies data resources according to their importance;

[0011] Step 6: The total trust value is obtained by weighted comprehensive calculation of the user's risk value and the importance of access resources. The total trust value is compared with the minimum acceptable trust (threshold trust value). If the total trust value is higher than the threshold trust value, access rights will be granted; if the total trust value is lower than the threshold trust value, considering that abnormal behavior does not necessarily bring security risks, in order to prevent system misjudgment and over-protection, a user credibility assessment based on the gradient boosting decision flow GBDS is proposed to conduct a rapid risk assessment of abnormal behavior for such users;

[0012] Step 7: After successfully accessing the system, permissions are distributed to tasks, tasks are assigned to roles, and roles are assigned to the principal user in a specific session. Users run various tasks through the assigned roles and are assigned permissions based on the tasks of the role. Permissions are revoked when the task is completed, so the role has permissions only when the task is being executed.

[0013] Step 8: After receiving the complete resource access request, the policy decision point performs policy matching in the policy management point and finds the most suitable policy;

[0014] Step 9: The policy decision point evaluates the most suitable policy based on the task information in the complete resource access request and returns the access control decision to the policy enforcement point;

[0015] Step 10: The policy enforcement point implements the returned access control decision. If the access control decision is allowed, the user terminal can access the requested resource, otherwise it cannot access it.

[0016] Furthermore, the total trust value is obtained by weighted comprehensive calculation of the user's risk value and the importance of access resources. The steps are:

[0017] Step 6.11: Considering the different importance of components or resources in the system, the degree of harm caused by threats is also different. Therefore, the importance of resources is divided into levels and assigned values. That is, the higher the level, the more important the resource is and the greater the weight it occupies.

[0018] Resource importance RW = {RW1, RW2, RW3, RW4}, where RW1 represents common data, RW2 represents general data, RW3 represents important data, and RW4 represents very important data;

[0019] Step 6.12: Consider the different degrees of impact of user threat behaviors on resources, and classify and assign values ​​to threat behaviors. The higher the threat behavior level, the more serious the user threat behavior.

[0020] Threatening behavior impact TB = {TB1, TB2, TB3, TB4}, where TB1 indicates a slight impact, TB2 indicates a moderate impact, TB3 indicates a severe impact, and TB4 indicates a very severe impact;

[0021] Step 6.13: Consider that the user's threatening behavior will increase its risk value, but the first threatening behavior may be a misoperation or a tentative attack. Therefore, when the threatening behavior occurs again, the user's behavior risk value will increase sharply; and when the user's behavior returns to normal and there is no dangerous operation, the risk value will gradually decrease;

[0022] Step 6.14: Define RU as the total trust value. To facilitate the evaluation and calculation of the trust value, map the value of RU to the interval [0,1],

[0023]

[0024] Among them, t is the number of threatening behaviors of the user, RU0 is the trust value of the user calculated most recently, and different threatening behavior frequencies will affect the risk value, that is, the higher the threatening behavior frequency, the greater the risk value.

[0025] Furthermore, user credibility assessment based on GBDS (Gradient Boosting Decision Stream) means that considering that abnormal behavior does not necessarily bring security risks, in order to prevent system misjudgment and over-protection, users with trust values ​​below the minimum acceptable trust will be subject to rapid risk assessment of abnormal behavior. Compared with the classic gradient boosting decision tree, the main concept of this method is to merge similar nodes after each split iteration. This process eliminates the problem of the gradual decrease in the amount of leaf node data in the decision tree, and can produce a deeper level. GB requires a total of M iterations, and each iteration produces a model. The gradient descent method is used to make the loss function smaller and smaller by moving in the negative gradient direction of the loss function at each iteration. The steps are as follows:

[0026] Step 6.21: First initialize the base learner:

[0027]

[0028] Among them, L(y i ,c) represents the loss function, N represents the number of decision trees, for multi-class classification. When the number of categories is i:

[0029]

[0030] Step 6.22: For the mth iteration, when m≤M, execute (1)-(4) repeatedly;

[0031] (1) Calculate the residual r mi ;

[0032]

[0033] (2) For r mi Fit a regression tree and get the leaf node region R of the mth tree mj ;

[0034] (3) For j = 1, 2, ..., J, linear search is performed to find the minimum value c of the loss function mj ;

[0035]

[0036] (4) Update f(x);

[0037]

[0038] Step 6.23: Get the regression tree;

[0039]

[0040] Step 6.24: For the problem that there are three types of samples: credible, generally credible, and uncredible, use a three-dimensional vector to mark the label of the sample. [1,0,0] indicates that the sample is credible, [0,1,0] indicates that the sample is generally credible, and [0,0,1] indicates that the sample is uncredible;

[0041] Step 6.25: Train a decision stream for each class independently;

[0042] Step 6.26: After the classic decision tree branches, the merging algorithm takes the leaf nodes generated in the current stage and the unsplit nodes previously obtained from the upper layer of the model as input and fuses similar nodes.

[0043] The calculation expression of correlation strength is:

[0044]

[0045] Step 6.27: Finally, to estimate the prediction accuracy of the K leaf nodes of class J, use Gini impurity to calculate:

[0046]

[0047] Furthermore, the gradient boosting algorithm refers to an approximate method using the steepest descent, that is, using the value of the negative gradient of the loss function in the current model as an approximate value of the residual of the boosting tree algorithm in the classification problem for fitting. The residual refers to the difference between the true value and the predicted value. In the learning process, first learn a regression tree, then use "true value-predicted value" to get the residual, and then use the residual as a learning target to learn the next regression tree, and so on, until the residual is less than a threshold close to 0 or the number of regression trees reaches a certain threshold. The core idea is to reduce the loss function by fitting the residual in each round.

[0048] Compared with the prior art, the method of the present invention has the following advantages:

[0049] 1) First, the present invention adopts a permission-task-role model, in which permissions are assigned to tasks, tasks are bound to roles, and roles are then assigned to users. When the task is completed, the corresponding permissions will be revoked, realizing the principle of least privilege and separation of duties. At the same time, the user's trust value is continuously calculated in units of time, and the trust value will change with the change of user behavior. Compared with traditional static coarse-grained access control technology, it can continuously monitor the user's access behavior, change entity access rights in real time, adjust policies in real time, and perform dynamic access authorization.

[0050] 2) The business system of the power grid is complex, involving a wide variety of business types and various types of terminals for access. The requirements of various terminals and businesses for access control vary greatly. After the introduction of the zero-trust security definition boundary security architecture, how to universally represent these differentiated entity identity attributes is one of the difficulties. How to evaluate the impact of changes in the security status of each entity and the security status of the environment on access control is another difficulty. Finally, how to change the access control results of the subject to the object based on entity behavior or environmental changes in the multi-cloud and big data era environment is also a huge challenge for establishing a zero-trust security architecture suitable for power business. The zero-trust dynamic access control method based on GBDS user credibility evaluation of the present invention can effectively solve the above problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of the zero-trust dynamic access control provided by the present invention.

[0052] Figure 2 This is an architectural diagram of the zero-trust dynamic access control provided by the present invention. DETAILED DESCRIPTION

[0053] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. However, the following description is only exemplary and is not intended to limit the scope of the present invention and its application.

[0054] like Figure 1 and Figure 2 A zero-trust dynamic access control method based on GBDS user credibility evaluation is shown, and the specific steps are as follows:

[0055] Step 1: The user sends an access request. If the user is a new user, he / she submits personal identity information for registration. The administrator reviews the user's identity. If the registration is successful, the user is assigned an initial trust value. If the registration fails, access is denied.

[0056] Step 2: The user logs in to the system. If the password is correct, the identity authentication is passed; if the password is incorrect, access is denied;

[0057] Step 3: The system obtains a data set of user behavior, including event logs, malicious activities, environment, and information security-related learning status triggered by the user, and calculates the user's risk value based on this data;

[0058] Step 4: The event log includes the content of the event and its severity. Based on the severity, the system calculates the event risk caused by the user and obtains the risk value for each accident level;

[0059] Step 5: The system classifies data resources according to their importance;

[0060] Step 6: The total trust value is obtained by weighted comprehensive calculation of the user's risk value and the importance of access resources. The total trust value is compared with the minimum acceptable trust (threshold trust value). If the total trust value is higher than the threshold trust value, access rights will be granted; if the total trust value is lower than the threshold trust value, considering that abnormal behavior does not necessarily bring security risks, in order to prevent system misjudgment and over-protection, a user credibility assessment based on the gradient boosting decision flow GBDS is proposed to conduct a rapid risk assessment of abnormal behavior for such users;

[0061] Step 7: After successfully accessing the system, permissions are distributed to tasks, tasks are assigned to roles, and roles are assigned to the principal user in a specific session. Users run various tasks through the assigned roles and are assigned permissions based on the tasks of the role. Permissions are revoked when the task is completed, so the role has permissions only when the task is being executed.

[0062] Step 8: After receiving the complete resource access request, the policy decision point performs policy matching in the policy management point and finds the most suitable policy;

[0063] Step 9: The policy decision point evaluates the most suitable policy based on the task information in the complete resource access request and returns the access control decision to the policy enforcement point;

[0064] Step 10: The policy enforcement point implements the returned access control decision. If the access control decision is allowed, the user terminal can access the requested resource, otherwise it cannot access it.

[0065] The total trust value is obtained by weighted comprehensive calculation of the user's risk value and the importance of access resources. The steps are:

[0066] Step 6.11: Considering the different importance of components or resources in the system, the degree of harm caused by threats is also different. Therefore, the importance of resources is divided into levels and assigned values. That is, the higher the level, the more important the resource is and the greater the weight it occupies.

[0067] Resource importance RW = {RW1, RW2, RW3, RW4}, where RW1 represents common data, RW2 represents general data, RW3 represents important data, and RW4 represents very important data;

[0068] Step 6.12: Consider the different degrees of impact of user threat behaviors on resources, and classify and assign values ​​to threat behaviors. The higher the threat behavior level, the more serious the user threat behavior.

[0069] Threatening behavior impact TB = {TB1, TB2, TB3, TB4}, where TB1 indicates a slight impact, TB2 indicates a moderate impact, TB3 indicates a severe impact, and TB4 indicates a very severe impact;

[0070] Step 6.13: Consider that the user's threatening behavior will increase its risk value, but the first threatening behavior may be a misoperation or a tentative attack. Therefore, when the threatening behavior occurs again, the user's behavior risk value will increase sharply; and when the user's behavior returns to normal and there is no dangerous operation, the risk value will gradually decrease;

[0071] Step 6.14: Define RU as the total trust value. To facilitate the evaluation and calculation of the trust value, map the value of RU to the interval [0,1],

[0072]

[0073] Among them, t is the number of threatening behaviors of the user, RU0 is the trust value of the user calculated most recently, and different threatening behavior frequencies will affect the risk value, that is, the higher the threatening behavior frequency, the greater the risk value.

[0074] User credibility assessment based on GBDS (Gradient Boosting Decision Stream) means that considering that abnormal behavior does not necessarily bring security risks, in order to prevent system misjudgment and over-protection, users with trust values ​​below the minimum acceptable trust will be subject to rapid risk assessment of abnormal behavior. Compared with the classic gradient boosting decision tree, the main concept of this method is to merge similar nodes after each split iteration. This process eliminates the problem of the gradual decrease in the amount of leaf node data in the decision tree and can produce a deeper level. GB requires a total of M iterations, and each iteration produces a model. The gradient descent method is used to make the loss function smaller and smaller by moving in the negative gradient direction of the loss function at each iteration. The steps are as follows:

[0075] Step 6.21: First initialize the base learner;

[0076]

[0077] Among them, L(y i ,c) represents the loss function, N represents the number of decision trees, for multi-class classification. When the number of categories is i:

[0078]

[0079] Step 6.22: For the mth iteration, when m≤M, execute (1)-(4) repeatedly;

[0080] (1) Calculate the residual r mi ;

[0081]

[0082] (2) For r miFit a regression tree and get the leaf node region R of the mth tree mj ;

[0083] (3) For j = 1, 2, ..., J, linear search is performed to find the minimum value c of the loss function mj ;

[0084]

[0085] (4) Update f(x);

[0086]

[0087] Step 6.23: Get the regression tree;

[0088]

[0089] Step 6.24: For the problem that there are three types of samples: credible, generally credible, and uncredible, use a three-dimensional vector to mark the label of the sample. [1,0,0] indicates that the sample is credible, [0,1,0] indicates that the sample is generally credible, and [0,0,1] indicates that the sample is uncredible;

[0090] Step 6.25: Train a decision stream for each class independently;

[0091] Step 6.26: After the classic decision tree branches, the merging algorithm takes the leaf nodes generated in the current stage and the unsplit nodes previously obtained from the upper layer of the model as input and fuses similar nodes.

[0092] The calculation expression of correlation strength is:

[0093]

[0094] Step 6.27: Finally, to estimate the prediction accuracy of the K leaf nodes of class J, use Gini impurity to calculate:

[0095]

[0096] The gradient boosting algorithm refers to an approximate method that uses the steepest descent, that is, using the negative gradient of the loss function in the current model as an approximate value of the residual of the boosting tree algorithm in the classification problem for fitting. The residual refers to the difference between the true value and the predicted value. In the learning process, first learn a regression tree, then use "true value-predicted value" to get the residual, and then use the residual as a learning target to learn the next regression tree, and so on, until the residual is less than a threshold close to 0 or the number of regression trees reaches a certain threshold. The core idea is to reduce the loss function by fitting the residual in each round.

[0097] In order to facilitate understanding of the above technical solutions of the present invention, the above technical solutions of the present invention are described in detail below through specific embodiments.

[0098] Example

[0099] Comparative test,Comparative test refers to the use of two or more related algorithms, through comparative analysis of the results, to explore the impact of different algorithms on the experimental results. In order to demonstrate the effectiveness of the various parts designed in the proposed method of fine-grained dynamic access control for visiting users, the classic classification algorithms Support Vector Machine (SVM) and Gradient Boosting Decision Tree (GBDT) are used to compare with the improved GBDS algorithm. In order to remove randomness, all experiments are repeated 5 times, and the average value of each indicator is taken. Then the indicators used to solve our research problem are given. First, the following variables are defined:

[0100] 1) True Positive TP (True Positive) indicates the number of policies that allow authorization to be identified as allowed, denoted as TP.

[0101] 2) False Positive (FP) indicates the number of policies that deny authorization and are identified as allowed policies, denoted as FP.

[0102] 3) False Negative FN (False Negative) indicates the number of policies that allow authorization but are identified as denying policies, denoted as FN.

[0103] 4) True Negative TN (true negative) indicates the number of policies that deny authorization and are identified as denied policies, denoted as TN.

[0104] We use accuracy to measure the performance of each category. The calculation expression of accuracy is:

[0105]

[0106] It represents the ratio of the number of all correctly judged samples to the total number. The higher the accuracy, the better the algorithm effect.

[0107] 2) The recall rate calculation expression is:

[0108]

[0109] It indicates how many positive examples in the original sample are predicted correctly. The higher the recall rate, the better the algorithm effect.

[0110] 3) The F1-score calculation expression is:

[0111]

[0112] F1-score is an indicator used in statistics to measure the accuracy of a binary classification model. It takes into account both the accuracy and recall of the classification model. It is a weighted average of the accuracy and recall of the model. The maximum value is 1 and the minimum value is 0. The larger the value, the better the model.

[0113] We compared the accuracy, recall, F1-score and running time respectively. The experimental results are shown in Table 1. The variance of these means is between 0.01 and 0.1.

[0114]

[0115]

[0116] In contrast, GBDS not only has stable and high-quality training results, but also begins to converge after about 200 training rounds, achieving the best training results.

[0117] The above only expresses the preferred embodiments of the present invention, and the description is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for ordinary technicians in this field, several modifications, improvements and substitutions can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. A zero-trust dynamic access control method based on gradient boosting decision flow GBDS user credibility evaluation, characterized in that: The following steps are involved: Step 1: The user sends an access request. If the user is a new user, he / she submits personal identity information for registration. The administrator reviews the user's identity. If the registration is successful, the user is assigned an initial trust value. Failure to register will result in access being denied; Step 2: The user logs in to the system. If the password is correct, the identity authentication is passed; if the password is incorrect, access is denied; Step 3: The system obtains a data set of user behavior, including event logs, malicious activities, environment, and information security-related learning status triggered by the user, and calculates the user's risk value based on this data; Step 4: The event log includes the content of the event and its severity. Based on the severity, the system calculates the event risk caused by the user and obtains the risk value of each accident level; Step 5: The system classifies data resources according to their importance; Step 6: The total trust value is calculated by weighted comprehensive calculation of the user's risk value and the importance of the accessed resource, and the total trust value is compared with the minimum acceptable trust value, i.e., the threshold trust value. If the total trust value is higher than the threshold trust value, access permission is granted. If the total trust value is lower than the threshold trust value, a user credibility assessment based on the gradient boosting decision flow GBDS is proposed to conduct a rapid risk assessment of abnormal behavior for this type of user; Step 7: After successfully accessing the system, permissions are distributed to tasks, tasks are assigned to roles, and roles are assigned to principal users in a specific session; users run various tasks through the assigned roles and are assigned permissions based on the tasks of the roles; Permissions are revoked when the task is completed, so the role only has permissions while the task is being executed; Step 8: After receiving the complete resource access request, the policy decision point performs policy matching in the policy management point and finds the most suitable policy; Step 9: The policy decision point evaluates the most suitable policy based on the task information in the complete resource access request and returns the access control decision to the policy enforcement point; Step 10: The policy enforcement point implements the returned access control decision. If the access control decision is allowed, the user terminal can access the requested resource, otherwise it cannot access it. Based on the user credibility evaluation of the gradient boosting decision flow GBDS, similar nodes are merged after each split iteration; the gradient boosting GB algorithm requires a total of M iterations, each iteration generates a model, and the gradient descent method is used to make the loss function smaller and smaller by moving in the negative gradient direction of the loss function at each iteration; the steps are as follows: Step 6.21: First initialize the base learner; Among them, L(y i ,c) represents the loss function, N represents the number of decision trees, for multi-class classification; when the number of categories is i: Step 6.22: For the mth iteration, when m≤M, execute (1)-(4) repeatedly; (1) Calculate the residual; (2) For r mi Fit a regression tree and get the leaf node region R of the mth tree mj ; (3) For j = 1, 2, ..., J, linear search is performed to find the minimum value of the loss function; (4) Update f(x); Step 6.23: Get the regression tree; Step 6.24: For the problem that there are three types of samples: credible, generally credible, and uncredible, use a three-dimensional vector to mark the label of the sample; [1,0,0] indicates that the sample is credible, [0,1,0] indicates that the sample is generally credible, and [0,0,1] indicates that the sample is uncredible; Step 6.25: Train a decision stream for each class independently; Step 6.26: After the classic decision tree branches, the merging algorithm takes the leaf nodes generated in the current stage and the undivided nodes previously obtained from the upper layer of the model as input and fuses similar nodes; the correlation strength calculation expression is: Step 6.27: Finally, to estimate the prediction accuracy of the K leaf nodes of class J, use Giniimpurity to calculate:

2. According to claim 1, the zero-trust dynamic access control method based on GBDS user credibility evaluation is characterized in that: The total trust value is obtained by weighted comprehensive calculation of the user's risk value and the importance of access resources. The steps are: Step 6.11: Classify and assign values ​​to resource importance. The higher the level, the more important the resource is and the greater its weight is. Resource importance RW = {RW1, RW2, RW3, RW4}, where RW1 represents common data, RW2 represents general data, RW3 represents important data, and RW4 represents very important data; Step 6.12: Classify and assign values ​​to threat behaviors. The higher the threat behavior level, the more serious the user's threat behavior. Threatening behavior impact TB = {TB1, TB2, TB3, TB4}, where TB1 indicates a slight impact, TB2 indicates a moderate impact, TB3 indicates a severe impact, and TB4 indicates a very severe impact; Step 6.13: Consider that the user's threatening behavior will increase its risk value, but the first threatening behavior may be a misoperation or a tentative attack. Therefore, when the threatening behavior occurs again, the user's behavior risk value will increase sharply; and when the user's behavior returns to normal and there is no dangerous operation, the risk value will gradually decrease; Step 6.14: Define RU as the total trust value and map the value of RU to the interval [0,1], Among them, t is the number of threatening behaviors of the user, RU0 is the trust value of the user calculated most recently, and different threatening behavior frequencies will affect the risk value, that is, the higher the threatening behavior frequency, the greater the risk value.

3. The zero-trust dynamic access control method based on GBDS user credibility evaluation according to claim 1 is characterized in that: The gradient boosting (GB) algorithm refers to the use of the steepest descent approximation method, that is, the use of the negative gradient of the loss function in the current model as the approximate value of the residual of the boosting tree algorithm in the classification problem for fitting; the residual refers to the difference between the true value and the predicted value. In the learning process, first learn a regression tree, then use "true value - predicted value" to get the residual, and then use the residual as a learning target to learn the next regression tree, and so on, until the residual is less than a threshold close to 0 or the number of regression trees reaches a certain threshold; the core idea is to reduce the loss function by fitting the residual in each round.

Images