Security protection effectiveness evaluation method based on reward model and chatGLM model

Abstract

The invention discloses a security protection effectiveness evaluation method based on a reward model and a chatGLM model, and the method comprises the steps: firstly capturing general evaluation rules, security protection features, security failure points and failure reasons as training data, forming a mapping pair of evaluation data and rule scoring through rough manual intervention, training the reward model, obtaining a scoring model, and carrying out the scoring of the scoring model; meanwhile, the mapping pairs are converted into question and answer form data to conduct fine adjustment on the chatGLM large model, after training is completed, the deep learning technology is adopted in the inference stage to extract input evaluation rules and semantic interaction features of protection evaluation results, then the large model subjected to fine adjustment is input for content expression optimization, and the evaluation rules and the protection evaluation results are obtained. And inputting the optimized evaluation rule and the protection evaluation result into a trained scoring model for scoring to obtain a safety protection effectiveness evaluation value, thereby realizing automatic and intelligent evaluation of a safety protection system, improving evaluation efficiency and accuracy, and meeting the demand of large-scale data processing.

Description

Technical Field

[0001] The present invention relates to the field of network security technology, and in particular to a security protection effectiveness evaluation method based on a reward model and a chatGLM model. Background Art

[0002] In the current field of network security, with the continuous advancement of technology and the increasing sophistication of network attack methods, the effectiveness evaluation of security protection systems has become increasingly important. Traditional security protection measures often focus on the deployment of protective equipment and the formulation of strategies, but whether these measures are truly effective and can resist real-world attacks requires scientific evaluation and verification.

[0003] Specifically, in the traditional security protection system, enterprises usually deploy security equipment such as firewalls, intrusion detection systems, antivirus software, and formulate corresponding security policies and operating procedures. However, the effectiveness of these measures often depends on the configuration of the equipment, the scientific nature of the policy, and the operation level of the personnel. In actual operation, due to configuration errors, improper policies or operational errors, the security protection system may have loopholes and cannot effectively defend against external attacks or internal threats.

[0004] In order to improve the effectiveness of security protection, it is necessary to conduct a comprehensive assessment of existing security protection measures. This includes systematic inspection and testing of the configuration of security equipment, the implementation of security policies, and the execution of security management systems. The actual effect of the security protection system can be verified through simulated attacks, penetration tests, etc., to identify potential security risks and propose improvement measures. However, traditional security protection effectiveness assessment mainly relies on expert experience and subjective judgment, and requires a lot of manual participation, which is not only inefficient, but also difficult to adapt to large-scale data processing needs. Summary of the invention

[0005] In view of the deficiencies of the prior art, the present invention aims to provide a security protection effectiveness evaluation method based on a reward model and a chatGLM model.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A security protection effectiveness evaluation method based on a reward model and a chatGLM model, including a training phase and an inference phase;

[0008] The specific process of the training phase is as follows:

[0009] S11, obtaining training data; the training data includes general evaluation rules, safety protection features, safety failure points and failure causes;

[0010] S12, pre-processing the training data obtained in step S11 by data classification and data cleaning, and then taking the safety protection features, safety failure points and failure causes as evaluation data, scoring and calibrating each evaluation data through rough manual intervention according to the obtained general evaluation rules, and grouping each evaluation data with its corresponding rule scoring to form an {evaluation data; rule scoring} mapping pair;

[0011] S13, training the reward model using the preprocessed training data obtained in step S12;

[0012] S14, converting the preprocessed training data obtained in step S12 into question-answering task data, and fine-tuning the chatGLM model based on the question-answering task data to obtain a fine-tuned chatGLM model. In the process of fine-tuning the chatGLM model, continue to iteratively fine-tune the reward model to obtain a scoring model;

[0013] The specific process of the inference stage is as follows:

[0014] S21. Obtain the assessment rules and protection assessment results of the safety protection system;

[0015] S22, semantically encode the evaluation rule and the protection evaluation result respectively to obtain an evaluation rule semantic encoding feature vector and a protection evaluation result semantic encoding feature vector;

[0016] S23, performing multi-scale semantic interaction on the semantic encoding feature vector of the assessment rule and the semantic encoding feature vector of the protection assessment result to obtain an assessment rule-protection assessment result semantic interaction joint matrix;

[0017] S24. Use the fine-tuned chatGLM model and the scoring model to process the assessment rule-protection assessment result semantic interaction joint matrix to generate a security protection effectiveness assessment value.

[0018] Furthermore, in step S14, in the process of fine-tuning the chatGLM model, the specific process of continuing to iteratively fine-tune the reward model to obtain the scoring model is: inputting the optimized evaluation rule expression and the optimized protection evaluation result expression generated in the chatGLM model fine-tuning process into the reward model for scoring, and updating the parameters of the reward model according to the scoring results of the reward model to better reflect the mapping relationship between the evaluation data and the rule scoring, repeating the process of chatGLM model fine-tuning and reward model updating until the performance of the chatGLM model and the reward model reaches a satisfactory level, and finally obtaining the fine-tuned chatGLM model and scoring model.

[0019] Furthermore, in step S22, the evaluation rules and the protection evaluation results are semantically encoded, and the words or phrases in the text are converted into dense vector representations through word embedding technology and self-attention mechanism, and the long-distance semantic dependencies between words are captured, so as to extract the key contextual semantic information in the evaluation rules and the protection evaluation results, and obtain the semantic encoding feature vector of the evaluation rules and the semantic encoding feature vector of the protection evaluation results.

[0020] Furthermore, the specific process of step S23 is:

[0021] S231, feature decoupling is performed on the semantic encoding feature vector of the evaluation rule to obtain a sequence of local semantic feature vectors of the evaluation rule, as specifically shown in the following formula:

[0022] Decouple(f2)={f 21 ,f 22 ,...,f 2i ,...,f 2l}

[0023] f2 is the semantic encoding feature vector of the evaluation rule, Decouple(·) represents the decoupling process, and f 21 ,f 22 ,...,f 2i ,...,f 2l respectively represent the first, second, i-th and l-th evaluation rule local semantic feature vectors in the sequence of the evaluation rule local semantic feature vectors, and l represents the total number of vectors in the sequence of the evaluation rule local semantic feature vectors;

[0024] S232, performing a nonlinear transformation on the semantic coding feature vector of the protection assessment result to obtain a semantic coding feature vector of the protection assessment result after nonlinear transformation, which is specifically shown in the following formula:

[0025] f 1t =W′f1+b

[0026] f1 is the semantic encoding feature vector of the protection assessment result, W′ is the weight matrix, b is the bias vector, and f 1t Semantically encode feature vectors for protection evaluation results after nonlinear transformation;

[0027] S233, respectively calculating the association matrix between the semantic coding feature vector of the protection evaluation result after nonlinear transformation and each evaluation rule local semantic feature vector in the sequence of the evaluation rule local semantic feature vector, to obtain a sequence of evaluation rule-protection evaluation result semantic interaction association matrices;

[0028] S234, calculating the characteristic interaction energy factor of each evaluation rule-protection evaluation result semantic interaction association matrix in the sequence of the evaluation rule-protection evaluation result semantic interaction association matrix to obtain a sequence of characteristic interaction energy factors;

[0029] S235. Based on the sequence of the feature interaction energy factors, the sequence of the evaluation rule-protection evaluation result semantic interaction association matrices is fused to obtain the evaluation rule-protection evaluation result semantic interaction joint matrix.

[0030] Furthermore, in step S233, the calculation process of the evaluation rule-protection evaluation result semantic interaction association matrix is: multiplying the transposed vector of the protection evaluation result semantic encoding feature vector after nonlinear transformation by the evaluation rule local semantic feature vector to obtain the corresponding evaluation rule-protection evaluation result semantic interaction association matrix, as shown in the following formula:

[0031]

[0032] (·) T represents the transpose of a vector, Represents matrix multiplication operation, M 12i It represents the semantic interaction association matrix of the ith evaluation rule and the protection evaluation result, that is, the association matrix between the semantic encoding feature vector of the protection evaluation result after nonlinear transformation and the local semantic feature vector of the ith evaluation rule.

[0033] Furthermore, in step S234, the calculation process of the feature interaction energy factor is as follows: the square of the difference between the kurtosis and the feature mean of the evaluation rule-protection evaluation result semantic interaction association matrix, the feature variance of the evaluation rule-protection evaluation result semantic interaction association matrix, and the weighted sum of the bias items of the evaluation rule-protection evaluation result semantic interaction association matrix are calculated, and then the feature variance of the evaluation rule-protection evaluation result semantic interaction association matrix and the bias item are added and divided by the weighted sum to obtain the feature interaction energy factor, as shown in the following formula:

[0034]

[0035]

[0036] E 12i represents the characteristic interaction energy factor of the semantic interaction association matrix of the i-th evaluation rule-protection evaluation result, k, μ and σ 2 denote the kurtosis, feature mean and feature variance of the semantic interaction matrix of the i-th evaluation rule-protection evaluation result, respectively, and λ is the bias term E{[M 12i (t)-μ] 4}, represents the fourth-order central moment of the semantic interaction association matrix of the ith evaluation rule-protection evaluation result, that is, the expected value of the fourth power of the difference between each eigenvalue in the ith evaluation rule-protection evaluation result semantic interaction association matrix and its eigenvalue mean, and t represents each eigenvalue in the semantic interaction association matrix of the evaluation rule-protection evaluation result.

[0037] Furthermore, the specific process of step S235 is: inputting the sequence of feature interaction energy factors into the Sigmoid function to obtain a sequence of normalized feature interaction energy factors; using the sequence of normalized feature interaction energy factors as a sequence of weights, calculating the position-weighted sum of the sequence of the evaluation rule-protection evaluation result semantic interaction association matrix, and obtaining the evaluation rule-protection evaluation result semantic interaction joint matrix; as shown in the following formula:

[0038] E norm,12i =sigmoid(E 12i )

[0039] M=∑ i M 12i E norm,12i

[0040] sigmoid represents the sigmoid activation function, E norm,12i represents the ith normalized feature interaction energy factor, and M represents the evaluation rule-protection evaluation result semantic interaction joint matrix.

[0041] Furthermore, the specific process of step S24 is as follows:

[0042] S241, inputting the evaluation rule-protection evaluation result semantic interaction joint matrix into the fine-tuned chatGLM model to obtain an optimized evaluation rule expression and an optimized protection evaluation result expression;

[0043] S242, inputting the optimized evaluation rule expression and the optimized protection evaluation result expression into the scoring model to obtain the safety protection effectiveness evaluation value;

[0044] The specific process of step S241 is: calculate the sum of the absolute values ​​of each eigenvalue of the evaluation rule-protection evaluation result semantic interaction joint matrix to obtain the first evaluation rule-protection evaluation result semantic interaction joint and modulation value, and calculate the square root of the sum of the squares of each eigenvalue of the evaluation rule-protection evaluation result semantic interaction joint matrix to obtain the second evaluation rule-protection evaluation result semantic interaction joint and modulation value; specifically as shown in the following formula:

[0045]

[0046] m i ∈M∈RW×H

[0047] n=W×H

[0048] Where M represents the evaluation rule-protection evaluation result semantic interaction matrix, α is the first evaluation rule-protection evaluation result semantic interaction matrix and modulation value, β is the second evaluation rule-protection evaluation result semantic interaction matrix and modulation value, m i represents the i-th eigenvalue of the evaluation rule-protection evaluation result semantic interaction joint matrix, n is the number of eigenvalues ​​in the evaluation rule-protection evaluation result semantic interaction joint matrix; W and H represent the number of columns and rows of the evaluation rule-protection evaluation result semantic interaction joint matrix, respectively, and R represents a real number set;

[0049] After performing point subtraction on the evaluation rule-protection evaluation result semantic interaction joint matrix and the second evaluation rule-protection evaluation result semantic interaction joint and modulation value, perform point multiplication with the number of eigenvalues ​​of the evaluation rule-protection evaluation result semantic interaction joint matrix and the inverse of the first evaluation rule-protection evaluation result semantic interaction joint and modulation value to calculate matrix one, and take the inverse of each eigenvalue in matrix one to obtain the first evaluation rule-protection evaluation result semantic interaction joint phase conversion matrix; specifically as shown in the following formula:

[0050]

[0051] M1 represents the first evaluation rule-protection evaluation result semantic interaction joint phase transformation matrix, ⊙ represents the dot product, Indicates point reduction, [·] ⊙-1 It means taking the inverse of each eigenvalue in the matrix;

[0052] After performing point subtraction on the evaluation rule-protection evaluation result semantic interaction joint matrix and the first evaluation rule-protection evaluation result semantic interaction joint and modulation value, perform point multiplication with the square root of the number of eigenvalues ​​of the evaluation rule-protection evaluation result semantic interaction joint matrix and the inverse of the second evaluation rule-protection evaluation result semantic interaction joint and modulation value to calculate the matrix 2, and take the inverse of each eigenvalue in the matrix 2 to obtain the second evaluation rule-protection evaluation result semantic interaction joint phase conversion matrix; specifically as shown in the following formula:

[0053]

[0054] M2 represents the second evaluation rule-protection evaluation result semantic interaction joint phase transformation matrix;

[0055] Calculate the dot product vector of the second evaluation rule-protection evaluation result semantic interaction joint phase transformation matrix and the weighted hyperparameter, and then perform dot subtraction between the first evaluation rule-protection evaluation result semantic interaction joint phase transformation matrix and the dot product vector to obtain the optimized evaluation rule-protection evaluation result semantic interaction joint matrix; specifically as shown in the following formula:

[0056]

[0057] M′ is the optimized evaluation rule-protection evaluation result semantic interaction joint moment, and ω is a weighted hyperparameter;

[0058] The optimized evaluation rule-protection evaluation result semantic interaction joint matrix is ​​input into the fine-tuned chatGLM model to obtain the optimized evaluation rule expression and the optimized protection evaluation result expression.

[0059] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program implements the above method when executed by a processor.

[0060] The present invention also provides a computer device, comprising a processor and a memory, wherein the memory is used to store a computer program; and when the processor is used to execute the computer program, the above method is implemented.

[0061] The beneficial effects of the present invention are as follows: the present invention first captures general evaluation rules, security protection features, security failure points and failure causes as training data, and forms a mapping pair of evaluation data and rule scores through rough manual intervention, so as to train the reward model and obtain a scoring model. At the same time, the mapping pair is converted into question-and-answer format data to fine-tune the chatGLM large model. After the training is completed, deep learning technology is used in the inference stage to extract the semantic interaction features of the input evaluation rules and protection evaluation results, and then the fine-tuned large model is input for content expression optimization, and then the optimized evaluation rules and protection evaluation results are input into the trained scoring model for scoring to obtain a security protection effectiveness evaluation value, thereby realizing automated and intelligent evaluation of the security protection system, improving evaluation efficiency and accuracy, and meeting the needs of large-scale data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a flow chart of a method for evaluating the effectiveness of security protection based on a reward model and a chatGLM model according to an embodiment of the present invention;

[0063] Figure 2 A data flow diagram of a security protection effectiveness evaluation method based on a reward model and a chatGLM model according to an embodiment of the present invention;

[0064] Figure 3 is a flowchart of the implementation of step S23 in the method of the embodiment of the present invention;

[0065] Figure 4 is a flowchart of the implementation of step S24 in the method of the embodiment of the present invention;

[0066] Figure 5 is a flowchart for implementing the method of evaluating potential in an embodiment of the present invention;

[0067] Figure 6 This is a training example diagram of an attack load response assessment model based on a large model technology in an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The present invention will be further described below in conjunction with the accompanying drawings. It should be noted that this embodiment is based on the technical solution and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to this embodiment.

[0069] This embodiment provides a security protection effectiveness evaluation method based on a reward model and a chatGLM model.

[0070] In the training phase, we first capture general evaluation rules, safety protection features, safety failure points and failure causes as training data, and form a mapping pair of evaluation data and rule scores through rough manual intervention, so as to train the reward model (RM model) and obtain the scoring model; at the same time, we convert the mapping pair into question-and-answer format data to fine-tune the chatGLM model;

[0071] In the inference stage, deep learning technology is used to extract the semantic interaction features of the input evaluation rules and protection evaluation results, and then the fine-tuned chatGLM model is input to optimize the content expression. The optimized evaluation rules and protection evaluation results are then input into the trained scoring model for scoring to obtain the security protection effectiveness evaluation value. In this way, the automated and intelligent evaluation of the security protection system can be achieved, the evaluation efficiency and accuracy can be improved, and the needs of large-scale data processing can be met.

[0072] Specifically, Figure 1 and Figure 2 As shown, the security protection effectiveness evaluation method based on the reward model and the chatGLM model of this embodiment includes a training phase and an inference phase.

[0073] The specific process of the training phase is as follows:

[0074] S11, obtaining training data. The training data includes general evaluation rules, safety protection features, safety failure points and failure causes;

[0075] S12, preprocessing the training data obtained in step S11 by data classification and data cleaning, wherein the preprocessed training data includes a predetermined number of calibrated mapping pairs, wherein the mapping pairs are {evaluation data; rule scoring};

[0076] S13, training the reward model using the preprocessed training data obtained in step S12;

[0077] S14. Convert the preprocessed training data obtained in step S12 into question-answering task data, and fine-tune the chatGLM model based on the question-answering task data to obtain a fine-tuned chatGLM model. During the process of fine-tuning the chatGLM model, continue to iteratively fine-tune the reward model to obtain a scoring model.

[0078] It should be noted that the general evaluation rules, security protection characteristics, security failure points and failure causes are the basic data for evaluating the effectiveness of security protection. Among them, the general evaluation rules provide standards or benchmarks on how to evaluate the effectiveness of security protection; security protection characteristics refer to specific attributes or working principles related to the security protection system; and security failure points and failure causes refer to weaknesses or loopholes in the security protection system and the causes of security failures.

[0079] Specifically, in the data classification of step S12, the training data is classified according to its type and content, for example, it can be classified according to different security protection types (such as physical security, network security, application security, etc.) or failure causes, and in the data cleaning of step S12, invalid data, duplicate data and erroneous data are removed, and the data is formatted in a unified manner to ensure data quality. After completing data classification and data cleaning, the security protection features, security failure points and failure causes are used as evaluation data, and each evaluation data is scored and calibrated through rough manual intervention according to the obtained general evaluation rules, and each evaluation data is grouped with its corresponding rule scoring group to form an {evaluation data; rule scoring} mapping pair.

[0080] Specifically, in step S13, each mapping pair obtained is used as a training data set of the reward model to train the reward model so that it can learn the mapping relationship between the evaluation data and the rule score, and then can accurately score the security protection system.

[0081] In this embodiment, in order to use machine learning technology to understand and parse natural language texts related to security protection, so as to improve the automation and intelligence level of security protection effectiveness evaluation, the chatGLM model is used as a language processing model to perform semantic parsing and content optimization processing on the text data in the security protection effectiveness evaluation. The chatGLM model is a pre-trained large-scale language model that already has the ability to handle complex language tasks. However, in order to make it perform better in a specific field, the method of this embodiment further fine-tunes it. Therefore, the method of this embodiment uses the description of the security incident in the evaluation data as a question (Question) and the corresponding rule score as an answer (Answer), thereby converting the mapping of the evaluation data and the rule score into a data pair of questions and answers. For example, if the evaluation data is a text describing firewall configuration Y, and the rule score is a score (such as 75 points), then a data pair similar to {Question: What is the effectiveness score of the given firewall configuration Y? ; Answer: 75 points} can be constructed. By fine-tuning the chatGLM model, it can learn domain-specific knowledge and patterns, so that it performs better in the specific task of security protection effectiveness evaluation.

[0082] Furthermore, in this embodiment, the reward model is iterated while fine-tuning the chatGLM model to achieve joint optimization of the two, ensuring that the two can work together during the evaluation process to provide more accurate and consistent results. The specific process is: the optimized evaluation rule expression and the optimized protection evaluation result expression generated during the fine-tuning of the chatGLM model are input into the reward model for scoring. According to the scoring results, the parameters of the reward model are updated to better reflect the mapping relationship between the evaluation data and the rule scoring. Repeat the above fine-tuning and updating process until the performance of the two models reaches a satisfactory level. In each iteration, the chatGLM model and the reward model are adjusted according to the latest data and feedback to gradually improve the accuracy and consistency of the evaluation. Both the fine-tuning chatGLM model and the iterative reward model use the same {evaluation data; rule scoring} mapping data to ensure that the two are optimized on the same data basis. At the same time, the reward model is updated by the optimized expression generated during the fine-tuning of the chatGLM model, forming a closed-loop optimization process, so that the two models can work together to provide more accurate and consistent evaluation results.

[0083] After the training phase is completed, it is necessary to further use the fine-tuned chatGLM model and the scoring model to score the effectiveness of security protection, that is, enter the inference phase. The specific process of the inference phase is as follows:

[0084] S21. Obtain the assessment rules and protection assessment results of the safety protection system;

[0085] S22, semantically encode the evaluation rule and the protection evaluation result respectively to obtain an evaluation rule semantic encoding feature vector and a protection evaluation result semantic encoding feature vector;

[0086] S23, performing multi-scale semantic interaction on the semantic encoding feature vector of the assessment rule and the semantic encoding feature vector of the protection assessment result to obtain an assessment rule-protection assessment result semantic interaction joint matrix;

[0087] S24. Use the fine-tuned chatGLM model and the scoring model to process the assessment rule-protection assessment result semantic interaction joint matrix to generate a security protection effectiveness assessment value.

[0088] Specifically, in step S21, the evaluation rules define the standards, indicators and methods used in the process of evaluating the effectiveness of the security protection system. The protection evaluation results are conclusive descriptions of the performance of the security protection system in actual applications.

[0089] It should be noted that in the process of conducting a quantitative evaluation of the effectiveness of security protection based on the evaluation rules and the protection evaluation results, in order to avoid misunderstandings caused by the different expressions of the evaluation rules and the protection evaluation results, the method of this embodiment expects to further utilize the interdependence between the evaluation rules and the protection evaluation results to optimize the content expression of the evaluation rules and the protection evaluation results, so as to ensure the semantic expression correlation between the two, thereby improving the accuracy of the evaluation.

[0090] To achieve this goal, in step S22, the evaluation rules and the protection evaluation results are semantically encoded, and the words or phrases in the text are converted into dense vector representations through word embeddings technology and self-attention mechanism, and the long-distance semantic dependencies between words are captured, so as to extract the key contextual semantic information in the evaluation rules and the protection evaluation results, thereby obtaining the evaluation rule semantic encoding feature vector and the protection evaluation result semantic encoding feature vector.

[0091] More specifically, the evaluation rule and the protection evaluation result may be respectively input into a semantic encoder based on a Transformer architecture for semantic encoding to obtain a semantic encoding feature vector of the evaluation rule and a semantic encoding feature vector of the protection evaluation result.

[0092] It should be noted that in order to comprehensively mine the semantic association relationship between the evaluation rules and the protection evaluation results, the method of this embodiment introduces the feature local-global multi-scale semantic interaction based on semantic feature decoupling to process the semantic encoding feature vector of the evaluation rules and the semantic encoding feature vector of the protection evaluation results. Through local semantic association interaction and global feature fusion, in-depth mining of the complex semantic relationship between the evaluation rules and the protection evaluation results is achieved.

[0093] Specifically, the specific process of step S23 is as follows: Figure 3 As shown:

[0094] S231, feature decoupling is performed on the semantic encoding feature vector of the evaluation rule to obtain a sequence of local semantic feature vectors of the evaluation rule, as specifically shown in the following formula:

[0095] Decouple(f2)={f 21 ,f 22 ,...,f 2i ,...,f 2l}

[0096] f2 is the semantic encoding feature vector of the evaluation rule, Decouple(·) represents the decoupling process, and f 21 ,f 22 ,...,f 2i ,...,f 2l They respectively represent the first, second, i-th and l-th evaluation rule local semantic feature vectors in the sequence of the evaluation rule local semantic feature vectors, and l represents the total number of vectors in the sequence of the evaluation rule local semantic feature vectors.

[0097] S232, performing a nonlinear transformation on the semantic coding feature vector of the protection assessment result to obtain a semantic coding feature vector of the protection assessment result after nonlinear transformation, which is specifically shown in the following formula:

[0098] f 1t =W′b1+f

[0099] f1 is the semantic encoding feature vector of the protection assessment result, W′ is the weight matrix, b is the bias vector, and f 1t Semantically encode the feature vector of the protection evaluation result after nonlinear transformation.

[0100] S233, respectively calculating the association matrix between the semantic coding feature vector of the protection evaluation result after nonlinear transformation and each evaluation rule local semantic feature vector in the sequence of the evaluation rule local semantic feature vector, to obtain a sequence of evaluation rule-protection evaluation result semantic interaction association matrices;

[0101] S234, calculating the characteristic interaction energy factor of each evaluation rule-protection evaluation result semantic interaction association matrix in the sequence of the evaluation rule-protection evaluation result semantic interaction association matrix to obtain a sequence of characteristic interaction energy factors;

[0102] S235. Based on the sequence of the feature interaction energy factors, the sequence of the evaluation rule-protection evaluation result semantic interaction association matrices is fused to obtain the evaluation rule-protection evaluation result semantic interaction joint matrix.

[0103] Specifically, in this embodiment, in step S233, the calculation process of the evaluation rule-protection evaluation result semantic interaction association matrix is: multiplying the transposed vector of the protection evaluation result semantic encoding feature vector after nonlinear transformation by the evaluation rule local semantic feature vector to obtain the corresponding evaluation rule-protection evaluation result semantic interaction association matrix, as shown in the following formula:

[0104]

[0105] (·) T represents the transpose of a vector, Represents matrix multiplication operation, M 12i It represents the semantic interaction association matrix of the ith evaluation rule and the protection evaluation result, that is, the association matrix between the semantic encoding feature vector of the protection evaluation result after nonlinear transformation and the local semantic feature vector of the ith evaluation rule.

[0106] Specifically, in step S234, the calculation process of the feature interaction energy factor is as follows: the square of the difference between the kurtosis and the feature mean of the evaluation rule-protection evaluation result semantic interaction association matrix, the feature variance of the evaluation rule-protection evaluation result semantic interaction association matrix, and the weighted sum of the bias items of the evaluation rule-protection evaluation result semantic interaction association matrix are calculated, and then the feature variance of the evaluation rule-protection evaluation result semantic interaction association matrix and the bias item are added and divided by the weighted sum to obtain the feature interaction energy factor, as shown in the following formula:

[0107]

[0108] E 12i represents the characteristic interaction energy factor of the semantic interaction association matrix of the i-th evaluation rule-protection evaluation result, k, μ and σ 2 denote the kurtosis, feature mean and feature variance of the semantic interaction matrix of the i-th evaluation rule-protection evaluation result, respectively, and λ is the bias term E{[M 12i (t)-μ] 4} represents the fourth-order central moment of the semantic interaction association matrix of the ith evaluation rule-protection evaluation result, that is, the expected value of the fourth power of the difference between each eigenvalue in the ith evaluation rule-protection evaluation result semantic interaction association matrix and its eigenvalue mean, and t represents each eigenvalue in the semantic interaction association matrix of the evaluation rule-protection evaluation result.

[0109] Specifically, the specific process of step S235 is: input the sequence of feature interaction energy factors into the Sigmoid function to obtain a sequence of normalized feature interaction energy factors; use the sequence of normalized feature interaction energy factors as a sequence of weights, calculate the position-weighted sum of the sequence of the evaluation rule-protection evaluation result semantic interaction association matrix, and obtain the evaluation rule-protection evaluation result semantic interaction joint matrix. As shown in the following formula:

[0110] E norm,12i =sigmoid(E 12i )

[0111] M=∑ i M 12i E norm,12i

[0112] sigmoid represents the sigmoid activation function, E norm,12i represents the ith normalized feature interaction energy factor, and M represents the evaluation rule-protection evaluation result semantic interaction joint matrix.

[0113] It should be noted that in step S23, the evaluation rule semantic encoding feature vector is decomposed into a sequence of evaluation rule local semantic feature vectors through a decoupling operation, which can reveal different dimensions or aspects within it and enhance the model's ability to capture local features. At the same time, the protection evaluation result semantic encoding feature vector is nonlinearly transformed, which can introduce more complex feature interactions and enhance the model's expressiveness. Next, the association matrix between the protection evaluation result semantic encoding feature vector after nonlinear transformation and the decoupled local semantic feature vectors of each evaluation rule is calculated, which can capture the interaction between feature vectors, reveal the correlation and information flow between features, and thus form a sequence of interactive association matrices. Then, by further calculating the feature interaction energy factors of each interactive association matrix, the intensity and importance of the interaction between feature vectors can be revealed, which helps to identify and strengthen the most critical feature interactions. Furthermore, the feature interaction energy factor sequence is normalized by a Sigmoid function so that its value range falls within the (0,1) interval, and is converted into an attention weight, so as to perform weighted summation on the interactive association matrix sequence, thereby integrating the interactive information of local features and obtaining the evaluation rule-protection evaluation result semantic interaction joint matrix. In this way, a comprehensive response representation of the assessment rules and protection assessment results is provided at the global level, which helps to enhance the model's understanding of the overall feature interaction patterns between the assessment rules and protection assessment results.

[0114] Specifically, in step S24, the evaluation rule-protection evaluation result semantic interaction joint matrix is ​​input into the fine-tuned chatGLM model for processing. The fine-tuned chatGLM model uses its powerful semantic understanding ability, combined with the pre-trained knowledge and the domain-specific knowledge learned in the fine-tuning stage, to conduct in-depth analysis and understanding of the semantic interaction information in the evaluation rule-protection evaluation result semantic interaction joint matrix, which can effectively improve the accuracy and clarity of the semantic expression of the evaluation rules and protection evaluation results, and generate more accurate and targeted optimized evaluation rule expressions and optimized protection evaluation result expressions.

[0115] Then, the optimized evaluation rule expression and the optimized protection evaluation result expression are input into the scoring model to obtain a safety protection effectiveness evaluation value. The scoring model can accurately score according to the optimized evaluation rule expression and the optimized protection evaluation result expression based on the mapping relationship between the evaluation data and the rule scoring learned in the training phase, thereby generating a quantitative safety protection effectiveness evaluation value.

[0116] Specifically, Figure 4 As shown, the specific process of step S24 is as follows:

[0117] S241, inputting the evaluation rule-protection evaluation result semantic interaction joint matrix into the fine-tuned chatGLM model to obtain an optimized evaluation rule expression and an optimized protection evaluation result expression;

[0118] S242. Input the optimized evaluation rule expression and the optimized protection evaluation result expression into the scoring model to obtain the security protection effectiveness evaluation value.

[0119] In this embodiment, the specific process of step S241 is: calculate the sum of the absolute values ​​of each eigenvalue of the evaluation rule-protection evaluation result semantic interaction joint matrix to obtain the first evaluation rule-protection evaluation result semantic interaction joint and modulation value, and calculate the square root of the sum of the squares of each eigenvalue of the evaluation rule-protection evaluation result semantic interaction joint matrix to obtain the second evaluation rule-protection evaluation result semantic interaction joint and modulation value; specifically as shown in the following formula:

[0120]

[0121]

[0122] m i ∈M∈R W×H

[0123] n=W×H

[0124] Where M represents the evaluation rule-protection evaluation result semantic interaction matrix, α is the first evaluation rule-protection evaluation result semantic interaction matrix and modulation value, β is the second evaluation rule-protection evaluation result semantic interaction matrix and modulation value, m i represents the i-th eigenvalue of the evaluation rule-protection evaluation result semantic interaction joint matrix, n is the number of eigenvalues ​​in the evaluation rule-protection evaluation result semantic interaction joint matrix; W and H represent the number of columns and rows of the evaluation rule-protection evaluation result semantic interaction joint matrix respectively, and R represents a real number set.

[0125] After performing point subtraction on the evaluation rule-protection evaluation result semantic interaction joint matrix and the second evaluation rule-protection evaluation result semantic interaction joint and modulation value, perform point multiplication with the number of eigenvalues ​​of the evaluation rule-protection evaluation result semantic interaction joint matrix and the inverse of the first evaluation rule-protection evaluation result semantic interaction joint and modulation value to calculate matrix one, and take the inverse of each eigenvalue in matrix one to obtain the first evaluation rule-protection evaluation result semantic interaction joint phase conversion matrix; specifically as shown in the following formula:

[0126]

[0127] M1 represents the first evaluation rule-protection evaluation result semantic interaction joint phase transformation matrix, ⊙ represents the dot product, Indicates point reduction, [·] ⊙-1 It means taking the inverse of each eigenvalue in the matrix.

[0128] After point-wise subtraction of the evaluation rule-protection evaluation result semantic interaction joint matrix and the first evaluation rule-protection evaluation result semantic interaction joint and modulation value, point-wise multiplication is performed with the square root of the number of eigenvalues ​​of the evaluation rule-protection evaluation result semantic interaction joint matrix and the inverse of the second evaluation rule-protection evaluation result semantic interaction joint and modulation value to calculate matrix two, and take the inverse of each eigenvalue in matrix two to obtain the second evaluation rule-protection evaluation result semantic interaction joint phase conversion matrix. Specifically shown in the following formula:

[0129]

[0130] M2 represents the second evaluation rule-protection evaluation result semantic interaction joint phase transformation matrix.

[0131] Calculate the dot product vector of the second evaluation rule-protection evaluation result semantic interaction joint phase transformation matrix and the weighted hyperparameter, and then perform dot subtraction between the first evaluation rule-protection evaluation result semantic interaction joint phase transformation matrix and the dot product vector to obtain the optimized evaluation rule-protection evaluation result semantic interaction joint matrix; specifically as shown in the following formula:

[0132]

[0133] M′ is the optimized evaluation rule-protection evaluation result semantic interaction joint moment, and ω is a weighted hyperparameter.

[0134] The optimized evaluation rule-protection evaluation result semantic interaction joint matrix is ​​input into the fine-tuned chatGLM model to obtain the optimized evaluation rule expression and the optimized protection evaluation result expression.

[0135] It should be noted that, in step S241, considering that the evaluation rule semantic coding feature vector and the protection evaluation result semantic coding feature vector respectively represent the semantic coding features of the evaluation rule and the protection evaluation result, when they are subjected to feature local-global interactive responses based on feature decoupling, the semantic association feature distribution pattern based on the source text distribution will also have local-global text-generation prediction interactive response differences. Therefore, it is expected to improve the detailed interactive response semantic aggregation expression effect of the evaluation rule-protection evaluation result semantic interaction joint matrix based on the association pattern representation differences.

[0136] Therefore, in this embodiment, the difference of the difference and modulation representation of the eigenvalue of the evaluation rule-protection evaluation result semantic interaction joint matrix relative to the overall feature set of the evaluation rule-protection evaluation result semantic interaction joint matrix is ​​used as the semantic change intensity information, and a phase-like transformation corresponding to the position-based intensity modulation is performed through different and modulation representations, so that the aggregation enhancement of the semantic change phase perception can enhance the axial aggregation receptive field along the feature aggregation direction by performing a spatial translation operation based on alternating stacking under the set scale equalization of the evaluation rule-protection evaluation result semantic interaction joint matrix, thereby improving the perception effect of the aggregated semantics of the evaluation rule-protection evaluation result semantic interaction joint matrix on the detailed semantic changes, so as to improve the expression effect of the evaluation rule-protection evaluation result semantic interaction joint matrix, and improve the accuracy of the optimized evaluation rule expression and the optimized protection evaluation result expression obtained by inputting it into the fine-tuned chatGLM model.

[0137] In the method of this embodiment, there is also a potential assessment step, which is intended to assess the defense and response capabilities of an organization in the field of cyberspace security. The potential assessment step is based on the attack restoration method, and implements a verification method combined with the assessment system to quickly identify the existing security protection equipment, the current defense status of the security protection system, and the potential assessment. It uses attack restoration assessment technology, large model analysis technology, security log comparison and analysis technology, and network security defense situation awareness to form an attack and defense potential assessment system, forming an attack and defense construction combined with a reverse assessment verification system, ensuring that the assessment system framework effectively identifies the effectiveness of security protection, and efficiently identifies and improves the shortcomings of the security protection system.

[0138] More specifically, Figure 5 As shown, the step of evaluating potential includes:

[0139] (1) Generation of attack restoration assessment cases and design of assessment rules, including: analyzing attack types and multi-source intelligence to obtain attack methods and attack events, restoring attack methods and attack events, and then inputting them into the assessment case and assessment rule generation framework.

[0140] (2) Assessment of attack perception capabilities of security protection equipment, including: accurately assessing the positioning response status through a large-scale model-based attack load response model, assessing attack loads, and specific semantic markers.

[0141] (3) Assessment of attack resistance potential, including: assessment of attack and defense threat models, assessment of typical network attack and defense scenarios, simulation of attack scenarios combined with multi-source intelligence, and assessment of potential target risk status.

[0142] In this embodiment, the attack load response model based on the large model implements security protection effectiveness verification for security failure point judgment and analysis, policy configuration scheme / special detection rules, such as Figure 6 As shown in the figure, the training data of the attack load response model based on the large model includes network security equipment protection features, network security equipment documentation manuals, general security detection rules, general detection rule-special rule mapping pairs, inspection case execution features and known security failure causes. The model used is: ChatGLM2-13b, the inference computing power is: A6000 80G*2, the training computing power is: A100 80G*2 (LORA fine-tuning), A800 80G*8 (full parameter fine-tuning).

[0143] More specifically, first, capture the general evaluation rules, obtain the security protection features, security failure points and failure causes for data preparation; second, classify the prepared data by clustering, form a mapping relationship through rough manual intervention, and discard the data that cannot be matched; then, after accumulating a certain number of calibrated mapping pairs, use the data to train the reward model to form a scoring model, and use the scoring model to score more evaluation data and rules, and record the data that meets the requirements for the next step of training. Finally, convert the mapping pair data into question-answering task data, fine-tune the chatGLM model, input the evaluation rules and protection evaluation results into the fine-tuned chatGLM model, and score the output results using the scoring model. In the process, the scoring model is iteratively fine-tuned synchronously. Finally, the training of the chatGLM model and the scoring model is completed, and the output results are retained in order of high and low scores, and work together with the marking, positioning and response status evaluation technology of the attack payload to achieve accurate judgment of the response status.

[0144] For those skilled in the art, various corresponding changes and modifications can be made according to the above technical solutions and concepts, and all of these changes and modifications should be included in the protection scope of the claims of the present invention.

Claims

1. A security protection effectiveness evaluation method based on a reward model and a chatGLM model, characterized in that: Includes training phase and inference phase; The specific process of the training phase is as follows: S11, obtaining training data; the training data includes general evaluation rules, safety protection features, safety failure points and failure causes; S12, pre-processing the training data obtained in step S11 by data classification and data cleaning, and then taking the safety protection features, safety failure points and failure causes as evaluation data, scoring and calibrating each evaluation data through rough manual intervention according to the obtained general evaluation rules, and grouping each evaluation data with its corresponding rule scoring to form an {evaluation data; rule scoring} mapping pair; S13, training the reward model using the preprocessed training data obtained in step S12; S14, converting the preprocessed training data obtained in step S12 into question-answering task data, and fine-tuning the chatGLM model based on the question-answering task data to obtain a fine-tuned chatGLM model. In the process of fine-tuning the chatGLM model, continue to iterate the reward model to obtain a scoring model; The specific process of the inference stage is as follows: S21. Obtain the assessment rules and protection assessment results of the safety protection system; S22, semantically encode the evaluation rule and the protection evaluation result respectively to obtain an evaluation rule semantic encoding feature vector and a protection evaluation result semantic encoding feature vector; S23, performing multi-scale semantic interaction on the semantic encoding feature vector of the assessment rule and the semantic encoding feature vector of the protection assessment result to obtain an assessment rule-protection assessment result semantic interaction joint matrix; S24. Use the fine-tuned chatGLM model and the scoring model to process the assessment rule-protection assessment result semantic interaction joint matrix to generate a security protection effectiveness assessment value.

2. The method according to claim 1, characterized in that In step S14, in the process of fine-tuning the chatGLM model, the specific process of continuing to iteratively fine-tune the reward model to obtain the scoring model is: inputting the optimized evaluation rule expression and the optimized protection evaluation result expression generated in the chatGLM model fine-tuning process into the reward model for scoring, and updating the parameters of the reward model according to the scoring results of the reward model to better reflect the mapping relationship between the evaluation data and the rule scoring, repeating the process of chatGLM model fine-tuning and reward model updating until the performance of the chatGLM model and the reward model reaches a satisfactory level, and finally obtaining the fine-tuned chatGLM model and scoring model.

3. The method according to claim 1, characterized in that In step S22, the evaluation rules and the protection evaluation results are semantically encoded, and the words or phrases in the text are converted into dense vector representations through word embedding technology and self-attention mechanism, and the long-distance semantic dependencies between words are captured, so as to extract the key contextual semantic information in the evaluation rules and the protection evaluation results, and obtain the semantic encoding feature vector of the evaluation rules and the semantic encoding feature vector of the protection evaluation results.

4. The method according to claim 1, characterized in that The specific process of step S23 is: S231, feature decoupling is performed on the semantic encoding feature vector of the evaluation rule to obtain a sequence of local semantic feature vectors of the evaluation rule, as specifically shown in the following formula: Decouple(f2)={f 21 ,f 22 ,...,f 2i ,...,f 2l } f2 is the semantic encoding feature vector of the evaluation rule, Decouple(·) represents the decoupling process, and f 21 ,f 22 ,...,f 2i ,...,f 2l respectively represent the first, second, i-th and l-th evaluation rule local semantic feature vectors in the sequence of the evaluation rule local semantic feature vectors, and l represents the total number of vectors in the sequence of the evaluation rule local semantic feature vectors; S232, performing a nonlinear transformation on the semantic coding feature vector of the protection assessment result to obtain a semantic coding feature vector of the protection assessment result after nonlinear transformation, which is specifically shown in the following formula: f 1t =W′f1+b f1 is the semantic encoding feature vector of the protection assessment result, W′ is the weight matrix, b is the bias vector, and f 1t Semantically encode feature vectors for protection evaluation results after nonlinear transformation; S233, respectively calculating the association matrix between the semantic coding feature vector of the protection evaluation result after nonlinear transformation and each evaluation rule local semantic feature vector in the sequence of the evaluation rule local semantic feature vector, to obtain a sequence of evaluation rule-protection evaluation result semantic interaction association matrices; S234, calculating the characteristic interaction energy factor of each evaluation rule-protection evaluation result semantic interaction association matrix in the sequence of the evaluation rule-protection evaluation result semantic interaction association matrix to obtain a sequence of characteristic interaction energy factors; S235. Based on the sequence of the feature interaction energy factors, the sequence of the evaluation rule-protection evaluation result semantic interaction association matrices is fused to obtain the evaluation rule-protection evaluation result semantic interaction joint matrix.

5. The method according to claim 4, characterized in that In step S233, the calculation process of the evaluation rule-protection evaluation result semantic interaction association matrix is: multiply the transposed vector of the protection evaluation result semantic coding feature vector after nonlinear transformation by the evaluation rule local semantic feature vector to obtain the corresponding evaluation rule-protection evaluation result semantic interaction association matrix, as shown in the following formula: (·) T represents the transpose of a vector, Represents matrix multiplication operation, M 12i It represents the semantic interaction association matrix of the ith evaluation rule and the protection evaluation result, that is, the association matrix between the semantic encoding feature vector of the protection evaluation result after nonlinear transformation and the local semantic feature vector of the ith evaluation rule.

6. The method according to claim 5, characterized in that In step S234, the calculation process of the feature interaction energy factor is as follows: the square of the difference between the kurtosis and the feature mean of the evaluation rule-protection evaluation result semantic interaction association matrix, the feature variance of the evaluation rule-protection evaluation result semantic interaction association matrix, and the weighted sum of the bias items of the evaluation rule-protection evaluation result semantic interaction association matrix are calculated, and then the feature variance of the evaluation rule-protection evaluation result semantic interaction association matrix and the bias item are added and divided by the weighted sum to obtain the feature interaction energy factor, as shown in the following formula: E 12i represents the characteristic interaction energy factor of the semantic interaction association matrix of the i-th evaluation rule-protection evaluation result, k, μ and σ 2 denote the kurtosis, feature mean and feature variance of the semantic interaction matrix of the i-th evaluation rule-protection evaluation result, respectively, and λ is the bias term E{[M 12i (t)-μ] 4 }, represents the fourth-order central moment of the semantic interaction association matrix of the ith evaluation rule-protection evaluation result, that is, the expected value of the fourth power of the difference between each eigenvalue in the ith evaluation rule-protection evaluation result semantic interaction association matrix and its eigenvalue mean; t represents each eigenvalue in the semantic interaction association matrix of the evaluation rule-protection evaluation result.

7. The method according to claim 6, characterized in that The specific process of step S235 is: inputting the sequence of feature interaction energy factors into the Sigmoid function to obtain a sequence of normalized feature interaction energy factors; using the sequence of normalized feature interaction energy factors as a sequence of weights, calculating the position-weighted sum of the sequence of the evaluation rule-protection evaluation result semantic interaction association matrix, and obtaining the evaluation rule-protection evaluation result semantic interaction joint matrix; As shown below: E norm,12i =sigmoid(E 12i ) M=∑ i M 12i E norm,12i sigmoid represents the sigmoid activation function, E norm,12i represents the i-th normalized feature interaction energy factor, and M represents the evaluation rule-protection evaluation result semantic interaction joint matrix.

8. The method according to claim 1, characterized in that The specific process of step S24 is as follows: S241, inputting the evaluation rule-protection evaluation result semantic interaction joint matrix into the fine-tuned chatGLM model to obtain an optimized evaluation rule expression and an optimized protection evaluation result expression; S242, inputting the optimized evaluation rule expression and the optimized protection evaluation result expression into the scoring model to obtain the safety protection effectiveness evaluation value; The specific process of step S241 is: calculate the sum of the absolute values ​​of each eigenvalue of the evaluation rule-protection evaluation result semantic interaction joint matrix to obtain the first evaluation rule-protection evaluation result semantic interaction joint and modulation value, and calculate the square root of the sum of the squares of each eigenvalue of the evaluation rule-protection evaluation result semantic interaction joint matrix to obtain the second evaluation rule-protection evaluation result semantic interaction joint and modulation value; specifically as shown in the following formula: m i ∈M∈R W×H n=W×H Where M represents the evaluation rule-protection evaluation result semantic interaction matrix, α is the first evaluation rule-protection evaluation result semantic interaction matrix and modulation value, β is the second evaluation rule-protection evaluation result semantic interaction matrix and modulation value, m i represents the i-th eigenvalue of the evaluation rule-protection evaluation result semantic interaction joint matrix, n is the number of eigenvalues ​​in the evaluation rule-protection evaluation result semantic interaction joint matrix; W and H represent the number of columns and rows of the evaluation rule-protection evaluation result semantic interaction joint matrix, respectively, and R represents a real number set; After performing point subtraction on the evaluation rule-protection evaluation result semantic interaction joint matrix and the second evaluation rule-protection evaluation result semantic interaction joint and modulation value, perform point multiplication with the number of eigenvalues ​​of the evaluation rule-protection evaluation result semantic interaction joint matrix and the inverse of the first evaluation rule-protection evaluation result semantic interaction joint and modulation value to calculate matrix one, and take the inverse of each eigenvalue in matrix one to obtain the first evaluation rule-protection evaluation result semantic interaction joint phase conversion matrix; specifically as shown in the following formula: M1 represents the first evaluation rule-protection evaluation result semantic interaction joint phase transformation matrix, ⊙ represents the dot product, Indicates point reduction, [·] ⊙-1 It means taking the inverse of each eigenvalue in the matrix; After performing point subtraction on the evaluation rule-protection evaluation result semantic interaction joint matrix and the first evaluation rule-protection evaluation result semantic interaction joint and modulation value, perform point multiplication with the square root of the number of eigenvalues ​​of the evaluation rule-protection evaluation result semantic interaction joint matrix and the inverse of the second evaluation rule-protection evaluation result semantic interaction joint and modulation value to calculate the matrix 2, and take the inverse of each eigenvalue in the matrix 2 to obtain the second evaluation rule-protection evaluation result semantic interaction joint phase conversion matrix; specifically as shown in the following formula: M2 represents the second evaluation rule-protection evaluation result semantic interaction joint phase transformation matrix; Calculate the dot product vector of the second evaluation rule-protection evaluation result semantic interaction joint phase transformation matrix and the weighted hyperparameter, and then perform dot subtraction between the first evaluation rule-protection evaluation result semantic interaction joint phase transformation matrix and the dot product vector to obtain the optimized evaluation rule-protection evaluation result semantic interaction joint matrix; specifically as shown in the following formula: M′ is the optimized evaluation rule-protection evaluation result semantic interaction joint moment, and ω is a weighted hyperparameter; The optimized evaluation rule-protection evaluation result semantic interaction joint matrix is ​​input into the fine-tuned chatGLM model to obtain the optimized evaluation rule expression and the optimized protection evaluation result expression.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

10. A computer device, characterized in that: The method comprises a processor and a memory, wherein the memory is used to store a computer program; when the processor is used to execute the computer program, the method described in any one of claims 1 to 8 is implemented.

Images