A method and system for generating a path of risk assessment of a differential mode attack on a power system based on a high-dimensional attack vector matrix

Abstract

The application discloses a power system differential mode attack risk assessment path generation method and system based on a high-dimensional attack vector matrix. The method obtains the current system state and the multi-source attack modeling information required for power system security assessment, forms a threat behavior primitive set for security testing, and forms or imports a high-dimensional attack vector matrix; generates a security test action space and syntax constraints based on the high-dimensional attack vector matrix, generates a legal action set or an action mask vector according to the current system state; inputs the current system state and the legal action set into a strategy generation model, and selects an abstract test action under controlled constraints by the strategy generation model; maps the abstract test action into an abstract simulation event in a power system soft simulation environment, and updates the system state, path history, candidate influence agent score and differential mode risk candidate attributes according to the soft simulation feedback; when the path termination condition is met, a risk assessment report is output, which is comprehensive, objective and accurate.

Description

Technical Field

[0001] This invention relates to a method and system for generating differential attack risk assessment paths in power systems based on high-dimensional attack vector matrices, belonging to the technical fields of power system network security simulation, risk assessment and security testing. Background Technology

[0002] As the digitalization, networking, and intelligence of power systems continue to improve, services such as dispatch control, station automation, distributed energy access, electricity consumption data collection, and charging facility management are increasingly reliant on complex information and communication networks and industrial control protocols. Power system cybersecurity assessments need to simultaneously consider the impact of potential threat behaviors on business links, protocol semantics, equipment topology, and defense response during the multi-stage chain evolution process.

[0003] Differential attacks refer to attacks where attackers exploit vulnerabilities or backdoors in specific versions or environments of target hardware, software, or protocols to launch differentiated attacks that are environment-dependent. As long as the specific attack method depends on and is limited by a specific hardware architecture, software version, or protocol implementation, it can be classified as a "differential attack." In power system security assessment scenarios, modeling such differentiated risks aims to identify potential security weaknesses under specific equipment models, software versions, protocol implementations, operating environments, or business links within a controlled software simulation environment. This provides a basis for protection strategy verification, risk localization, test case design, and system hardening. Existing power system security testing methods often rely on static rule bases, expert experience, or manually orchestrated test paths. These methods typically struggle to uniformly express protocol semantics, business impact, traffic characteristics, device topology, access capabilities, implementation environment dependencies, and defense feedback. They also lack the ability to uniformly constrain the phase sequence, capability transfer, topology reachability, protocol preconditions, mutual exclusion relationships, and unlocking relationships between test actions. Therefore, there is an urgent need for a method that can represent power system security assessment knowledge in a high-dimensional structure and generate differential attack risk assessment paths under syntactic constraints and soft simulation feedback, so as to improve the automation and standardization of power system network security testing, risk assessment and protection capability verification. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for generating differential attack risk assessment paths for power systems based on a high-dimensional attack vector matrix. This method addresses problems in existing power system security testing, such as insufficient dimensions of attack knowledge representation, reliance on manual experience in attack path construction, lack of unified grammatical constraints on attack actions, difficulty in labeling differential attack candidate attributes, and lack of soft simulation feedback loop in the attack path generation process. This improves the accuracy of power system risk assessment, thereby enhancing the effectiveness of power grid security early warning and defense, and improving the security and stability of the power system.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A method for generating a path for assessing differential mode attacks in power systems based on a high-dimensional attack vector matrix, characterized by the following steps: 1) Obtain the current system state and the multi-source attack modeling information required for power system differential attack modeling. The current system state is used to characterize at least the topology state, access state, defense state and historical action state in the power system soft simulation environment. The multi-source attack modeling information includes at least attack investigation information, protocol semantic information, business link information, traffic spatiotemporal information, equipment topology information and soft simulation capability information. 2) Normalize the multi-source attack modeling information and generate a set of attack action primitives based on the normalized information, wherein the i-th attack action primitive is represented as:

[0006] In the formula, Indicates the first Each attack action primitive; Indicates the identity information of the action; Indicates information about the action phase; Represents information about the target object; This indicates the preconditions for the action; Represents protocol semantic information; Indicates information about the impact on business; Indicates the topological scope information; This indicates test mapping information; 3) Construct or import a high-dimensional attack vector matrix, which includes multiple rows of attack vectors corresponding to the attack action primitives. Each row of attack vectors corresponds to an abstract attack action and includes 12 sets of standardized fields: action identity field group, attack stage and access capability field group, protocol semantic field group, business behavior field group, traffic spatiotemporal field group, device topology field group, software and hardware implementation specific field group, differential mode impact field group, defense response field group, software simulation mapping field group, policy interface field group, and test case generation field group. The standardized attack vector corresponding to the i-th abstract attack action is obtained by fusing field group encodings:

[0007] In the formula, This represents the fused feature vector of the i-th abstract attack action; , , , , and These represent the encoding results of the attack phase, protocol semantics, business behavior, traffic spatiotemporal, device topology, and implementation environment fields, respectively. This represents the fusion weight matrix for the corresponding field group; 4) Generate an attack action space and attack syntax constraints based on the high-dimensional attack vector matrix, wherein the comprehensive feasibility evaluation value of the abstract attack action is calculated according to the constraint attributes of different field groups in the attack vector:

[0008] In the formula, This represents the comprehensive feasibility evaluation value of the i-th abstract attack action; This represents the constraint feature value of the k-th field group corresponding to the i-th attack action; This represents the weight coefficient of the corresponding field group; m represents the number of field groups involved in the calculation. 5) Combining the current system state, attack action space, and attack syntax constraints, perform a six-fold joint legality judgment on each candidate abstract attack action in the attack action space to generate a set of legal actions or an action mask vector. The six conditions are: access capability condition, target existence condition, target type matching condition, topology reachability condition, defense status condition, and repeated execution restriction condition. When a candidate abstract attack action meets all the conditions, the corresponding candidate abstract attack action is added to the set of legal actions, or the value of the corresponding index position of the candidate abstract attack action in the action mask vector is set to a first preset value to indicate that the candidate abstract attack action is in an optional state; when any condition is not met, the corresponding candidate abstract attack action is excluded from the set of legal actions, or the value of the corresponding index position of the candidate abstract attack action in the action mask vector is set to a second preset value to indicate that the candidate abstract attack action is in a blocked state, and the corresponding illegal reason is recorded.

[0009] The mask value of the i-th candidate abstract attack action in the action mask vector at time t is determined by the following formula:

[0010] In the formula, Indicates the index of the candidate abstract attack action. Indicates the current decision-making moment. This represents the first preset value used to characterize the optional state. This represents a second preset value used to characterize the shielding state, and ; , , , , and They represent the first One candidate abstract attack action Whether the following conditions are met at any given time: access capability condition, target existence condition, target type matching condition, topology reachability condition, defense status condition, and repeated execution restriction condition; the value of each determination result is either a first preset value or a second preset value, where the first preset value is 1 and the second preset value is 0; When the result of the judgment of the above six conditions is 1, Add the corresponding candidate abstract attack action to the set of legal actions; when the result of any condition is 0, The corresponding candidate abstract attack actions are excluded from the set of legal actions, and the corresponding illegal reasons are recorded. Through the above joint calculation, only candidate abstract attack actions that simultaneously satisfy constraints such as access capability, target matching, topology reachability, and defense status are allowed to enter the subsequent strategy selection. By using the above action mask calculation method, the six legality conditions are jointly judged, so that candidate abstract attack actions that do not meet any of the legality conditions are deterministically blocked before entering the strategy generation model, thereby reducing invalid actions participating in path generation and improving the legality, interpretability and generation efficiency of risk assessment paths. 6) Input the current system state and the set of legal actions or action mask vector into the policy generation model. The policy generation model calculates the evaluation value of each candidate abstract attack action based on the current system state and action constraint information, and selects the abstract attack action with the highest evaluation value from the set of legal actions.

[0011] In the formula, This represents the abstract attack action selected at time t; This represents the i-th candidate abstract attack action; This represents the set of legal actions after filtering based on the action mask vector; This indicates that the policy generation model is in the current system state. Next action The calculated action evaluation value; 7) Map the abstract attack action to an abstract simulation event in the power system soft simulation environment, and obtain the soft simulation feedback; 8) Update the current system state, path history, composite candidate influence agent score, and differential attack candidate attributes based on the soft simulation feedback; The current system state is iteratively updated based on the action execution results and software simulation feedback:

[0012] In the formula, This represents the current system state at time t. This represents the abstract attack action executed at time t. This represents the feedback information from the soft simulation. This represents the state update function; X （t+1） Characterizes the system state at time t+1 after iterative update; 9) When the path termination condition is met, output the differential attack risk assessment path, attack test cases, and result report.

[0013] In step 3) above, the field groups of the high-dimensional attack vector matrix include at least: action identity field group, attack stage and access capability field group, protocol semantic field group, business behavior field group, traffic spatiotemporal field group, device topology field group, hardware and software implementation specificity field group, differential impact field group, defense response field group, software simulation mapping field group, policy interface field group, and test case generation field group. Through these field groups, the system can uniformly represent dispersed attack knowledge, protocol semantics, business impact, topology relationships, capability conditions, defense feedback, and simulation feedback into a computable, constrainable, simulable, and policy-generating data structure.

[0014] In step 2) above, the attack action primitives include: action name, action identifier, attack stage, required access capability, target node type, target protocol type, preconditions, post-execution state change, candidate impact description, detection risk description, defense response description, soft simulation event mapping relationship, and test case generation fields. The attack action primitives are abstract attack action descriptions in a software simulation environment, and do not contain real attack payloads, real device addresses, real control messages, or real vulnerability exploitation scripts.

[0015] The specific steps of step 5) above are: to determine the legality of each candidate abstract attack action in the attack action space; The legality determination includes at least access capability conditions, target existence conditions, target type matching conditions, topology reachability conditions, defense status conditions, and repeated execution restriction conditions; When the candidate abstract attack action simultaneously satisfies the access capability condition, target existence condition, target type matching condition, topology reachability condition, defense status condition, and repeated execution restriction condition, the value of the corresponding index position of the candidate abstract attack action in the action mask vector is set to the first preset value to indicate that the candidate abstract attack action is in an optional state. When the candidate abstract attack action does not meet any of the following conditions: access capability condition, target existence condition, target type matching condition, topology reachability condition, defense status condition, and repeated execution restriction condition, the value of the corresponding index position of the candidate abstract attack action in the action mask vector is set to a second preset value to indicate that the candidate abstract attack action is in a blocked state, and the corresponding illegal reason is recorded.

[0016] In step 5) above, the current system state includes: the current set of accessible nodes, the current set of controllable nodes, the current set of reachable nodes, node semantic features, topological adjacency relationships, historical action sequences, defense status, detection status, business operation status, protocol semantic status, and candidate influence proxy scoring status. The current system state is initialized based on the power system soft simulation environment at the beginning and is iteratively updated with the soft simulation feedback.

[0017] In step 6) above, the policy generation model is a reinforcement learning policy model. The reinforcement learning policy model performs action value estimation or policy selection based on graph structure state representation. The graph structure state representation includes the node semantic features, the topological adjacency relationship, the global state, and the action mask vector. The reinforcement learning policy model outputs the action value or action selection result of the candidate abstract attack actions, and selects the next abstract attack action under the constraint of the action mask vector.

[0018] In step 6) above, the strategy generation model does not make unconstrained selections from all attack actions, but rather selects actions from the set of legal actions, thereby reducing invalid actions that do not meet the attack phase, access capability, topology reachability, or protocol preconditions.

[0019] In step 7) above, the abstract attack action is mapped to an abstract simulation event in the power system soft simulation environment, including: Based on the software simulation mapping field corresponding to the abstract attack action, determine its target object, action stage, target node type, protocol semantic category, business impact category, and state change category; Trigger the corresponding abstract simulation event in the power system soft simulation environment; The abstract simulation event is used to change the topology state, access state, service state, protocol semantic state, defense state, or candidate influence agent scoring state in the power system software simulation environment, without generating real attack messages, real control commands, or real vulnerability exploitation behaviors.

[0020] In step 8) above, the composite candidate impact proxy score includes at least one of the following: observation difference score, business impact score, physical impact score, protocol semantic impact score, detection risk score, and defense response score; The differential attack candidate attributes are determined based on the candidate impact proxy score, affected node type, affected business link, protocol semantic changes, and defense response results, and are used to characterize the degree to which candidate paths have a differentiated impact in the soft simulation environment.

[0021] In step 8) above, the formula for calculating the composite candidate influence proxy score is:

[0022] In the formula, express The candidate impact of time on agent score. This indicates that the observations were desynchronized. Indicates the business impact items, Indicates the physical effects. This represents the semantic components of the protocol, and the scoring results are cropped to the [0,1] interval.

[0023] In step 9) above, the path termination condition includes at least one of the following: the differential attack candidate influence condition reaches a preset threshold, the candidate path reaches the maximum number of steps, the legal action set is empty, the candidate path is blocked by the defense state, the detection risk reaches a preset threshold, or the strategy generation process meets the preset stop condition.

[0024] In step 9) above, the attack test cases and result reports are generated based on the following information: Differential attack risk assessment path, high-dimensional attack vector matrix fields corresponding to each abstract attack action in the candidate path, soft simulation feedback at each step, candidate impact proxy score, detection risk record, defense response record, path termination reason and capability boundary description. The attack test cases include preconditions, abstract attack steps, expected soft simulation feedback, and result judgment conditions. The result reports include candidate path reports, technical personnel experiment reports, non-technical personnel summaries, high-dimensional matrix extracts, and capability boundary descriptions.

[0025] A power system differential mode attack risk assessment path generation system based on high-dimensional attack vector matrix, and a multi-source attack modeling information management module, are used to acquire, organize and normalize the multi-source attack modeling information required for power system differential mode attack modeling, and generate a set of attack action primitives based on the normalization results. The high-dimensional attack vector matrix management module is used to construct or import high-dimensional attack vector matrices, and to perform field verification, encoding, and indexing on the high-dimensional attack vector matrices. The attack action space management module is used to generate an attack action space based on the high-dimensional attack vector matrix. 5) Combining the current system state, attack action space, and attack syntax constraints, perform a six-fold joint legality judgment on each candidate abstract attack action in the attack action space to generate a set of legal actions or an action mask vector. The six conditions are: access capability condition, target existence condition, target type matching condition, topology reachability condition, defense status condition, and repeated execution restriction condition. When a candidate abstract attack action meets all the conditions, the corresponding candidate abstract attack action is added to the set of legal actions, or the value of the corresponding index position of the candidate abstract attack action in the action mask vector is set to a first preset value to indicate that the candidate abstract attack action is in an optional state; when any condition is not met, the corresponding candidate abstract attack action is excluded from the set of legal actions, or the value of the corresponding index position of the candidate abstract attack action in the action mask vector is set to a second preset value to indicate that the candidate abstract attack action is in a blocked state, and the corresponding illegal reason is recorded.

[0026] By using the above action mask calculation method, the six legality conditions are jointly judged, so that candidate abstract attack actions that do not meet any of the legality conditions are deterministically blocked before entering the strategy generation model, thereby reducing invalid actions participating in path generation and improving the legality, interpretability and generation efficiency of risk assessment paths. The strategy generation module is used to select abstract attack actions based on the current system state and the set of legal actions or action mask vector; A soft simulation adaptation module is used to map the abstract attack action into an abstract simulation event in a power system soft simulation environment. The simulation feedback and impact assessment module is used to obtain soft simulation feedback and update the current system state, path history, candidate impact proxy scores, and defense response status. A differential attack candidate labeling module is used to determine differential attack candidate attributes based on the candidate influence proxy score and soft simulation feedback. The test case and result report generation module is used to output differential attack risk assessment path, attack test cases, and result reports when the path termination condition is met.

[0027] The strategy generation module includes a graph structure state encoding unit, a candidate action value evaluation unit, and a mask constraint action selection unit. The graph structure state encoding unit is used to encode the semantic features of nodes, topological adjacency relationships, and global state; The candidate action value evaluation unit is used to output the action value of candidate abstract attack actions; The mask constraint action selection unit is used to block illegal actions according to the action mask vector and select the next abstract attack action from the set of legal actions. The simulation feedback and impact assessment module is used to calculate or update at least one of the following based on the state changes returned by the soft simulation environment: observation difference score, business impact score, physical impact score, protocol semantic impact score, detection risk score, and defense response score. The updated results are then fed back to the strategy generation module and the test case and result report generation module.

[0028] This invention enables structured modeling, controlled verification, and closed-loop assessment of environment-dependent, differentiated security risks in power systems. It improves the systematic nature, interpretability, and reproducibility of generated security test cases for power systems, providing support for protection strategy evaluation, risk localization, and system security hardening. The method is only designed for abstract security testing and risk assessment in a software simulation environment and does not generate real attack payloads, real control messages, real device commands, or real vulnerability exploit scripts.

[0029] The present invention can also provide a computing device, a computer-readable storage medium, or a computer program product. The computing device includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, implements the above-described method. The computer-readable storage medium stores a computer program that, when executed by the processor, implements the above-described method. The computer program product, when executed by the computing device, implements the above-described method.

[0030] Any techniques not mentioned in this invention are based on existing technologies.

[0031] Compared with the prior art, the present invention has at least the following beneficial effects: First, this invention constructs a high-dimensional attack vector matrix to provide a unified, multi-dimensional, and structured model of the protocol semantics, business links, traffic characteristics, device topology, access permissions, hardware and software attributes, defense response characteristics, and simulation feedback information of power system attacks. This overcomes the limitations of traditional attack knowledge representation, which is fragmented, rule-based, and incalculable, achieving standardized, quantifiable, and simulable data representation of attack elements. This lays the data and model foundation for refined risk assessment of differential attacks on power systems.

[0032] Second, this invention integrates an attack syntax constraint mechanism to construct a set of legitimate attack actions. Based on attack phase logic, device access capabilities, power grid topology reachability, protocol preconditions, and attack target matching rules, it can automatically eliminate invalid, unreasonable, and unfeasible attack actions. This effectively improves the compliance and rationality of attack path generation, avoids interference from invalid scenarios, and further ensures the logic and credibility of differential attack risk assessment results.

[0033] Third, this invention introduces a strategy generation model combined with a soft simulation feedback closed-loop iterative mechanism. Relying on the simulation environment, the system operation status and attack impact agent score are updated in real time. It autonomously selects attack paths with actual threat and testing value, realizing the intelligent and automated generation of differential attack risk paths. This solves the drawback of traditional solutions that rely on manual experience to construct paths, and greatly improves the objectivity and comprehensiveness of risk assessment.

[0034] Fourth, the present invention sets up an attack impact proxy scoring and differential attack candidate attribute marking mechanism, which can accurately identify, screen and mark features of attack paths with environmental dependence and implementation condition constraints, realize differentiated and refined attack risk classification and judgment, and effectively improve the power system differential attack risk assessment system.

[0035] Fifth, based on high-dimensional attack vector matrices, optimal differential attack risk assessment paths, and simulation closed-loop feedback results, this invention automatically generates standardized attack test cases and quantitative risk assessment reports, which can provide reliable and complete data support for power system security performance testing, simulation verification, and iterative optimization of security protection strategies, significantly improving the standardization and engineering practicality of power system network security risk assessment. Attached Figure Description

[0036] Figure 1 A schematic diagram of the overall architecture of a power system differential attack risk assessment path generation system based on a high-dimensional attack vector matrix, provided for an embodiment of the present invention; Figure 2 A schematic diagram of the overall process for generating a path for assessing differential attacks on power systems based on a high-dimensional attack vector matrix, provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the high-dimensional attack vector matrix construction process provided in an embodiment of the present invention; Figure 4 A schematic diagram illustrating the process of generating attack syntax constraints and a set of legal actions provided in an embodiment of the present invention; Figure 5 A flowchart illustrating the interaction between the reinforcement learning strategy and the power system soft simulation environment to generate a differential attack risk assessment path, as provided in this embodiment of the invention. Figure 6 A schematic diagram illustrating the differential attack risk assessment path, attack test cases, and result report generation process provided in an embodiment of the present invention.

[0037] Figure 7 This is a schematic diagram of a differential attack risk assessment path generated in an embodiment of the present invention. Detailed Implementation

[0038] The following is combined Figures 1 to 7The technical solutions in the embodiments of the present invention are clearly and completely described. It should be understood that the following embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. The attack actions described herein are all abstract attack actions oriented towards power system software simulation environments, and do not include attack payloads, port scanning instructions, vulnerability exploitation processes, real packet construction, or real control instructions that can be directly applied to real power systems.

[0039] A method for generating a path for assessing differential mode attacks in power systems based on a high-dimensional attack vector matrix includes the following steps: 1) Obtain the current system state and the multi-source attack modeling information required for power system differential attack modeling. The current system state is used to characterize at least the topology state, access state, defense state and historical action state in the power system soft simulation environment. The multi-source attack modeling information includes at least attack investigation information, protocol semantic information, business link information, traffic spatiotemporal information, equipment topology information and soft simulation capability information. 2) Normalize the multi-source attack modeling information and generate a set of attack action primitives based on the normalized information, wherein the i-th attack action primitive is represented as:

[0040] In the formula, Indicates the first Each attack action primitive; Indicates the identity information of the action; Indicates information about the action phase; Represents information about the target object; This indicates the preconditions for the action; Represents protocol semantic information; Indicates information about the impact on business; Indicates the topological scope information; This indicates test mapping information; 3) Construct or import a high-dimensional attack vector matrix, which includes multiple rows of attack vectors corresponding to the attack action primitives. Each row of attack vectors corresponds to an abstract attack action and includes 12 sets of standardized fields: action identity field group, attack stage and access capability field group, protocol semantic field group, business behavior field group, traffic spatiotemporal field group, device topology field group, software and hardware implementation specific field group, differential mode impact field group, defense response field group, software simulation mapping field group, policy interface field group, and test case generation field group. The standardized attack vector corresponding to the i-th abstract attack action is obtained by fusing field group encodings:

[0041] In the formula, This represents the fused feature vector of the i-th abstract attack action; , , , , and These represent the encoding results of the attack phase, protocol semantics, business behavior, traffic spatiotemporal, device topology, and implementation environment fields, respectively. This represents the fusion weight matrix for the corresponding field group; 4) Generate an attack action space and attack syntax constraints based on the high-dimensional attack vector matrix, wherein the comprehensive feasibility evaluation value of the abstract attack action is calculated according to the constraint attributes of different field groups in the attack vector:

[0042] In the formula, This represents the comprehensive feasibility evaluation value of the i-th abstract attack action; This represents the constraint feature value of the k-th field group corresponding to the i-th attack action; This represents the weight coefficient of the corresponding field group; m represents the number of field groups involved in the calculation. 5) Combining the current system state, attack action space, and attack syntax constraints, perform a six-fold joint legality judgment on each candidate abstract attack action in the attack action space to generate a set of legal actions or an action mask vector. The six conditions are: access capability condition, target existence condition, target type matching condition, topology reachability condition, defense status condition, and repeated execution restriction condition. When a candidate abstract attack action meets all the conditions, the corresponding candidate abstract attack action is added to the set of legal actions, or the value of the corresponding index position of the candidate abstract attack action in the action mask vector is set to a first preset value to indicate that the candidate abstract attack action is in an optional state; when any condition is not met, the corresponding candidate abstract attack action is excluded from the set of legal actions, or the value of the corresponding index position of the candidate abstract attack action in the action mask vector is set to a second preset value to indicate that the candidate abstract attack action is in a blocked state, and the corresponding illegal reason is recorded. The mask value of the i-th candidate abstract attack action in the action mask vector at time t is determined by the following formula:

[0043] In the formula, Indicates the index of the candidate abstract attack action. Indicates the current decision-making moment. This represents the first preset value used to characterize the optional state. This represents a second preset value used to characterize the shielding state, and ; , , , , and They represent the first One candidate abstract attack action Whether the following conditions are met at any given time: access capability condition, target existence condition, target type matching condition, topology reachability condition, defense status condition, and repeated execution restriction condition; the value of each determination result is either a first preset value or a second preset value, where the first preset value is 1 and the second preset value is 0; When the result of the judgment of the above six conditions is 1, Add the corresponding candidate abstract attack action to the set of legal actions; when the result of any condition is 0, The corresponding candidate abstract attack actions are excluded from the set of legal actions, and the corresponding illegal reasons are recorded. Through the above joint calculation, only candidate abstract attack actions that simultaneously satisfy constraints such as access capability, target matching, topology reachability, and defense status are allowed to enter the subsequent strategy selection. 6) Input the current system state and the set of legal actions or action mask vector into the policy generation model. The policy generation model calculates the evaluation value of each candidate abstract attack action based on the current system state and action constraint information, and selects the abstract attack action with the highest evaluation value from the set of legal actions.

[0044] In the formula, This represents the abstract attack action selected at time t; This represents the i-th candidate abstract attack action; This represents the set of legal actions after filtering based on the action mask vector; This indicates that the policy generation model is in the current system state. Next action The calculated action evaluation value; 7) Map the abstract attack action to an abstract simulation event in the power system soft simulation environment, and obtain the soft simulation feedback; 8) Update the current system state, path history, composite candidate influence agent score, and differential attack candidate attributes based on the soft simulation feedback; The current system state is iteratively updated based on the action execution results and software simulation feedback:

[0045] In the formula, This represents the current system state at time t. This represents the abstract attack action executed at time t. This represents the feedback information from the soft simulation. This represents the state update function; X （t+1） Characterizes the system state at time t+1 after iterative update; 9) When the path termination condition is met, output the differential attack risk assessment path, security test cases, and result report.

[0046] In step 3) above, the field groups of the high-dimensional attack vector matrix include at least: action identity field group, attack stage and access capability field group, protocol semantic field group, business behavior field group, traffic spatiotemporal field group, device topology field group, hardware and software implementation specificity field group, differential impact field group, defense response field group, software simulation mapping field group, policy interface field group, and test case generation field group. Through these field groups, the system can uniformly represent dispersed attack knowledge, protocol semantics, business impact, topology relationships, capability conditions, defense feedback, and simulation feedback into a computable, constrainable, simulable, and policy-generating data structure.

[0047] In step 2), the attack action primitives include: action name, action identifier, attack stage, required access capabilities, target node type, target protocol type, preconditions, post-execution state changes, candidate impact description, detection risk description, defense response description, software simulation event mapping relationship, and test case generation fields; wherein, the attack action primitives are abstract attack action descriptions in the software simulation environment and do not contain real attack payloads, real device addresses, real control messages, or real vulnerability exploitation scripts.

[0048] In step 5), the current system state includes: the current set of accessible nodes, the current set of controllable nodes, the current set of reachable nodes, node semantic features, topological adjacency relationships, historical action sequences, defense status, detection status, service operation status, protocol semantic status, and candidate influence agent scoring status; wherein, the current system state is initialized according to the power system soft simulation environment at the beginning and is iteratively updated with the soft simulation feedback.

[0049] In step 6), the policy generation model is a reinforcement learning policy model. This model estimates action value or selects a policy based on a graph structure state representation, which includes the node semantic features, topological adjacency relationships, global state, and the action mask vector. The reinforcement learning policy model outputs the action value or action selection result of candidate abstract attack actions and selects the next abstract attack action under the constraint of the action mask vector. In step 6), the policy generation model does not select from all attack actions without constraints, but rather selects actions from the set of legal actions, thereby reducing invalid actions that do not meet the attack phase, access capability, topological reachability, or protocol preconditions.

[0050] In step 7), the abstract attack action is mapped to an abstract simulation event in the power system software simulation environment, including: determining the target object, action stage, target node type, protocol semantic category, business impact category, and state change category according to the software simulation mapping field corresponding to the abstract attack action; triggering the corresponding abstract simulation event in the power system software simulation environment; wherein, the abstract simulation event is used to change the topology state, access state, business state, protocol semantic state, defense state, or candidate impact proxy scoring state in the power system software simulation environment, without generating real attack packets, real control commands, or real vulnerability exploitation behavior.

[0051] In step 8), the composite candidate impact proxy score includes at least one of the following: observation difference score, business impact score, physical impact score, protocol semantic impact score, detection risk score, and defense response score; the differential attack candidate attribute is determined based on the candidate impact proxy score, affected node type, affected business link, protocol semantic change, and defense response result, and is used to characterize the degree of differential impact of the candidate path in the soft simulation environment.

[0052] In step 9), the path termination condition includes at least one of the following: the differential attack candidate influence condition reaches a preset threshold, the candidate path reaches the maximum number of steps, the legal action set is empty, the candidate path is blocked by the defense state, the detection risk reaches a preset threshold, or the strategy generation process meets the preset stop condition.

[0053] In step 9), the attack test cases and result reports are generated based on the following information: differential attack risk assessment path, high-dimensional attack vector matrix fields corresponding to each abstract attack action in the candidate path, soft simulation feedback at each step, candidate impact proxy score, detection risk record, defense response record, path termination reason, and capability boundary description; wherein, the attack test cases include preconditions, abstract attack steps, expected soft simulation feedback, and result judgment conditions, and the result reports include candidate path reports, technical personnel experiment reports, non-technical personnel summaries, high-dimensional matrix extracts, and capability boundary descriptions.

[0054] A power system differential mode attack risk assessment path generation system based on a high-dimensional attack vector matrix includes: The multi-source attack modeling information management module is used to acquire, organize, and normalize the multi-source attack modeling information required for differential mode attack modeling of power systems, and generate a set of attack action primitives based on the normalization results. The high-dimensional attack vector matrix management module is used to construct or import high-dimensional attack vector matrices, and to perform field verification, encoding, and indexing on the high-dimensional attack vector matrices. The attack action space management module is used to generate an attack action space based on the high-dimensional attack vector matrix. The attack syntax constraint module is used to perform a six-condition joint legality judgment on each candidate abstract attack action in the attack action space based on the current system state, attack action space, and attack syntax constraints, generating a set of legal actions or an action mask vector. The six conditions are access capability condition, target existence condition, target type matching condition, topology reachability condition, defense status condition, and repeated execution restriction condition. When a candidate abstract attack action meets all conditions, the corresponding candidate abstract attack action is added to the set of legal actions, or the value of the corresponding index position of the candidate abstract attack action in the action mask vector is set to a first preset value to indicate that the candidate abstract attack action is in an optional state. When any condition is not met, the corresponding candidate abstract attack action is excluded from the set of legal actions, or the value of the corresponding index position of the candidate abstract attack action in the action mask vector is set to a second preset value to indicate that the candidate abstract attack action is in a blocked state, and the corresponding illegal reason is recorded. The strategy generation module is used to select abstract attack actions based on the current system state and the set of legal actions or action mask vector; A soft simulation adaptation module is used to map the abstract attack action into an abstract simulation event in a power system soft simulation environment. The simulation feedback and impact assessment module is used to obtain soft simulation feedback and update the current system state, path history, candidate impact proxy scores, and defense response status. A differential attack candidate labeling module is used to determine differential attack candidate attributes based on the candidate influence proxy score and soft simulation feedback. The test case and result report generation module is used to output differential attack risk assessment path, attack test cases, and result reports when the path termination condition is met.

[0055] The strategy generation module includes a graph structure state encoding unit, a candidate action value evaluation unit, and a mask constraint action selection unit. The graph structure state encoding unit is used to encode the semantic features of nodes, topological adjacency relationships, and global state; The candidate action value evaluation unit is used to output the action value of candidate abstract attack actions; The mask constraint action selection unit is used to block illegal actions according to the action mask vector and select the next abstract attack action from the set of legal actions. The simulation feedback and impact assessment module is used to calculate or update at least one of the following based on the state changes returned by the soft simulation environment: observation difference score, business impact score, physical impact score, protocol semantic impact score, detection risk score, and defense response score. The updated results are then fed back to the strategy generation module and the test case and result report generation module.

[0056] In one specific embodiment, this invention uses a soft simulation environment for the coordinated operation of power system station control and process control as an example. This soft simulation environment includes 20 abstract nodes and 11 communication links. Node types include access_gateway, breaker, control_center, electrical_bus, feeder, gateway, HMI, ied, RTU, and server. The soft simulation environment includes control center nodes, SCADA server nodes, communication gateway nodes, process control nodes, measurement and control execution entity nodes, and circuit breaker object nodes. Protocol semantic abstraction includes industrial control protocols or business communication semantics such as IEC 60870-5-104, IEC 61850 GOOSE, Modbus, and DL / T 698.

[0057] like Figure 1 As shown in the figure, in this specific embodiment, the power system differential attack risk assessment path generation system based on high-dimensional attack vector matrix includes a multi-source attack modeling information management module, a high-dimensional attack vector matrix management module, an attack action space management module, an attack syntax constraint module, a strategy generation module, a soft simulation adaptation module, a simulation feedback and impact assessment module, a differential attack candidate labeling module, and a test case and result report generation module.

[0058] The multi-source attack modeling information management module is used to receive and organize attack research data, protocol semantic information, business link information, traffic spatiotemporal information, device topology information, software and hardware implementation specific information, and software simulation environment capability information.

[0059] The high-dimensional attack vector matrix management module is used to receive, store, verify, encode, and index the organized high-dimensional attack vector matrix, and provide the matrix to subsequent action space, syntax constraints, strategy generation, and test case generation processes.

[0060] In this specific implementation, the system's data flow includes a static knowledge flow, a dynamic decision flow, and a simulation feedback flow. The static knowledge flow refers to the multi-source attack modeling information management module first organizing attack research data, protocol semantic information, business link information, traffic spatiotemporal information, device topology information, hardware and software implementation specific information, and software simulation environment capability information to form attack action primitives or receive a set of manually organized attack action primitives. The high-dimensional attack vector matrix management module further performs field verification, encoding, and indexing on the attack action primitives to form a high-dimensional attack vector matrix that can be subsequently invoked. The attack action space management module and the attack syntax constraint module respectively read the action identity field, stage field, access capability field, target type field, protocol semantic field, topology constraint field, and simulation mapping field from the high-dimensional attack vector matrix to generate an abstract set of attack actions and corresponding attack syntax constraints.

[0061] Dynamic decision flow refers to the process where, at each moment of path generation, the system constructs the current system state based on the current node state, the current set of controllable nodes, the current set of reachable nodes, the set of acquired capabilities, historical action sequences, defense response status, and global state information. The attack syntax constraint module filters the attack action space for legality based on the current system state, generating a set of legal actions or an action mask for the current moment. The policy generation module receives the current system state and the set of legal actions, selects the next abstract attack action from the set of legal actions, and appends the selected abstract attack action to the differential attack risk assessment path.

[0062] The simulation feedback stream refers to the process by which the software simulation adaptation module maps the abstract attack actions selected by the strategy generation module to abstract simulation events in the software simulation environment. After executing these abstract simulation events, the software simulation environment returns information on node state changes, link state changes, business link impacts, protocol semantic anomalies, physical state changes, detection results, defense responses, and information on affected objects. The simulation feedback and impact assessment module updates the candidate impact proxy scores, current system state, and subsequent legal action set based on the feedback information. The differential attack candidate marking module updates the differential attack candidate attributes of the candidate paths based on hardware / software implementation specificity, protocol stack specificity, device model specificity, selective action marking, and impact scope marking. After the path terminates, the test case and result report generation module combines the differential attack risk assessment path, high-dimensional attack vector matrix fields, and software simulation feedback information to generate attack test cases and result reports. Thus, the system forms a closed-loop data flow from structured representation of attack knowledge, legal action generation, strategy selection, software simulation feedback to test case output.

[0063] like Figure 2 As shown, the above-mentioned static knowledge flow, dynamic decision flow, and simulation feedback flow can be further expanded into the following method flow: acquire multi-source information, form or import a high-dimensional attack vector matrix, generate an attack action space and attack syntax constraints, generate a set of legal actions based on the current state, select the next abstract attack action by the strategy generation module, map the abstract attack action to an abstract simulation event in the soft simulation environment, receive soft simulation feedback and update the candidate impact agent score, and finally generate a differential attack risk assessment path, attack test cases, and result report.

[0064] like Figure 3 As shown, in this specific implementation, the high-dimensional attack vector matrix is ​​denoted as M, and M consists of multiple attack vectors. Each attack vector corresponds to a discrete abstract attack action, and each field group corresponds to the attack action's identity, stage, protocol semantics, business semantics, traffic spatiotemporal, topology constraints, hardware / software specificity, differential candidate impact, defense response, software simulation mapping, policy interface, or test case generation attribute. In the current V1 implementation, the high-dimensional attack vector matrix includes 33 rows, with each row corresponding to a discrete abstract action in the action space. It can be represented as:

[0065]

[0066] In the formula, M represents the high-dimensional attack vector matrix; n represents the total number of abstract attack actions contained in the high-dimensional attack vector matrix, which is 33 in this embodiment; i represents the sequence number of the abstract attack action, and the value of i ranges from 1 to n. Let represent the attack vector corresponding to the i-th abstract attack action; T represents the transpose of the vector sequence composed of each attack vector, so that each attack vector forms a row in the high-dimensional attack vector matrix.

[0067] This represents the action identity field group for the i-th abstract attack action, used to describe the action identifier, action name, and action category; This represents the attack phase and access capability field group for the i-th abstract attack action, used to describe the phase to which the action belongs, the required access capabilities, and the prerequisite state conditions. The protocol semantic field group represents the i-th abstract attack action and is used to describe the associated protocol family, protocol object type, and protocol semantic category. The business behavior field group represents the i-th abstract attack action, which describes the business process, business chain, and business impact type associated with the action; The traffic spatiotemporal field group represents the i-th abstract attack action, which is used to describe the communication timing, traffic intensity, write operation ratio, burst degree and semantic offset characteristics related to the action; The device topology field group representing the i-th abstract attack action is used to describe the target node, target device type, device region, topology role, and topology reachability conditions; This represents the implementation environment dependency field group for the i-th abstract attack action, which describes the dependency of the action on a specific hardware architecture, software version, protocol stack implementation, device model, operating environment, or configuration method. This represents the differential candidate impact field group for the i-th abstract attack action, used to describe the differential attack candidate flag, candidate impact type, and expected candidate impact degree; The defense response field group represents the i-th abstract attack action, which describes the expected detection risk, defense response type, and action restriction status. The group of soft simulation mapping fields represents the i-th abstract attack action, which describes the mapping relationship between the abstract attack action and the soft simulation event; The policy interface field group represents the i-th abstract attack action, which describes the action index, action mask association conditions, and interface attributes required by the policy generation model. This represents the test case generation field group for the i-th abstract attack action, which describes the role of the action in the security test path, as well as the test steps, preconditions, and expected result generation attributes.

[0068] Table 1. Example of a high-dimensional attack vector matrix field schema.

[0069]

[0070] In this specific implementation, the differential attack candidate flag indicates that the corresponding attack action has subsequent differential verification value in the software simulation environment. The current V1 matrix contains a `differential_candidate_flag` field, and 24 out of 33 abstract actions are marked as differential attack candidate actions. This flag is determined primarily based on a comprehensive analysis of fields such as hardware / software implementation specificity, protocol stack specificity, device model specificity, selective action flag, scope of influence, and subsequent verification role.

[0071] Furthermore, taking the IEC104 information object redirection class abstract action as an example, its encoding in the high-dimensional attack vector matrix does not contain the real IEC104 message address or real control field, but only the action semantics, target type, access conditions, candidate differential mode attributes and simulation mapping information for the soft simulation environment.

[0072] Table 2 Examples of concrete abstract action coding

[0073] like Figure 4 As shown, in this specific implementation, the attack syntax constraint module generates a set of legal actions before each path generation step. The set of legal actions can be represented as an action mask vector with a length equal to the number of actions. In the current V1 implementation, the number of actions is 33, therefore the length of the action mask vector is 33. The system traverses each abstract action in the action space and determines whether the action is valid based on the current system state, access capabilities, target node, target type, topology reachability, defense isolation state, blocking edges, and repeated execution restrictions.

[0074] In one implementation, the action mask can be generated according to the following logic:

[0075] in, Indicates the first An abstract attack action, express The current system state at any given moment. Indicates as of The historical sequence of actions at any given moment. Indicates access capability conditions, used to determine the current system status. Does the access capability in the middle satisfy the action? The required access capabilities; This indicates the conditions under which the target exists, used to determine the action. Does the target node exist in the current software simulation topology? This indicates the target type matching condition, used to determine whether the target node type belongs to an action. The set of target types that are allowed to be used; This indicates topological reachability conditions, used to determine whether a target node is within the current controllable or reachable range; This indicates the defense status conditions, used to determine whether the target node has not been isolated, whether the relevant communication edges have not been blocked, and whether the current defense status does not prohibit the execution of this action; This indicates a condition for repeated execution, used to determine the action. Does it not violate the historical action sequence? The following are restrictions on repeated execution or mutual exclusion constraints. When both of the above conditions are met, A value of 1 indicates an action. This is a legal action at the current moment; if any condition is not met, then... A value of 0 indicates an action. It is currently blocked. In the current implementation, the phase sequence constraint defaults to a non-strict mode and does not necessarily block the action; in other implementations, this constraint can also be set as a strong blocking constraint.

[0076] Table 3 Examples of illegal reasons for attacking syntax constraints

[0077] like Figure 5 As shown, in this specific implementation, the policy generation module adopts the GraphDQN model based on graph structure state encoding as an implementation method. The input of this model includes a node feature matrix, an adjacency matrix, a global state vector, and an action mask. Among them, the node feature matrix has a shape of [20, 44], the adjacency matrix has a shape of [20, 20], the global state vector has a shape of

[16] , and the action mask has a shape of

[33] . The GraphDQN model outputs a 33-dimensional action Q-value vector, and each output component corresponds to a discrete abstract action in the action space.

[0078] In this specific implementation, the GraphDQN model includes two simple graph convolutional layers for extracting graph structure features from node features and topological adjacency relationships. The node embeddings after graph convolution are then subjected to mean pooling and max pooling to obtain graph-level representations, which are then concatenated with the global state vector and input into the head of a multilayer perceptron, outputting the Q-values ​​of each candidate action. Action masks are not used as direct inputs to the forward computation of GraphDQN, but are used in the policy selection stage to shield illegal actions, preventing them from being selected as candidate actions for the next step.

[0079] The global state vector can include step count ratio, access capability, detection score, candidate impact agent score, stealth level, illegal action ratio, observation desynchronization score, business impact score, physical impact score, protocol semantic score, defense response severity, alarm node ratio, affected node ratio, controlled node ratio, reachable node ratio, and current stage identifier. These states collectively describe the current software simulation environment, attack history, and defense response status.

[0080] In this specific implementation, after each abstract simulation event is executed, the soft simulation environment feeds back the aggregation module's aggregation of node state changes, business state changes, protocol semantic anomalies, physical state changes, detection results, and defense responses. The candidate impact proxy score is obtained by weighting the observation desynchronization sub-item, business impact sub-item, physical impact sub-item, and protocol semantic sub-item.

[0081]

[0082] express The candidate impact of time on agent score. This indicates that the observations were desynchronized. Indicates the business impact items, Indicates the physical effects. This represents the semantic components of the protocol.

[0083] The above The score is cropped to the [0,1] interval. It should be noted that in this implementation, this score represents the candidate influence proxy score in the soft simulation environment and is not directly equivalent to the differential attack verification result.

[0084] During the reinforcement learning path generation process, the system also calculates the single-step reward based on factors such as access capability gain, candidate impact increment, observation desynchronization sub-item, business impact sub-item, physical impact sub-item, protocol semantic sub-item, detection risk, illegal action penalty, defense response penalty, stage advancement reward, success reward, and failure penalty. One implementation is as follows:

[0085] in, Indicates the increment of access capacity. This indicates the incremental impact of effective candidates. This indicates that the observations were desynchronized. Indicates the business impact items, Indicates the physical effects. Represents the semantic components of the protocol. This indicates an increase in the risk of detection. This indicates penalties related to illegal actions, defense responses, blocking, detection failures, or timeouts. This represents the reward when the path meets the preset success conditions. This reward function is only one implementation in a specific embodiment; in other implementations, the weights or sub-items can be adjusted according to the test objectives.

[0086] like Figure 6 As shown, in this specific implementation, the system terminates the current path generation process when the path reaches the preset candidate influence conditions, reaches the maximum number of path steps, is blocked by the defense response, or there is no legitimate action in the current state. The test case and result report generation module structurally combines the differential attack risk assessment path, static fields in the high-dimensional attack vector matrix, soft simulation feedback at each step, candidate influence proxy scores, defense response records, and capability boundary descriptions to form attack test cases and result reports.

[0087] The attack test cases may include test case identifiers, candidate path steps, abstract attack actions, target node types, preconditions, expected feedback, actual simulation feedback, candidate impact proxy scores, and capability boundary descriptions. The results report may include an experimental report for technical personnel, a summary for non-technical personnel, a candidate path report, a high-dimensional attack vector matrix extract, and patent specification materials. Both the test cases and reports are used for security testing, software simulation verification, or subsequent verification input and do not represent real system attack scripts.

[0088] like Figure 7 As shown in this specific implementation, under a soft simulation scenario with a business_load_jitter perturbation strength of 0.15, the system generates a differential attack risk assessment path consisting of eight abstract actions using the DQN strategy. This path sequentially includes maintenance_plane_access, asset_enumeration, lateral_to_process_layer, setpoint_tamper, iec104_cot_spoof, iec104_value_tamper, iec104_ioa_redirect, and breaker_trip_command. The generation process of this path is shown in the table below.

[0089] Table 4. Example of the evolution process of differential attack risk assessment path

[0090] In the above path, each action was filtered by an action mask before being selected. Historical execution records show that no illegal action flag was triggered at any step of this 8-step path, and no detection or defense blocking was triggered. After the 8th step was executed, the candidate impact proxy score reached the preset success condition, the system returned `success_differential_impact_reached`, and path generation terminated. The final candidate impact proxy score for this path was 0.758305, the final detection score was 0.0, and the total reward was 15.785222.

[0091] It should be noted that, Figure 7 The paths shown in Table 4 are merely one specific implementation of the present invention, used to illustrate how the system generates differential attack risk assessment paths based on a high-dimensional attack vector matrix, attack syntax constraints, a policy generation model, and soft simulation feedback. The present invention does not limit candidate paths to including the aforementioned actions, nor does it limit them to having the same number of actions, protocol type, topology scale, model structure, or scoring weights. In other implementations, other differential attack risk assessment paths can be generated based on different power business scenarios, different soft simulation topologies, different protocol semantics, different policy models, or different test objectives.

[0092] In summary, this specific implementation method constructs a power system soft simulation environment with 20 nodes and 11 connections to integrate and import multi-source attack knowledge and construct a high-dimensional attack vector matrix. It generates a set of legal attack actions based on standardized attack syntax constraints and optimizes abstract attack actions within the legal action range using a strategy generation model. Simultaneously, by leveraging the real-time feedback of attack candidate impacts from the soft simulation environment, it iteratively optimizes the attack path in a closed loop, ultimately generating a precise power system differential attack risk assessment path, standardized attack test cases, and quantitative assessment results. This achieves a precise and comprehensive quantitative assessment of the power system differential attack risk. It should be noted that the above specific implementation method is only used to clearly illustrate the technical solution of this invention and is not intended to limit the scope of protection of this invention. Those skilled in the art can make reasonable improvements and optimizations based on the core technical ideas of this invention, all of which fall within the scope of protection of this invention.

Claims

1. A method for generating a path for assessing differential mode attack risk in a power system based on a high-dimensional attack vector matrix, characterized in that, Includes the following steps: 1) Obtain the current system state and the multi-source attack modeling information required for power system differential attack modeling. The current system state is used to characterize at least the topology state, access state, defense state and historical action state in the power system soft simulation environment. The multi-source attack modeling information includes at least attack investigation information, protocol semantic information, business link information, traffic spatiotemporal information, equipment topology information and soft simulation capability information. 2) Normalize the multi-source attack modeling information and generate a set of attack action primitives based on the normalized information, wherein the i-th attack action primitive is represented as: In the formula, Indicates the first Each attack action primitive; Indicates the identity information of the action; Indicates information about the action phase; Represents information about the target object; This indicates the preconditions for the action. Represents protocol semantic information; Indicates information about the impact on business; Indicates the topological scope information; This indicates test mapping information; 3) Construct or import a high-dimensional attack vector matrix, which includes multiple rows of attack vectors corresponding to the attack action primitives. Each row of attack vectors corresponds to an abstract attack action and includes 12 sets of standardized fields: action identity field group, attack stage and access capability field group, protocol semantic field group, business behavior field group, traffic spatiotemporal field group, device topology field group, software and hardware implementation specific field group, differential mode impact field group, defense response field group, software simulation mapping field group, policy interface field group, and test case generation field group. The standardized attack vector corresponding to the i-th abstract attack action is obtained by fusing field group encodings: In the formula, This represents the fused feature vector of the i-th abstract attack action; , , , , and These represent the encoding results of the attack phase, protocol semantics, business behavior, traffic spatiotemporal, device topology, and implementation environment fields, respectively. This represents the fusion weight matrix for the corresponding field group; 4) Generate an attack action space and attack syntax constraints based on the high-dimensional attack vector matrix, wherein the comprehensive feasibility evaluation value of the abstract attack action is calculated according to the constraint attributes of different field groups in the attack vector: In the formula, This represents the comprehensive feasibility evaluation value of the i-th abstract attack action; This represents the constraint feature value of the k-th field group corresponding to the i-th attack action; This represents the weight coefficient of the corresponding field group; m represents the number of field groups involved in the calculation. 5) Combining the current system state, attack action space, and attack syntax constraints, perform a joint legality determination on each candidate abstract attack action in the attack action space, considering access capability conditions, target existence conditions, target type matching conditions, topology reachability conditions, defense state conditions, and repeated execution restriction conditions, to generate a set of legal actions or an action mask vector; the mask value of the i-th candidate abstract attack action in the action mask vector at time t is determined according to the following formula: In the formula, Indicates the index of the candidate abstract attack action. Indicates the current decision-making moment. This represents the first preset value used to characterize the optional state. This represents a second preset value used to characterize the shielding state, and ; , , , , and They represent the first One candidate abstract attack action in Whether the following conditions are met at any given time: access capability condition, target existence condition, target type matching condition, topology reachability condition, defense status condition, and repeated execution restriction condition; the value of each determination result is either a first preset value or a second preset value, where the first preset value is 1 and the second preset value is 0; When the result of the judgment of the above six conditions is 1, Add the corresponding candidate abstract attack action to the set of legal actions; when the result of any condition is 0, The corresponding candidate abstract attack actions are excluded from the set of legal actions, and the corresponding illegal reasons are recorded. Through the above joint calculation, only candidate abstract attack actions that simultaneously satisfy constraints such as access capability, target matching, topology reachability, and defense status are allowed to enter the subsequent strategy selection. 6) Input the current system state and the set of legal actions or action mask vector into the policy generation model. The policy generation model calculates the evaluation value of each candidate abstract attack action based on the current system state and action constraint information, and selects the abstract attack action with the highest evaluation value from the set of legal actions. In the formula, This represents the abstract attack action selected at time t; This represents the i-th candidate abstract attack action; This represents the set of legal actions after filtering based on the action mask vector; This indicates that the policy generation model is in the current system state. Next action The calculated action evaluation value; 7) Map the abstract attack action to an abstract simulation event in the power system soft simulation environment, and obtain the soft simulation feedback; 8) Update the current system state, path history, composite candidate influence agent score, and differential attack candidate attributes based on the soft simulation feedback; The current system state is iteratively updated based on the action execution results and software simulation feedback: In the formula, This represents the current system state at time t. This represents the abstract attack action executed at time t. This represents the feedback information from the soft simulation. This represents the state update function; X （t+1） Characterizes the system state at time t+1 after iterative update; 9) When the path termination condition is met, output the differential attack risk assessment path, attack test cases, and result report.

2. The method according to claim 1, characterized in that, In step 2), the attack action primitives include: action name, action identifier, attack stage, required access capability, target node type, target protocol type, preconditions, post-execution state change, candidate impact description, detection risk description, defense response description, soft simulation event mapping relationship, and test case generation fields. The attack action primitives are abstract attack action descriptions in a software simulation environment, and do not contain real attack payloads, real device addresses, real control messages, or real vulnerability exploitation scripts.

3. The method according to claim 1 or 2, characterized in that, In step 5), the current system state includes: the current set of accessible nodes, the current set of controllable nodes, the current set of reachable nodes, node semantic features, topological adjacency relationships, historical action sequences, defense status, detection status, business operation status, protocol semantic status, and candidate influence proxy scoring status. The current system state is initialized based on the power system soft simulation environment at the beginning and is iteratively updated with the soft simulation feedback.

4. The method according to claim 1 or 2, characterized in that, In step 6), the policy generation model is a reinforcement learning policy model. The reinforcement learning policy model performs action value estimation or policy selection based on graph structure state representation. The graph structure state representation includes the node semantic features, the topological adjacency relationship, the global state, and the action mask vector. The reinforcement learning policy model outputs the action value or action selection result of the candidate abstract attack actions, and selects the next abstract attack action under the constraint of the action mask vector.

5. The method according to claim 1 or 2, characterized in that, In step 7), the abstract attack action is mapped to an abstract simulation event in the power system soft simulation environment, including: Based on the software simulation mapping field corresponding to the abstract attack action, determine its target object, action stage, target node type, protocol semantic category, business impact category, and state change category; Trigger the corresponding abstract simulation event in the power system soft simulation environment; The abstract simulation event is used to change the topology state, access state, service state, protocol semantic state, defense state, or candidate influence agent scoring state in the power system software simulation environment, without generating real attack messages, real control commands, or real vulnerability exploitation behaviors.

6. The method according to claim 1 or 2, characterized in that, In step 8), the composite candidate impact proxy score includes at least one of the following: observation difference score, business impact score, physical impact score, protocol semantic impact score, detection risk score, and defense response score; The differential attack candidate attributes are determined based on the candidate impact proxy score, affected node type, affected business link, protocol semantic changes, and defense response results, and are used to characterize the degree to which candidate paths have a differentiated impact in the soft simulation environment.

7. The method according to claim 6, characterized in that, In step 8), the formula for calculating the composite candidate influence proxy score is: In the formula, express The candidate impact of time on agent score. This indicates that the observations were desynchronized. Indicates the business impact items, Indicates the physical effects. The numbers 0.32, 0.24, 0.28, and 0.16 represent the semantic components of the protocol, respectively, and the weighting coefficients for each component are 0.32, 0.24, 0.28, and 0.

16. and The calculation results are restricted to the [0,1] interval. Through the above weighted aggregation and interval restriction, the soft simulation feedback of different dimensions is converted into a scoring result with a unified dimension, which is used to compare the influence of candidate paths and determine the path termination condition.

8. The method according to claim 1 or 2, characterized in that, In step 9), the path termination condition includes at least one of the following: the differential attack candidate influence condition reaches a preset threshold, the candidate path reaches the maximum number of steps, the legal action set is empty, the candidate path is blocked by the defense state, the detection risk reaches a preset threshold, or the strategy generation process meets the preset stop condition. In step 9), the attack test cases and result reports are generated based on the following information: Differential attack risk assessment path, high-dimensional attack vector matrix fields corresponding to each abstract attack action in the candidate path, soft simulation feedback at each step, candidate impact proxy score, detection risk record, defense response record, path termination reason and capability boundary description. The attack test cases include preconditions, abstract attack steps, expected soft simulation feedback, and result judgment conditions. The result reports include candidate path reports, technical personnel experiment reports, non-technical personnel summaries, high-dimensional matrix extracts, and capability boundary descriptions.

9. A power system differential mode attack risk assessment path generation system based on a high-dimensional attack vector matrix, characterized in that, include: The multi-source attack modeling information management module is used to acquire, organize, and normalize the multi-source attack modeling information required for differential mode attack modeling of power systems, and generate a set of attack action primitives based on the normalization results. The high-dimensional attack vector matrix management module is used to construct or import high-dimensional attack vector matrices, and to perform field verification, encoding, and indexing on the high-dimensional attack vector matrices. The attack action space management module is used to generate an attack action space based on the high-dimensional attack vector matrix. The attack syntax constraint module is used to perform a six-condition joint legality determination on each candidate abstract attack action in the attack action space based on the current system state, attack action space, and attack syntax constraints, generating a set of legal actions or an action mask vector. The six conditions are access capability condition, target existence condition, target type matching condition, topology reachability condition, defense status condition, and repeated execution restriction condition. When a candidate abstract attack action satisfies all conditions, the corresponding candidate abstract attack action is added to the set of legal actions, or the value of the corresponding index position of the candidate abstract attack action in the action mask vector is set to a first preset value to indicate that the candidate abstract attack action is in an optional state. If any condition is not met, the corresponding candidate abstract attack action is excluded from the set of legal actions, or the value of the corresponding index position of the candidate abstract attack action in the action mask vector is set to a second preset value to indicate that the candidate abstract attack action is in a masked state, and the corresponding illegal reason is recorded. The strategy generation module is used to select abstract attack actions based on the current system state and the set of legal actions or action mask vector; A soft simulation adaptation module is used to map the abstract attack action into an abstract simulation event in a power system soft simulation environment. The simulation feedback and impact assessment module is used to obtain soft simulation feedback and update the current system state, path history, candidate impact proxy scores, and defense response status. A differential attack candidate labeling module is used to determine differential attack candidate attributes based on the candidate influence proxy score and soft simulation feedback. The test case and result report generation module is used to output differential attack risk assessment path, attack test cases, and result reports when the path termination condition is met.

10. The system according to claim 9, characterized in that, The strategy generation module includes a graph structure state encoding unit, a candidate action value evaluation unit, and a mask constraint action selection unit. The graph structure state encoding unit is used to encode the semantic features of nodes, topological adjacency relationships, and global state; The candidate action value evaluation unit is used to output the action value of candidate abstract attack actions; The mask constraint action selection unit is used to block illegal actions according to the action mask vector and select the next abstract attack action from the set of legal actions. The simulation feedback and impact assessment module is used to calculate or update at least one of the following based on the state changes returned by the soft simulation environment: observation difference score, business impact score, physical impact score, protocol semantic impact score, detection risk score, and defense response score. The updated results are then fed back to the strategy generation module and the test case and result report generation module.

Images