Artificial general intelligent safety assessment method and system based on AHP and genetic algorithm

Abstract

The invention discloses an artificial universal intelligent security assessment method and system based on an AHP and a genetic algorithm. According to the scheme, firstly, a data acquisition module extracts data features related to security from an AGI model; then the index hierarchy construction module establishes a safety index system based on an analytic hierarchy process; the weight calculation module calculates an initial weight and performs consistency verification; if the consistency ratio (CR) does not meet the threshold value, the genetic algorithm module executes optimization and correction; the optimized weight is input into a grey clustering module to calculate a clustering coefficient and a comprehensive safety score; and finally, the result output module generates a safety assessment report and a risk level. According to the scheme of the invention, a hierarchical structure system of AGI security risks is established through AHP, and consistency correction and global optimization are carried out by using a genetic algorithm, so that evaluation weight calculation is more objective and stable; and meanwhile, a grey clustering model and a whitening weight function are combined to realize quantitative classification of different risk types, so that a comprehensive safety score with dynamic adjustability and interpretability can be output.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence security assessment, specifically to a general artificial intelligence (AGI) security assessment method GASA (Genetic-AHP SecurityAssessment) based on the combination of the Analytic Hierarchy Process (AHP) and the Genetic Algorithm (GA), which is used to achieve multi-dimensional and quantifiable security analysis of AGI systems. Background Technology

[0002] The development of Artificial General Intelligence (AGI) has become a key focus of current artificial intelligence research. With the rapid evolution of Large Language Models (LLMs), AGI systems are exhibiting characteristics approaching human intelligence in areas such as reasoning, generation, and decision-making. However, this generalized intelligence brings complex security risks, such as adversarial attacks, jailbreak attacks, prompt injection, privacy breaches, and data corruption. These risks extend beyond the scope of traditional cybersecurity and AI security, placing higher demands on the reliability, interpretability, compliance, and controllability of AGI. Therefore, establishing a scientific and systematic AGI security assessment model has become a crucial prerequisite for research and application.

[0003] Existing security assessment research mainly focuses on three directions: first, attack model analysis centered on security risk classification; second, hierarchical assessment methods based on indicator system construction; and third, dynamic assessment models driven by intelligent algorithms. Nong et al. proposed a system failure risk framework for network attack classification, but its structural definition is unclear, making it difficult to refine the quantitative indicators for different attack categories. Alcaraz et al. proposed a classification framework based on availability, integrity, and confidentiality, and assessed the impact of network attacks on critical infrastructure. Subsequently, Cazorla et al. expanded this framework, adding an anomaly detection dimension, forming the AICAn (Availability, Integrity, Confidentiality, and Anomaly) classification system. However, these studies emphasize traditional information system security and do not adequately consider the complexity of AGI multidimensional risks.

[0004] On the other hand, AHP is widely used in cybersecurity and risk assessment scenarios. Jia et al. used AHP to construct a network attack indicator system, but failed to perform consistency correction on the pairwise comparison matrix (PCM), leading to bias in the evaluation weights. Al-Zewairi et al. and Bashaiwth et al. introduced artificial neural networks (ANN) and long short-term memory networks (LSTM) for attack classification, but these methods rely on a large amount of training data and are difficult to interpret the weight relationships. Ahmed et al. proposed a deep learning-based botnet detection model, but it is only applicable to specific attack types. Aldhaheri et al. combined deep learning with the dendritic cell algorithm (DeepDCA) for IoT attack detection, but the model's feature extraction was insufficient and its generalization ability was weak. Kim et al. proposed a TTP classification model for network threat intelligence classification, but it suffers from overfitting.

[0005] In addition, some studies have attempted to introduce multi-indicator comprehensive evaluation methods. For example, Huang et al. proposed a federated execution and evaluation dual network framework (EEFED) to achieve personalized intrusion detection, but this requires additional global detection model support. Li et al. constructed a network attack evaluation model based on variable weight theory and ideal solution ranking technique (TOPSIS), incorporating spatiotemporal factors into risk assessment, but it is computationally complex and has limited applicability. Kumar et al. proposed a web application security evaluation scheme based on hesitant fuzzy set AHP and TOPSIS, improving the hierarchical level of the evaluation. Jiskani et al. used gray clustering to evaluate the security of mining engineering, verifying the applicability of AHP and gray systems in uncertain environments.

[0006] Furthermore, Chinese invention patent application CN 120337234 A discloses a method for evaluating the security performance of large models based on the analytic hierarchy process (AHP) and Shapley value fusion. This method, designed for large models, focuses on static weight fusion and is suitable for relatively fixed risk assessment scenarios. It uses AHP and Shapley value fusion as its main approach, and the process includes: setting a security assessment index vector, calculating subjective weights using AHP, calculating objective weights using Shapley values, linearly fusing subjective and objective weights, and finally outputting a security score. However, this method relies on mathematical fusion and lacks a dynamic correction mechanism. Such a solution has the following problems in practical applications: 1. Static evaluation framework: It relies on fixed weight fusion and cannot dynamically adapt to the risk evolution during the operation of the AGI system, resulting in a lag in evaluation results; 2. Weak consistency handling: AHP matrix correction relies on manual intervention, CR verification is simplistic, and weight bias is easily generated; 3. Low computational efficiency: Shapley values ​​require enumerating all subsets of indicators. When the number of indicators m is large, the computation becomes infeasible, limiting its application in large-scale artificial intelligence systems. 4. Limited applicability: This solution is designed only for large models and does not take into account the multidimensional interaction risks of AGI, making it difficult to apply directly to more artificial intelligence scenarios.

[0007] 5. Lack of uncertainty mechanism: This method cannot effectively handle the incompleteness of indicator data, which reduces the robustness of the evaluation.

[0008] Overall, the existing technologies still have the following main shortcomings in AGI security assessment: (1) The assessment process relies on manual judgment, which is highly subjective and has poor consistency; (2) The indicator system lacks a dynamic weight adjustment mechanism, making it difficult to reflect complex and multidimensional risks; (3) There is a lack of automated consistency correction methods at the algorithm level, and AHP results are prone to deviation; (4) The model has insufficient universality and is difficult to extend to AGI full-domain security risk assessment.

[0009] Therefore, how to combine the hierarchical analysis advantages of AHP with the global search and optimization capabilities of genetic algorithms to achieve the objectivity and computability of AGI safety assessment has become an urgent technical problem to be solved. Furthermore, existing technologies generally lack research on interval judgment matrices, multi-objective consistency optimization, and dynamic risk feedback mechanisms. A comprehensive framework that can balance expert experience, algorithm stability, and data reliability in assessment has not yet been established. These issues have become key bottlenecks restricting the interpretability and reproducibility of AGI safety assessment. Summary of the Invention

[0010] To address the problems of strong subjectivity, poor evaluation consistency, lack of dynamic optimization mechanisms, and difficulty in comprehensively quantifying complex security risks in existing AGI security assessment methods, this invention aims to propose an AGI security assessment scheme based on the combination of AHP and genetic algorithms. By establishing a generalized and scalable AGI security assessment model, the scheme can achieve a scientific, objective, and computable security assessment process, and can be applied to different types of AGI systems and complex security scenarios, providing a reliable quantitative basis for subsequent model enhancement, security supervision, and risk prevention and control.

[0011] To achieve the above objectives, this invention provides a method for assessing the safety of general artificial intelligence based on AHP and genetic algorithms, comprising: Step 1: Extract security-related data features from the AGI model; Step 2: Based on the data features extracted in Step 1, establish a security indicator system using the Analytic Hierarchy Process (AHP).

[0012] Step 3: Generate a pairwise comparison matrix for the security indicator system constructed in Step 2, and further calculate the initial weights; Step 4: Perform consistency verification and dual-index constraint processing on the pairwise comparison matrix generated in Step 3 and the calculated initial weight vector; Step 5: For the pairwise comparison matrix that failed the consistency check in Step 4, take the initial pairwise comparison matrix and the consistency index as the correction objects, and generate the corrected P pairwise comparison matrix based on the multi-objective genetic algorithm, so that it satisfies the dual index constraint, and recalculate the weight vector. Step 6: Perform credibility fusion weight correction on the weight vector calculated in Step 3 or Step 5 and the credibility factor input from the outside; Step 7: Based on the whitening weight function, use the fusion weights calculated in Step 6 to weight the safety assessment data of each indicator, and perform gray clustering calculation to obtain the clustering coefficient of each indicator; Step 8: Generate output security assessment results.

[0013] Furthermore, in step 3, when generating the pairwise comparison matrix, the indicators at the same level are compared pairwise based on the indicator system in step 2, and values ​​are assigned according to the set importance to generate a positive-reciprocal matrix.

[0014] Furthermore, when performing dual verification on the pairwise comparison matrix generated in step 3 in step 4, it includes both traditional consistency ratio verification and K-consistency index verification.

[0015] Furthermore, in step 5, a multi-objective fitness function is constructed with the consistency ratio CR, the difference from the original matrix, and the local sensitivity as objectives. A genetic algorithm is then used to perform a global search and automatic correction on the pairwise comparison matrix until the consistency constraint is met and the optimal closeness to the expert judgment is maintained.

[0016] Furthermore, in step 6, the initial weights calculated by AHP are exponentially fused with the confidence factor q obtained from actual security observation data at a fixed ratio α to form a new fused weight vector.

[0017] Furthermore, in step 7, three types of whitening weight functions are designed for the security data of each indicator.

[0018] Furthermore, in step 8, the weights are dynamically updated using an exponential forgetting factor β.

[0019] Furthermore, the method also includes an online adaptive update step, which uses real-time safety monitoring data obtained from continuous monitoring of the AGI model to perform online update calculations on the weight vector obtained in step 6.

[0020] To achieve the above objectives, the present invention also provides an artificial general intelligent security assessment system based on AHP and genetic algorithms. The system includes a data acquisition module, an indicator hierarchy construction module, a weight calculation module, a consistency check and dual indicator constraint module, a multi-objective genetic algorithm correction module, a credibility fusion weight correction module, a gray clustering calculation module, and a result output module. The data acquisition module is configured to extract security-related data features from the AGI model; The indicator hierarchy construction module is configured to interact with the data acquisition module, and can use the analytic hierarchy process to establish a safety indicator system based on the relevant safety data features extracted from the AGI model by the data acquisition module. The weight calculation module is configured to interact with the indicator hierarchy construction module, and can generate a pairwise comparison matrix and calculate the initial weights based on the security indicator system and / or the corresponding expert evaluation results constructed by the indicator hierarchy construction module. The consistency verification and dual-index constraint module is configured to interact with the weight calculation module, and can perform consistency verification on the pairwise comparison matrix (PCM) generated by the weight calculation module and the calculated initial weight vector. The multi-objective genetic algorithm correction module is configured to interact with the consistency test and dual-index constraint module. It can target the pairwise comparison matrix (PCM) marked as inconsistent by the consistency test and dual-index constraint module, and use the original pairwise comparison matrix (PCM) and consistency index as correction objects. It then uses the multi-objective genetic algorithm to correct the PCM and generate a corrected PCM that satisfies the dual-index constraint, while recalculating the weight vector. The credibility fusion weight correction module is configured to interact with the multi-objective genetic algorithm correction module and the weight calculation module, and can perform credibility fusion weight correction on the weight vector calculated by the multi-objective genetic algorithm correction module and / or the weight calculation module, as well as the credibility factor input from the outside. The gray clustering calculation module is configured to interact with the credibility fusion weight correction module. It can use the whitening weight function and the fusion weight calculated by the credibility fusion weight correction module to weight the security assessment data of each indicator and perform gray clustering calculation to obtain the clustering coefficient of each indicator. The result output module is configured to generate and output security assessment results.

[0021] Furthermore, the system also includes an online adaptive update module, which is configured to interact with the credibility fusion weight correction module, the gray clustering calculation module, and the weight calculation module. It can perform online update calculations on the weight vectors obtained by the credibility fusion weight correction module based on the real-time security monitoring data continuously monitored by the AGI module.

[0022] This invention presents AGISE, an AGI security assessment method based on the combination of AHP and genetic algorithm. It constructs a multi-layer security index system through the analytic hierarchy process, incorporates weight optimization and grey clustering calculation, and uses a genetic algorithm to automatically correct the consistency of the pairwise comparison matrix (PCM), ensuring the objectivity and stability of the weight solution. This enables a comprehensive quantitative assessment of AGI systems in multiple dimensions, including network security, model reliability, service security, and data security, effectively overcoming the problems existing in the prior art. Attached Figure Description

[0023] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0024] Figure 1 This is a flowchart of the artificial general intelligence safety assessment method based on AHP and genetic algorithm in this invention.

[0025] Figure 2 This is a system framework diagram of the artificial general intelligent safety assessment system based on AHP and genetic algorithm in this invention.

[0026] Figure 3 This is an example diagram of the AGI safety risk classification and indicator system in this invention. Detailed Implementation

[0027] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below with reference to specific illustrations.

[0028] While existing technologies like AHP (Automatic Hierarchical Analysis) possess strong hierarchical analysis capabilities and can handle complex multi-dimensional indicators, their results heavily rely on human judgment, easily leading to consistency biases in PCM (Programmatical Process Management) and consequently affecting the accuracy of weight calculation. Traditional deep learning-based or TOPSIS-based assessment methods, although possessing automated feature extraction capabilities, lack theoretical interpretability and hierarchical logic, making it difficult to establish a reasonable weight allocation mechanism in multi-indicator, multi-level AGI security assessments. Furthermore, existing methods generally lack adaptive optimization strategies, failing to dynamically adjust indicator weights based on changes in the security posture of the AGI model under different scenarios, thus causing assessment results to lag behind risk evolution.

[0029] To address this, this invention constructs a hybrid evaluation framework that integrates AHP and genetic algorithms: AHP is used to establish a hierarchical structure of AGI safety risks, and genetic algorithms are used for consistency correction and global optimization, making the evaluation weight calculation more objective and stable; at the same time, by combining a gray clustering model and a whitening weight function, different risk types can be quantitatively classified, thereby outputting a comprehensive safety score with dynamic adjustability and interpretability.

[0030] Meanwhile, the present invention introduces a multi-objective genetic algorithm (MO-GA) optimization mechanism, a dual-constraint model of interval judgment matrix and K consistency index, and a credibility prior and online adaptive update mechanism to improve the robustness, dynamism and compliance of the evaluation system.

[0031] Based on the above scheme, this invention specifically forms a general artificial intelligence safety assessment method based on AHP and genetic algorithms. Combined with... Figure 1 As shown, the artificial general intelligence safety assessment method based on AHP and genetic algorithm presented in this invention mainly consists of the following steps: Step 1: Extract security-related data features from the AGI model; Step 2: Establish a security indicator system based on the data features extracted in Step 1 using the Analytic Hierarchy Process (AHP).

[0032] This step, based on the technical features established in Step 1, extracts relevant security data features from the master and slave AGI systems as the data source for the constructed security indicator system. This data includes security indicators that may be generated during the operation of the AGI model. This step is used to structure complex security risks, laying the foundation for subsequent quantitative analysis. The constructed indicator system will directly serve as the basic framework for subsequent weight calculations.

[0033] Step 3: Generate pairwise comparison matrices and calculate initial weights.

[0034] This step generates a pairwise comparison matrix based on the security indicator system and / or the corresponding expert evaluation results constructed in step 2, and further calculates the initial weights.

[0035] This step directly uses the indicator system constructed in step 2 as the source of elements for PCM, thereby transforming subjective judgments into quantitative weights.

[0036] Step 4: Consistency verification and dual-index constraints.

[0037] This step performs consistency verification and dual-index constraint processing on the pairwise comparison matrix (PCM) generated in step 3 and the calculated initial weight vector ω. The PCM serves as the direct object of consistency verification, while the weight vector is used to assist in calculating the consistency index. This consistency verification ensures that subsequent evaluations are based on mathematically consistent judgments, avoiding evaluation distortion caused by subjective bias.

[0038] This step uses the PCM and initial weights from step 3 as input. If the verification passes, the weights are directly used for the credibility fusion in step 6; if it fails, the genetic algorithm correction in step 5 is triggered.

[0039] Step 5: Multi-objective genetic algorithm correction.

[0040] This step targets the PCM marked as inconsistent in step 4, and uses the original PCM and consistency index as correction objects. It generates a corrected PCM based on a multi-objective genetic algorithm, ensuring that it meets the dual index constraints, and recalculates the weight vector ω. The corrected PCM and the recalculated weights are then passed to step 6 for credibility fusion.

[0041] As a further explanation, if the correction still fails, this step will be executed iteratively.

[0042] This step uses a genetic algorithm for correction, which enables automation and global optimization, and solves the subjective consistency problem of AHP.

[0043] Step 6: Confidence fusion weight correction.

[0044] This step performs credibility fusion weight correction on the weight vector ω calculated in step 3 or step 5, as well as the externally input credibility factor. This credibility fusion step enhances the robustness of the weights, makes the evaluation results more in line with the actual data quality, and reduces the absolute dominance of expert judgment.

[0045] This step further allows the corrected weights to be directly used in the grey clustering calculation in step 7 as a basis for weighting.

[0046] Step 7: Construct the whitening weight function to calculate the gray clustering factor.

[0047] This step uses the safety assessment data for each indicator and the fusion weight ω calculated in step 6. Based on the whitening weight function, it weights the safety assessment data for each indicator using the fusion weight ω calculated in step 6, and performs grey clustering calculation to obtain the clustering coefficients for each indicator. These coefficients are used to quantify the risk level, and finally, a comprehensive clustering coefficient is obtained. This step further passes the calculated clustering coefficients to step 8 or step 9 to calculate the overall score.

[0048] This step effectively handles the uncertainty in the assessment through grey clustering calculation, transforming multidimensional data into interpretable risk categories, which is the basis for visualizing the assessment results.

[0049] Step 8: Online adaptive update.

[0050] This step uses real-time safety monitoring data obtained from continuous monitoring by the AGI system to update the weight vector ω calculated in step 6 online, so that the updated weight vector ω can adapt to changes in the AGI safety situation.

[0051] The updated weights can be fed back into step 7 for grey clustering calculation, or used directly in the weight calculation in step 3 in the next round of evaluation.

[0052] The online updates based on this step ensure the dynamism of the present invention, enabling the evaluation model to self-adjust as the AGI system evolves, thus avoiding the lag of static models.

[0053] Step 9: Generate output security assessment results.

[0054] The following details the specific implementation scheme of each step in the artificial general intelligence safety assessment method proposed in this invention.

[0055] In some embodiments of the present invention, step 1 of this method specifically extracts structured features that can be mapped to a safety indicator system from quantifiable signals such as AGI interaction data and risk-triggered behaviors.

[0056] In this method, the extracted structured features are not used for classification or detection, but are directly used to drive the weight optimization of AHP and GA, forming a common constraint between expert judgment and the model's real behavior.

[0057] Furthermore, this method performs normalization, structuring, and time smoothing on the extracted raw data so that the processed data can participate in weight calculation and grey clustering. The processed data is also used to generate credibility factors to correct expert weights and support the dynamic update mechanism.

[0058] In some embodiments of the present invention, step 2 of this method constructs a hierarchical indicator system for AGI security risks using the analytic hierarchy process (AHP). First, the security data of the AGI system obtained in step 1 is used as the basis for indicator design; then, security risks are divided into four primary dimensions: network security (A), model reliability (B), service security (C), and data security and privacy protection (D).

[0059] Therefore, this method can structurally decompose complex AGI security risks into multi-level indicators. For example, Figure 3The proposed solution decomposes complex AGI security risks into primary and secondary indicators in a structured manner. This avoids omissions or overlaps in assessment dimensions and effectively addresses the lack of a dynamic mechanism in the construction of indicator systems in existing technologies. At the same time, the indicator design is tailored to the specific risks of AGI, thus differentiating it from traditional information system security assessments.

[0060] The system constructed by this method can comprehensively map the AGI security situation, providing standardized input for subsequent weight calculation and reducing subjective human intervention.

[0061] In some embodiments of the present invention, step 3 of this method is specifically implemented by having experts compare the indicators at the same level pairwise based on the indicator system constructed in step 2, assigning values ​​according to the importance scale (1-9 scale) in Table 1, and generating a positive reciprocal matrix PCM.

[0062] This step effectively transforms subjective judgments into mathematical matrices, providing calculable inputs for algorithm optimization, thereby effectively overcoming the limitations of existing technologies where the evaluation process relies on manual judgment. At the same time, compared with the traditional eigenvector method, LLSM is more sensitive to consistency deviations and can expose PCM inconsistency problems in advance, laying the foundation for subsequent genetic algorithm correction.

[0063] The weight calculation here is highly efficient, and the initial consistency ratio of the PCM constructed by experts often exceeds 0.1, which verifies the necessity of subsequent corrections and avoids evaluation distortion caused by weight bias.

[0064] In some embodiments of the present invention, step 4 of this method involves performing a dual verification on the PCM constructed in step 3, including a traditional consistency ratio (CR) verification and a K-consistency index verification. If both constraints are satisfied, the verification passes; otherwise, step 5 is triggered for correction.

[0065] This step introduces the K-index to supplement the global consistency check of CR, enhancing local logical consistency verification and addressing the issue of single CR checks being insensitive to high-order matrices. The K-threshold adapts to the matrix order, avoiding overly stringent checks that could lead to correction failures. This dual constraint reduces the false positive rate, automates verification to minimize human intervention, and provides explicit triggering conditions for genetic algorithm correction.

[0066] In some embodiments of the present invention, step 5 of this method is specifically implemented by constructing a multi-objective fitness function with the consistency ratio CR, the difference from the original matrix, and the local sensitivity as objectives, and using a genetic algorithm to perform a global search and automatic correction on the pairwise comparison matrix until the consistency constraint is met and the optimal closeness to the expert judgment is maintained.

[0067] This implementation scheme unifies the consistency correction of AHP, the preservation of expert semantics, and the robustness of sensitivity into a genetic optimization framework in a multi-objective manner, realizing an automated and convergently controllable correction mechanism for PCM. This allows the optimized matrix to achieve higher consistency and stability while maintaining the semantics of expert judgment, thereby significantly improving the objectivity, reliability, and reproducibility of the evaluation results of weight calculation.

[0068] In some embodiments of the present invention, step 6 of this method is specifically implemented by exponentially fusing the initial weights calculated by AHP with the credibility factor q obtained from actual security observation data at a fixed ratio α to form a new fused weight vector, so that expert judgment and objective data jointly participate in the determination of the final weights.

[0069] This implementation incorporates "expert semantic weights" and "model behavior credibility" into the same mathematical framework, using an exponential fusion method to maintain their non-linear influence and avoid biases caused by simple linear weighting. This allows the fused weights to maintain the interpretability of AHP while improving their alignment with real-world risk situations, resulting in a more stable, objective, and dynamically adaptable assessment outcome that reflects actual security conditions.

[0070] In some embodiments of the present invention, step 7 of this method specifically involves designing three types of whitening weight functions for the security data of each indicator. This better characterizes the nonlinear features of security risks, outperforming traditional linear functions and avoiding subjective setting biases; and gray clustering is used to effectively handle missing or uncertain data.

[0071] In some embodiments of this invention, step 8 of this method is preferably implemented by dynamically updating the weights using an exponential forgetting factor β. By balancing the influence of historical weights with the latest data using the β factor, drastic fluctuations in the evaluation results are avoided, ensuring the continuity of the scores. Simultaneously, noise is added during the update process to meet AGI data compliance requirements, aligning with the "compliance" requirements in the background art. This allows the solution to respond quickly to emerging risks.

[0072] In some embodiments of the present invention, step 9 of this method specifically generates a safety assessment report and risk level.

[0073] The artificial general intelligence security assessment scheme based on AHP and genetic algorithms presented in this example can be configured into a corresponding software program to form a corresponding artificial general intelligence security assessment system based on AHP and genetic algorithms. When running, this software program will execute the aforementioned artificial general intelligence security assessment method based on AHP and genetic algorithms, and simultaneously store the results in a suitable storage medium for the processor to retrieve and execute.

[0074] See Figure 2 The resulting Artificial General Intelligence Safety Assessment System 100 based on AHP and genetic algorithms mainly includes the following functional architecture: a data acquisition module 110, an indicator hierarchy construction module 120, a weight calculation module 130, a consistency verification and dual indicator constraint module 140, a multi-objective genetic algorithm correction module 150, a credibility fusion weight correction module 160, a gray clustering calculation module 170, an online adaptive update module 180, and a result output module 190. Furthermore, each module forms a logical closed loop through data flow and parameter interfaces to realize the entire AGI safety assessment process.

[0075] The data acquisition module 110 in the system is configured to extract security-related data features from the AGI model.

[0076] The indicator hierarchy construction module 120 in the system is configured to interact with the data acquisition module 110, and can use the analytic hierarchy process to establish a safety indicator system based on the relevant safety data features extracted by the data acquisition module 110 from the AGI system.

[0077] The data involved here includes safety metrics that may be generated during the operation of the AGI model.

[0078] This module enables the structuring of complex security risks, laying the foundation for subsequent quantitative analysis.

[0079] The weight calculation module 130 in the system is configured to interact with the indicator hierarchy construction module 120, and can generate a pairwise comparison matrix and calculate the initial weights based on the security indicator system and / or the corresponding expert evaluation results constructed by the indicator hierarchy construction module 120.

[0080] This module uses the indicator system constructed by the indicator hierarchy construction module 120 as the source of elements for PCM, thereby enabling subjective judgments to be transformed into quantitative weights.

[0081] The consistency verification and dual-index constraint module 140 in the system is configured to interact with the weight calculation module 130, and can perform consistency verification on the pairwise comparison matrix (PCM) generated by the weight calculation module 130 and the calculated initial weight vector ω.

[0082] This module uses PCM as the direct object of consistency verification, and the weight vector is used to assist in calculating the consistency index. Through the consistency verification of this module, it can ensure that the subsequent evaluation is based on mathematically consistent judgments and avoid evaluation distortion caused by subjective bias.

[0083] The multi-objective genetic algorithm correction module 150 in the system is configured to interact with the consistency test and dual-index constraint module 140. It can identify PCMs marked as inconsistent by the consistency test and dual-index constraint module 140, and use the original PCM and consistency index as correction objects. It can then use the multi-objective genetic algorithm to correct the PCM and generate a corrected PCM that satisfies the dual-index constraint, while recalculating the weight vector ω.

[0084] This module specifically employs a genetic algorithm for correction, enabling automation and global optimization, and solving the subjective consistency problem of AHP.

[0085] The credibility fusion weight correction module 160 in the system is configured to interact with the multi-objective genetic algorithm correction module 150 and the weight calculation module 130, and can perform credibility fusion weight correction on the weight vector ω calculated by the multi-objective genetic algorithm correction module 150 and / or the weight calculation module 130, as well as the externally input credibility factor.

[0086] The gray clustering calculation module 170 in the system is configured to interact with the credibility fusion weight correction module 160. It can weight the security assessment data of each indicator based on the whitening weight function and the fusion weight ω calculated by the credibility fusion weight correction module 160, and perform gray clustering calculation to obtain the clustering coefficient of each indicator. This coefficient is used to quantify the risk level and finally aggregate to obtain the comprehensive clustering coefficient.

[0087] The online adaptive update module 180 in the system is configured to interact with the credibility fusion weight correction module 160, the gray clustering calculation module 170, and the weight calculation module 130. It can update the weight vector ω obtained by the credibility fusion weight correction module 160 online based on the real-time security monitoring data obtained by the continuous monitoring of the AGI system, so that the updated weight vector ω can adapt to the changes in the AGI security situation.

[0088] This online adaptive update module 180 can further feed the updated weights back to the gray clustering calculation module 170 for gray clustering calculation, or directly use them for weight calculation in the weight calculation module 130 in the next round of evaluation.

[0089] The system's result output module 190 is configured to interact with the gray clustering calculation module 170 to generate and output security assessment results.

[0090] The implementation process of the AGI safety assessment system (GASA) based on AHP and genetic algorithms is divided into four core stages, forming a complete closed loop: First, the indicator system is constructed. The data acquisition module extracts security feature data from the AGI system, and the indicator hierarchy construction module establishes a four-level risk assessment system (dimensions such as network security and model reliability) based on this, laying a structured foundation for quantitative assessment. Second, the weight optimization and correction: the weight calculation module generates a pairwise comparison matrix and calculates the initial weights through expert judgment. After being verified by the consistency check module (CR<0.1 and K index double constraint), the genetic algorithm module performs multi-objective optimization on the unqualified matrix, and then the credibility fusion module balances the subjective and objective weights. Third, dynamic evaluation and calculation: the gray clustering module uses the whitening weight function to calculate the clustering coefficient of each risk level, and combines the optimized weights to obtain a comprehensive score; the online adaptive update module dynamically adjusts the weights through the exponential forgetting factor to ensure that the model evolves with the security situation. Fourth, the results output module aggregates all data to generate quantitative safety scores and risk level reports, supporting continuous monitoring and decision-making by the AGI system. The entire process, through modular collaboration, achieves objectivity, adaptability, and interpretability in the assessment process.

[0091] This system achieves tight coupling between modules through data flow and parameter constraints. The data acquisition module first standardizes the extracted model behavior features and then passes them to the indicator hierarchy construction module, enabling the indicator system to align with real risk performance. The constructed hierarchy is input into the weight calculation module to generate a pairwise comparison matrix, which is then processed by the consistency check module and the genetic algorithm correction module to form a closed loop of detection and optimization. The optimized weights, along with the objective values ​​output by the credibility fusion module, enter the gray clustering module to achieve quantitative classification of multidimensional risks. The clustering results are dynamically adjusted by the online adaptive update module and then written back to the weight pool to support the next cycle of evaluation iteration. Finally, the calculation results of all modules are uniformly summarized by the output module to form a traceable and interpretable safety scoring chain.

[0092] The following specific examples further illustrate the implementation process and corresponding technical features of the artificial general intelligence safety assessment scheme based on AHP and genetic algorithms proposed in this invention.

[0093] For ease of explanation, this section will first provide a detailed description of the data units, algorithms, and functional concepts involved in the specific implementation of this solution.

[0094] (1) Paired Comparison Matrix (PCM): Positive and negative matrices obtained by pairwise comparison of elements at the same level , satisfy , And for all ,have ; Indicates the importance degree of element i relative to element j.

[0095] (2)Importance scale and description [as shown in Table 1]: If matrix A is a PCM, The value is determined according to Table 1.

[0096] Table 1 Importance scale and its description

[0097] In this example, the use of the pairwise comparison matrix can transform the complex, multi-dimensional and difficult-to-directly-quantify AGI security risk relationship into a computable structured numerical form, enabling expert experience to participate in the weight reconstruction model in a consistent and comparable manner; at the same time, the reciprocity and consistency mechanisms of the matrix provide a mathematical basis for subsequent genetic algorithm optimization and weight correction, thus significantly enhancing the stability, objectivity and interpretability of the evaluation results.

[0098] (3)Consistency and random consistency index: Define the calculation method of the consistency index (CI) of PCM as: (1) where is the main eigenvalue of the PCM and n is the order of the PCM.

[0099] Another defined random consistency index RI is the average random consistency index, and the specific values are shown in Table 2.

[0100] The consistency ratio (CR) is obtained by comparing with the appropriate RI value, and the calculation formula is: (2) For PCM, if CR = 0, then the PCM is completely consistent. If 0 < CR < 0.1, the consistency of the PCM is acceptable. Otherwise, the PCM should be modified until .

[0101] Table 2 Matrix order and corresponding RI values

[0102] In this example, the introduction of the consistency index CI, the random consistency index RI and the consistency ratio CR is used to conduct a mathematical verification of the pairwise comparison matrix given by experts, so as to ensure the logical consistency and reliability of the weight calculation. Through the constraint of CR < 0.1, the system can quickly discover the subjective biases and contradictory relationships in the judgment matrix and trigger the genetic algorithm for automatic correction, making the final weights both conform to the expert judgment logic and maintain mathematical consistency; thereby enhancing the stability, interpretability of the weight distribution and the credibility of the evaluation results.

[0103] (4) Weight calculation method: Using the log-least-square method (LLSM), let... For one Determine the matrix, Let be a uniform matrix given by the following formula: (3) Required priority vector This can be represented as an optimization constraint problem: (4) (5) In this example, the logarithmic least squares method is used to calculate the weights, which can obtain the optimal and approximately consistent weight vector even with slight judgment bias. This makes the weight calculation optimal in the sense of minimizing error, thereby improving the accuracy and stability of the weights of each indicator. At the same time, this calculation method can also automatically smooth out the local inconsistency problem in expert judgment, making the final weights more consistent with the overall judgment trend, and providing high-quality initial values ​​for subsequent genetic algorithm optimization.

[0104] (5) Genetic Algorithm: The fitness function is a mapping of the optimal solution considering two types of constraints, expressed as: (6) in It is a conditional function. It is a constraint function.

[0105] The genetic algorithm used in this example performs a global search and correction of the pairwise comparison matrix through a multi-objective fitness function. This simultaneously minimizes consistency deviation and the amount of modification to the original expert matrix, achieving a balance between consistency, reliability, and stability in the corrected PCM. Its randomness and population search mechanism avoid getting trapped in local optima, thus significantly improving the reliability and automation of matrix consistency repair.

[0106] (6) AGI safety risk classification and indicator system See Figure 3 Based on the main security risks during the operation of the AGI system, this example divides security risks into four primary dimensions: network security, model reliability, service security, and data security and privacy protection. Each dimension is further refined into several secondary indicators. For example, network security includes indicators such as DDoS attacks, jailbreaking, and leaked prompt words; model reliability includes indicators such as interpretability, availability, and robustness. By establishing a hierarchical model, a multi-level mapping of the indicator system is achieved.

[0107] Based on the above technologies, this example constructs an artificial general intelligent security assessment system based on AHP and genetic algorithms. The system includes a data acquisition module, an indicator hierarchy construction module, a weight calculation module, a consistency check and dual indicator constraint module, a multi-objective genetic algorithm correction module, a credibility fusion weight correction module, a gray clustering calculation module, an online adaptive update module, and a result output module. Furthermore, the modules form a logical closed loop through data flow and parameter interfaces to realize the entire AGI security assessment process.

[0108] During operation, the main control module uniformly schedules the various functional modules within the system and executes them in the following order: data input → index system loading → pairwise comparison matrix generation → weight calculation → consistency check and dual index constraint → multi-objective genetic algorithm correction → credibility fusion weight correction → grey clustering calculation → online adaptive update → score output and report generation.

[0109] Specifically, the process for implementing AGI security assessment in this example system is as follows: (1) Constructing a PCM for AGI safety risk types and safety indicators: According to the appendix Figure 3 The designed AGI safety assessment classification constructs a pairwise comparison matrix based on expert evaluation results. ,in The value of represents the importance of indicator i and indicator j, and its value is taken from Table 1; Simultaneously, based on the designed AGI safety index system, a series of pairwise comparison matrices were constructed using expert evaluation results. ,in The value of represents the importance of indicator i and indicator j, and its value is taken from Table 1.

[0110] (2) Weight calculation and consistency check: For each PCM, the priority vector ω is calculated using LLSM to represent the weight of different security risks at different times. Then, CI and CR are calculated according to formulas (1) and (2), and a consistency index is added in addition to the traditional consistency ratio CR. : ; Improve the fitness function so that the optimization objective includes not only minimizing It also minimizes the difference from the expert's original matrix. and sensitivity measurement ,accomplish: ; Where CR(A) is the consistency ratio of the pairwise comparison matrix A; This indicates the degree of difference between the corrected matrix and the original expert judgment matrix, used to maintain the validity of expert judgments; This is the sensitivity of the matrix to small perturbations, used to measure the stability of the weights and prevent the results from being excessively affected by fluctuations in a small amount of data.

[0111] Based on this, by further sorting and selecting the results of multi-objective optimization, the matrix that simultaneously meets the consistency requirements, has the smallest deviation from expert judgment, and has the lowest sensitivity can be selected first, so that the final modified matrix achieves the comprehensive optimality among the three indicators of mathematical consistency, weight stability, and expert credibility.

[0112] In addition, a credibility factor is introduced for each indicator. The weights are adjusted as follows: ; in This is a balancing coefficient (0.6~0.8), used to adjust the proportion of influence between expert judgment and data credibility.

[0113] (3) Consistency correction in genetic algorithm: Construct constraint functions f and condition functions g in the fitness function: ; Where P1 and P2 are: ; P1 is used to constrain whether the weight vector meets the basic probability distribution requirements (such as the weight sum being close to 1) to ensure the validity of the search solution; P2 is the differential weighting term of adjacent weights, where φ is the position weight function and λ is the smoothing factor, used to control the smoothness of weight changes and avoid unreasonable jumps.

[0114] (4) Construction of whitening weight function and grey clustering: Based on the different attributes of each indicator, a corresponding exponential whitening weight function is designed. Let p be the gray interval item and s be the number of gray intervals, then the whitening weight function for the first gray class is expressed as: ; In the last gray class, the whitening weight function is expressed as: ; In the remaining gray classes, the whitening weight function is expressed as: ; in This represents the boundary value of the i-th indicator in the p-th gray level interval, used to determine the range to which the indicator value belongs in different risk levels.

[0115] (5) Online adaptive update mechanism: During AGI operation, when new safety observation data flows in, the system uses an exponential forgetting factor. Dynamically update weights: ； Ensure the model can respond quickly to new situations while maintaining scoring continuity and smoothness.

[0116] (6) Calculation of comprehensive clustering coefficient and overall score: After the aforementioned weight optimization and grey clustering calculations are completed, this step aggregates the clustering results by risk type and performs the final scoring. Specifically, here, secondary indicators are aggregated according to risk type to obtain the comprehensive clustering coefficients for the four risk types. Then, weight the risk type layer. The weighted sum is used to obtain the overall score. The comprehensive clustering coefficient for an AGI security risk type under each gray class is: ； The final overall result is: ; according to The value can determine the level of the attack.

[0117] The overall score calculated from this It can be called by the results output module as the final quantitative output of the GASA evaluation.

[0118] As a further explanation, the AGI security risk classification and indicator system in this example includes four categories of primary risks: network security (A), model reliability (B), service security (C), and data security and privacy protection (D).

[0119] Each risk category defines several secondary indicators, such as: Category A (A1–A5): DDoS attacks, defense bypass, information leakage, etc. Category B (B1–B6): Interpretability, robustness, access control, etc.; Category C (C1–C2): Content responsiveness management, abuse detection; Category D (D1–D4): Data storage security, privacy protection, etc.

[0120] As an example, the code for forming an extensible data structure is as follows: RiskCategories = { "A": ["A1", "A2", "A3", "A4", "A5"], "B": ["B1", "B2", "B3", "B4", "B5", "B6"], "C": ["C1", "C2"], "D": ["D1", "D2", "D3", "D4"] } As an example, the code for generating pairwise comparison matrices in this instance is as follows: def calc_weights(A): # A is a pairwise comparison matrix logA = np.log(A) n = A.shape[0] w = np.exp(np.sum(logA, axis=1) / n) return w / np.sum(w) Using sorting and selection, ensure that the matrix satisfies : def fitness_multi(A, A0): cr = calc_CR(A) mod = np.linalg.norm(np.log(A) - np.log(A0), ord=1) sens = local_sensitivity(A) return (cr, mod, sens) Next, credibility fusion is performed, the code is as follows: def fuse_prior_weights(w_llsm, q, alpha=0.7): w = (w_llsm ** alpha) * (q ** (1 - alpha)) return w / w.sum() Furthermore, a whitening weight function is then constructed based on the attribute type of the indicator and the evaluation data to calculate the gray clustering factor for each indicator. The code is as follows: def mu_first(x: float, b0: float, b1: float) -> float: if x <= b0: return 0.0 if x >= b1: return 1.0 return (x - b0) / (b1 - b0) def mu_last(x: float, bP_1: float, bP: float) -> float: if x <= bP_1: return 1.0 if x >= bP: return 0.0 return (bP - x) / (bP - bP_1) def mu_middle(x: float, bm1: float, b: float, bp1: float) -> float: if x <= bm1 or x >= bp1: return 0.0 if x < b: return (x - bm1) / (b - bm1) return (bp1 - x) / (bp1 - b) def cluster_factor(xs: List[float], ws: List[float], bounds: Dict[int, List[float]], p: int) -> Tuple[float, float]: assert len(xs) == len(ws) total = 0.0 for i, (x, w) in enumerate(zip(xs, ws)): bs = bounds[i] if p == 1: mu = mu_first(x, bs[0], bs[1]) elif p == len(bs) - 1: mu = mu_last(x, bs[-2], bs[-1]) else: mu = mu_middle(x, bs[p-1], bs[p], bs[p+1]) total += w * mu denom = float(np.sum(ws)) Gamma_p = total gamma_p = Gamma_p / denom return Gamma_p, gamma_p Furthermore, an online adaptive update is then performed, with the following code: def update_streaming(prev_w, new_w, beta=0.9): w = beta * prev_w + (1 - beta) * new_w return w / w.sum() def dp_perturb(x, eps, delta_sens): noise = np.random.laplace(0.0, delta_sens / eps, size=len(x)) return x + noise Finally, a weighted summation is performed based on the weights to obtain a comprehensive score, ranging from 1 to 10, corresponding to security risk levels from low to high. The clustering coefficients and comprehensive scores for the four types of security risks are then output.

[0121] As can be seen from the above technical solutions and specific examples, the GASA method provided by this invention can realize the safety assessment of artificial general intelligence (AGI) systems. It has produced significant technical progress and practical effects in addressing the problems of strong subjectivity, poor consistency and unstable assessment results of traditional AHP methods.

[0122] I. Improved technical performance and computing efficiency This GASA method combines AHP and GA to achieve consistency correction and objectification in the calculation of security assessment weights for AGI systems. Based on actual experimental results, the final consistency ratio of each pairwise comparison matrix meets the criterion of CR < 0.1, indicating that the GA-corrected matrix has good consistency and stability and can accurately reflect the relative importance of different indicators. Through a hierarchical analysis structure, the model achieves quantifiable assessment of four primary risk types: network security, model reliability, service security, and data security and privacy protection. Therefore, this GASA method maintains good mathematical consistency and interpretability in the weight determination process.

[0123] II. Experimental Verification and Analysis Results Experiments were conducted using two datasets to demonstrate that the proposed GASA method can output reasonable comprehensive security scores on both datasets. The clustering coefficients γ for the four security risk categories showed significant differences across the different datasets: the clustering coefficients for network security and data security were relatively low in Dataset 1, while they were significantly higher in Dataset 2, corresponding to a higher comprehensive score. This result validates that the GASA model can effectively distinguish the security differences between different types of AGI systems and comprehensively reflect multi-dimensional risks.

[0124] Furthermore, experimental results show that the weight distribution of each risk type and its subordinate indicators is consistent with the actual risk performance. For example, cybersecurity and data security have relatively high weights in the overall assessment, corresponding to the most significant dominance in the comprehensive score, indicating that the model can reflect the key factors of system security posture from the perspective of weight allocation. By combining grey clustering and whitening weight functions, GASA has good adaptability and robustness in handling the uncertainty of multidimensional indicators, and can maintain stable classification results under different data input conditions.

[0125] III. Applications and Scientific Value This GASA method can be embedded into AI security assessment and monitoring systems to automate security risk calculation and classification, reduce human judgment errors, and improve the objectivity and repeatability of the assessment process. In scientific research applications, this GASA method provides a unified mathematical model and verifiable framework for the security assessment of artificial intelligence systems; in industry practice, it can be used for AI system security compliance assessment, model deployment review, and continuous monitoring, demonstrating significant potential for widespread application.

[0126] Based on the above-mentioned artificial general intelligence security assessment scheme based on AHP and genetic algorithm, this embodiment of the invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the steps of the above-mentioned artificial general intelligence security assessment method based on AHP and genetic algorithm.

[0127] This invention also provides a processor for running a program, wherein the program executes the steps of the above-described artificial general intelligent security assessment method based on AHP and genetic algorithms.

[0128] This invention also provides a terminal device, which includes a processor, a memory, and a program stored in the memory and executable on the processor. The program code is loaded and executed by the processor to implement the steps of the above-described artificial general intelligent security assessment method based on AHP and genetic algorithms.

[0129] The present invention also provides a computer program product, which, when executed on a data processing device, is adapted to perform the steps of the above-described artificial general intelligent security assessment method based on AHP and genetic algorithms.

[0130] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0131] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0132] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0133] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0134] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0135] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0136] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0137] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0138] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0139] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0140] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0141] The method, specific system unit, or part thereof of the present invention described above is a pure software architecture. It can be deployed via program code on physical media, such as hard disks, optical discs, or any electronic device (such as smartphones or computer-readable storage media). When a machine loads and executes the program code (e.g., a smartphone loads and executes it), the machine becomes a device for implementing the present invention. The method and device of the present invention can also be transmitted in program code form via transmission media, such as cables, optical fibers, or any other transmission method. When the program code is received, loaded, and executed by a machine (e.g., a smartphone), the machine becomes a device for implementing the present invention.

[0142] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A general artificial intelligence security assessment method based on AHP and genetic algorithms, characterized in that, include: Step 1: Extract security-related data features from the AGI model; Step 2: Based on the data features extracted in Step 1, establish a security indicator system using the Analytic Hierarchy Process (AHP). Step 3: Generate a pairwise comparison matrix for the security indicator system constructed in Step 2, and further calculate the initial weights; Step 4: Perform consistency verification and dual-index constraint processing on the pairwise comparison matrix generated in Step 3 and the calculated initial weight vector; Step 5: For the pairwise comparison matrix that failed the consistency check in Step 4, take the initial pairwise comparison matrix and the consistency index as the correction objects, and generate the corrected P pairwise comparison matrix based on the multi-objective genetic algorithm, so that it satisfies the dual index constraint, and recalculate the weight vector. Step 6: Perform credibility fusion weight correction on the weight vector calculated in Step 3 or Step 5 and the credibility factor input from the outside; Step 7: Based on the whitening weight function, use the fusion weights calculated in Step 6 to weight the safety assessment data of each indicator, and perform gray clustering calculation to obtain the clustering coefficient of each indicator; Step 8: Generate output security assessment results.

2. The artificial general intelligence safety assessment method based on AHP and genetic algorithm according to claim 1, characterized in that, In step 3, when generating the pairwise comparison matrix, the indicators at the same level are compared pairwise based on the indicator system in step 2, and values ​​are assigned according to the set importance to generate a positive and negative matrix.

3. The artificial general intelligence safety assessment method based on AHP and genetic algorithm according to claim 1, characterized in that, When performing dual verification on the pairwise comparison matrix generated in step 3 in step 4, it includes both traditional consistency ratio verification and K-consistency index verification.

4. The artificial general intelligence safety assessment method based on AHP and genetic algorithm according to claim 1, characterized in that, In step 5, a multi-objective fitness function is constructed with the consistency ratio CR, the difference from the original matrix, and the local sensitivity as objectives. A genetic algorithm is then used to perform a global search and automatic correction on the pairwise comparison matrix until the consistency constraint is met and the optimal closeness to the expert judgment is maintained.

5. The artificial general intelligence safety assessment method based on AHP and genetic algorithm according to claim 1, characterized in that, In step 6, the initial weights calculated by AHP are exponentially fused with the confidence factor q obtained from actual security observation data at a fixed ratio α to form a new fused weight vector.

6. The artificial general intelligence safety assessment method based on AHP and genetic algorithm according to claim 1, characterized in that, In step 7, three types of whitening weight functions are designed for the security data of each indicator.

7. The artificial general intelligence safety assessment method based on AHP and genetic algorithm according to claim 1, characterized in that, In step 8, the weights are dynamically updated using the exponential forgetting factor β.

8. The artificial general intelligence security assessment method based on AHP and genetic algorithm according to claim 1, characterized in that, The method further includes an online adaptive update step, which uses real-time safety monitoring data obtained from continuous monitoring of the AGI model to perform online update calculations on the weight vector obtained in step 6.

9. A general-purpose artificial intelligence safety assessment system based on AHP and genetic algorithms, characterized in that, The system includes a data acquisition module, an indicator hierarchy construction module, a weight calculation module, a consistency check and dual indicator constraint module, a multi-objective genetic algorithm correction module, a credibility fusion weight correction module, a gray clustering calculation module, and a result output module. The data acquisition module is configured to extract security-related data features from the AGI model; The indicator hierarchy construction module is configured to interact with the data acquisition module, and can use the analytic hierarchy process to establish a safety indicator system based on the relevant safety data features extracted from the AGI model by the data acquisition module. The weight calculation module is configured to interact with the indicator hierarchy construction module, and can generate a pairwise comparison matrix and calculate the initial weights based on the security indicator system and / or the corresponding expert evaluation results constructed by the indicator hierarchy construction module. The consistency verification and dual-index constraint module is configured to interact with the weight calculation module, and can perform consistency verification on the pairwise comparison matrix (PCM) generated by the weight calculation module and the calculated initial weight vector. The multi-objective genetic algorithm correction module is configured to interact with the consistency test and dual-index constraint module. It can target the pairwise comparison matrix (PCM) marked as inconsistent by the consistency test and dual-index constraint module, and use the original pairwise comparison matrix (PCM) and consistency index as correction objects. It then uses the multi-objective genetic algorithm to correct the PCM and generate a corrected PCM that satisfies the dual-index constraint, while recalculating the weight vector. The credibility fusion weight correction module is configured to interact with the multi-objective genetic algorithm correction module and the weight calculation module, and can perform credibility fusion weight correction on the weight vector calculated by the multi-objective genetic algorithm correction module and / or the weight calculation module, as well as the credibility factor input from the outside. The gray clustering calculation module is configured to interact with the credibility fusion weight correction module. It can use the whitening weight function and the fusion weight calculated by the credibility fusion weight correction module to weight the security assessment data of each indicator and perform gray clustering calculation to obtain the clustering coefficient of each indicator. The result output module is configured to generate and output security assessment results.

10. The artificial general intelligent safety assessment system based on AHP and genetic algorithm according to claim 9, characterized in that, The system also includes an online adaptive update module, which is configured to interact with the credibility fusion weight correction module, the gray clustering calculation module, and the weight calculation module. It can perform online update calculations on the weight vectors obtained by the credibility fusion weight correction module based on the real-time security monitoring data continuously monitored by the AGI module.

Images