An AI security risk assessment method based on industrial cloud computing

Abstract

The application discloses an AI security risk assessment method based on industrial cloud computing, and relates to the technical field of risk assessment, comprising the following steps: identifying AI model instance decision behaviors actually triggered by AI model instances, and associating the AI model instance decision behaviors with corresponding industrial execution objects to generate decision execution association results; judging influence extension paths of the AI model instance decision behaviors according to the decision execution association results, determining the action ranges of the AI model instance decision behaviors, generating industrial influence level results, judging amplification states of the AI model instance decision behaviors based on the industrial influence level results and the current operation situation of an industrial cloud computing platform, and generating decision amplification state results; and the AI decision amplification effect is effectively determined by combining the industrial influence level results and the operation situation of the industrial cloud computing platform, so that the security risk is prevented from being gradually amplified in the operation process.

Description

Technical Field

[0001] This invention relates to the field of risk assessment technology, and in particular to an AI security risk assessment method based on industrial cloud computing. Background Technology

[0002] With the continuous development of the Industrial Internet, industrial big data, and cloud computing technologies, industry is gradually evolving from traditional closed automated control architectures towards open, platform-based, and intelligent directions. Industrial cloud computing provides a unified operating environment for equipment management, process optimization, and business scheduling in industrial production processes through centralized computing resources, elastic resource scheduling capabilities, and cross-platform data integration capabilities. Based on this, artificial intelligence technology is increasingly being introduced into industrial cloud computing platforms to achieve functions such as equipment status prediction, production plan optimization, anomaly detection, and autonomous decision-making control. In current industrial scenarios, AI models are typically deployed in the industrial cloud computing environment as model instances, and through continuous operation, they output decision results to industrial business processes, thereby directly or indirectly participating in the control process of industrial execution objects. As AI model instances are deployed on a large scale in industrial cloud environments, the scope of their decision-making behavior continues to expand, and the decision-making execution chain continues to extend. The industrial cloud computing platform has evolved from a simple computing power support role into a key hub carrying AI decision-making behavior, industrial business processes, and underlying resource scheduling.

[0003] However, existing methods still have room for improvement. First, current technologies generally overlook the amplification effect of AI decision-making in industrial cloud computing environments. This means that AI decisions may lead to continuous superposition, repeated triggering, or loop amplification problems under the action of mechanisms such as resource scheduling, task queuing, and execution feedback, resulting in the gradual amplification of security risks during operation. Second, existing technologies mainly focus on model algorithm security, data access control, or platform resource isolation, often evaluating AI decisions as isolated events. They lack the characterization of the relationship between the decision-making behavior of AI model instances and specific industrial execution objects, making it difficult to accurately reflect the actual execution path and scope of AI decisions in industrial business processes. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an AI security risk assessment method based on industrial cloud computing to address the problem of ignoring the amplification effect of AI decision-making behavior in industrial cloud computing environments.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides an AI security risk assessment method based on industrial cloud computing, comprising:

[0008] In an industrial cloud computing environment, deployed AI model instances are parsed to identify the actual AI model instance decision-making behaviors triggered by the AI ​​model instances, and associated with the corresponding industrial execution objects to generate decision execution association results.

[0009] Based on the results of decision execution correlation, determine the impact extension path of AI model instance decision behavior, determine the scope of AI model instance decision behavior, and generate industrial impact hierarchy results;

[0010] Based on the results of the industrial impact hierarchy and the current operating status of the industrial cloud computing platform, the amplification status of the decision-making behavior of the AI ​​model instance is determined, and the decision amplification status result is generated.

[0011] Based on the results of the decision amplification state, the results of the industrial impact hierarchy, and the controllability of the current cloud environment, the degree of security risk of the decision-making behavior of AI model instances is classified and determined, and a risk level result is generated.

[0012] Based on the risk level results, corresponding risk response management strategies are generated to control the security risks of AI model instance decision-making behavior in the industrial cloud environment.

[0013] As a preferred embodiment of the AI ​​security risk assessment method based on industrial cloud computing described in this invention, the generation of decision execution correlation results includes:

[0014] The runtime status information and decision output records of deployed AI model instances are analyzed to extract the decision output content generated by the AI ​​model instances during operation, and the decision behavior and decision triggering conditions of the AI ​​model instances are determined based on the decision output content.

[0015] The decision-making behavior of AI model instances is associated with predefined industrial execution objects in industrial business processes to generate decision execution association results between AI model instance decision-making behavior and industrial execution objects.

[0016] As a preferred embodiment of the AI ​​security risk assessment method based on industrial cloud computing described in this invention, wherein: determining the scope of the decision-making behavior of the AI ​​model instance includes:

[0017] Analyze the results of decision execution to extract the industrial execution objects and industrial business process locations corresponding to the decision-making behavior of AI model instances;

[0018] Based on the location of industrial business processes, the transmission order and correlation of AI model instance decision-making behavior in industrial business processes are analyzed to determine the impact extension path of AI model instance decision-making behavior in industrial business processes.

[0019] Based on the location of industrial business processes in the influence expansion path, the covered business process levels are identified layer by layer, and the identified business process levels are used as the scope of the AI ​​model instance decision-making behavior.

[0020] As a preferred embodiment of the AI ​​security risk assessment method based on industrial cloud computing described in this invention, the generation of industrial impact level results includes: hierarchically dividing the scope of the decision-making behavior of AI model instances to form corresponding industrial impact level identifiers, recording the industrial impact level identifiers in correspondence with the decision-making behavior of AI model instances, and generating industrial impact level results.

[0021] As a preferred embodiment of the AI ​​security risk assessment method based on industrial cloud computing described in this invention, wherein: determining the amplified state of the decision-making behavior of the AI ​​model instance includes:

[0022] Obtain the current operational status information of the industrial cloud computing platform, analyze the operational status information, and extract the computing power scheduling rhythm, resource isolation boundary status, execution buffer queue depth, and task queuing sequence features directly triggered by the decision-making behavior of AI model instances.

[0023] The computing power scheduling rhythm, resource isolation boundary status, execution buffer queue depth, and task queuing sequence characteristics are organized in chronological order to form a set of operational status characteristics;

[0024] Based on the set of operational status features, the feedback change characteristics of AI model instance decision-making behavior on the operational status of industrial cloud computing platform are identified, forming a set of decision-making status feedback features.

[0025] The carrying capacity status result is generated by combining the set of operational status characteristics, the set of decision-making status feedback characteristics, and the results of industrial impact levels.

[0026] As a preferred embodiment of the AI ​​security risk assessment method based on industrial cloud computing described in this invention, the generation of decision amplification state results includes:

[0027] The results of the carrying capacity status are analyzed to extract the change information of the carrying capacity status results over a continuous time period and generate a carrying capacity change sequence.

[0028] Based on the carrying capacity change sequence, identify the states of carrying capacity decline, increased carrying capacity fluctuation and delayed carrying capacity recovery within the corresponding industrial impact level;

[0029] By using the sequence of changes in carrying capacity and the set of feedback features of decision-making status, we can identify the execution state of repeated triggering, continuous superposition, and loop amplification of AI model instance decision-making behavior in the industrial cloud computing environment.

[0030] Based on the states of decreased carrying capacity, increased carrying capacity fluctuation, delayed carrying capacity recovery, and the execution state of loop amplification, the amplification state corresponding to the decision-making behavior of AI model instances is identified, and decision amplification state results are generated.

[0031] As a preferred embodiment of the AI ​​security risk assessment method based on industrial cloud computing described in this invention, the step of classifying and determining the security risk level of AI model instance decision-making behavior includes:

[0032] By analyzing the execution status of AI model instance decision-making behavior in the industrial cloud computing environment through the decision amplification status results, the amplification status level corresponding to the AI ​​model instance decision-making behavior is obtained, and the influence level range of AI model instance decision-making behavior in industrial business processes is determined based on the industrial influence level results.

[0033] Based on the results of the carrying capacity status and the decision amplification status, the controllability of the industrial cloud computing environment to the decision-making behavior of AI model instances under the current operating conditions is quantitatively evaluated, and the intervention capability status of the industrial cloud computing environment to the decision-making behavior of AI model instances is extracted.

[0034] The risk assessment criteria are formed by combining the amplified status level, the scope of impact, and the intervention capability status.

[0035] As a preferred embodiment of the AI ​​security risk assessment method based on industrial cloud computing described in this invention, the risk level result is obtained by classifying and identifying the degree of security risk of the decision-making behavior of AI model instances based on risk judgment criteria.

[0036] As a preferred embodiment of the AI ​​security risk assessment method based on industrial cloud computing described in this invention, the step of extracting the intervention capability status of the industrial cloud computing environment on the decision-making behavior of AI model instances specifically includes:

[0037] By analyzing the carrying capacity status results, we can obtain the resource scheduling margin status and execution buffer availability status of the industrial cloud computing environment within the corresponding industrial impact level. At the same time, by analyzing the decision amplification status results, we can obtain the execution continuity status and execution interruptibility status of the AI ​​model instance decision behavior in the industrial cloud computing environment.

[0038] According to the preset state correspondence, the resource scheduling margin state, execution buffer available state, execution continuity state, and execution interruptibility state are combined and mapped to the intervention capability state.

[0039] As a preferred embodiment of the AI ​​security risk assessment method based on industrial cloud computing described in this invention, the risk response management strategy is generated by limiting the execution frequency, execution order, and execution triggering conditions of AI model instance decision-making behavior in the industrial cloud computing environment through risk level results and intervention capability status.

[0040] The beneficial effects of this invention are as follows: By analyzing deployed AI model instances, the decision-making behavior of AI model instances is associated with specific industrial execution objects, clarifying the actual execution path and scope of AI decisions in industrial business processes. This solves the problem in existing technologies of evaluating AI decisions as isolated events and making it difficult to locate business impacts. At the same time, by combining the results of industrial impact hierarchy with the operational status of industrial cloud computing platforms, the repeated triggering, continuous superposition, and loop amplification states formed by the decision-making behavior of AI model instances under the influence of resource scheduling, task queuing, and execution feedback are identified. This enables effective determination of the amplification effect of AI decisions and avoids the gradual amplification of security risks during operation. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart of an AI security risk assessment method based on industrial cloud computing.

[0043] Figure 2 This diagram illustrates the relationship between the decision-making behavior of an AI model instance and the industrial execution object.

[0044] Figure 3 This is a schematic diagram of the decision amplification state determination process.

[0045] Figure 4 This is a diagram illustrating risk level assessment and risk response management. Detailed Implementation

[0046] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0047] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0048] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0049] Reference Figures 1-4 This is one embodiment of the present invention, which provides an AI security risk assessment method based on industrial cloud computing, including the following steps:

[0050] S1: In an industrial cloud computing environment, the deployed AI model instances are analyzed to identify the decision-making behaviors actually triggered by the AI ​​model instances, and associated with the corresponding industrial execution objects to generate decision execution association results.

[0051] S1.1: In an industrial cloud computing environment, the industrial cloud computing platform monitors the operational status of deployed AI model instances and obtains operational status information and decision output records corresponding to the AI ​​model instances from the industrial cloud computing platform. The operational status information includes the deployment identifier, operation timestamp, and execution context identifier of the AI ​​model instance in the industrial cloud computing environment. The decision output records include control instructions, scheduling instructions, or policy instructions output by the AI ​​model instance to the industrial business process during operation.

[0052] It should also be noted that AI model instances are machine learning or deep learning models trained and deployed in industrial cloud computing platforms to perform specific decision-making tasks in industrial business processes.

[0053] The industrial cloud computing platform parses and processes operational status information and decision output records, extracting the output content related to industrial business processes from the decision output records as decision output content. This related output content refers to control instructions, scheduling instructions, or strategy instructions generated by the AI ​​model instance based on the current industrial business status, process requirements, and task objectives. These instructions influence industrial execution objects or guide the operation of industrial processes. The decision output content corresponds to the specific decision results made by the AI ​​model instance in the industrial cloud computing environment for the industrial business process. Based on the instruction type identifier and industrial business process location identifier contained in the decision output content, the industrial cloud computing platform determines the actual AI model instance decision-making behavior triggered during operation, and simultaneously determines the decision triggering conditions. Decision triggering conditions include the industrial business status, industrial business events, or industrial business parameter changes corresponding to the decision output content.

[0054] S1.2: The industrial cloud computing platform reads industrial execution object information from the industrial business process configuration. The industrial execution object information is used to represent the execution entities in the industrial business process that can be directly affected by the decision-making behavior of AI model instances. The industrial execution objects include industrial equipment control objects, industrial process control nodes, and industrial business scheduling nodes.

[0055] The industrial cloud computing platform associates AI model instance decision-making behaviors with predefined industrial execution objects within industrial business processes. The association process involves the platform reading the industrial business process location identifier carried in the decision output and comparing it with the industrial business process location identifiers already bound in the industrial business process configuration. If the comparison results match, the industrial execution object bound to the industrial business process location identifier is identified as the corresponding industrial execution object for the AI ​​model instance decision-making behavior. If the industrial business process location identifier in the decision output cannot be matched with a corresponding industrial execution object in the industrial business process configuration, the platform marks the AI ​​model instance decision-making behavior as unassociated and records the corresponding industrial business process location identifier. Through this association process, a one-to-one correspondence is established between AI model instance decision-making behaviors and industrial execution objects.

[0056] To further explain, industrial execution objects are predefined when configuring industrial business processes on the industrial cloud computing platform. During the configuration process, the platform assigns corresponding industrial execution objects to the executable nodes within the industrial business process and sets industrial execution object identifiers, industrial business process location identifiers, and execution capability attributes for each object. These attributes represent the object's position within the process and the types of AI model instance decision-making behaviors it can receive and execute. The platform then binds these industrial execution objects to their corresponding positions within the industrial business process, thus establishing an industrial execution object configuration relationship.

[0057] After completing the corresponding association, a decision execution association result is obtained, which represents the correspondence between the decision-making behavior of AI model instances and industrial execution objects. The decision execution association result is used to describe the industrial execution objects that the decision-making behavior of AI model instances actually affects in the industrial cloud computing environment.

[0058] Preferably, unlike existing methods that analyze AI decisions merely as model outputs, this invention introduces an intermediate layer of industrial execution objects within the industrial cloud computing environment. This allows AI model instance decision-making behavior to establish a clear correspondence with specific execution entities in industrial business processes, effectively mapping AI decisions from the abstract model layer to the industrial execution layer. In this way, the actual objects and locations of AI model instance decision-making behavior within the industrial cloud computing environment become locatable and traceable, providing a stable and clear foundation for subsequent assessments of the impact, amplification, and security risk levels of AI model instance decision-making behavior. Compared to existing AI security analysis methods that lack constraints on decision execution objects, this invention avoids the problem of difficulty in pinpointing and locating AI decision risks, thus improving the engineering applicability of AI security risk assessment in the industrial cloud computing environment.

[0059] S2: Determine the impact extension path of AI decision-making based on the results of decision execution association, determine the scope of the AI ​​model instance decision behavior, and generate industrial impact hierarchy results.

[0060] S2.1: The decision execution association results include the AI ​​model instance decision behavior identifier, the industrial execution object identifier, and the industrial business process location identifier. The industrial cloud computing platform reads the industrial execution object corresponding to the AI ​​model instance decision behavior from the decision execution association results, and at the same time reads the industrial business process location of the industrial execution object in the industrial business process.

[0061] The industrial cloud computing platform reads the connection relationships between various industrial execution objects in the industrial business process based on the location of the industrial business process and utilizes the pre-configured process structure within the industrial business process. This pre-configured process structure is defined during the industrial business process modeling phase to clarify the connection relationships and execution dependencies between industrial execution objects. Based on the location of the industrial business process, the platform locates the industrial execution objects directly affected by the AI ​​model instance's decision-making behavior within the industrial business process structure, and then, following the connection relationships within the industrial business process structure, identifies subsequent industrial execution objects that have execution dependencies on the original industrial execution objects.

[0062] In determining subsequent industrial execution objects, the industrial cloud computing platform identifies the transmission relationships between industrial execution objects based on the process structure dependencies pre-defined during the configuration of industrial business processes. The process structure dependencies are used to limit the sequential execution constraints between industrial execution objects in the industrial business process, clarifying that subsequent industrial execution objects can only enter the executable state after the preceding industrial execution object completes its corresponding decision execution. The process structure dependencies are different from connection relationships that only represent physical or logical connections. The process structure dependencies are used to describe the execution order and decision transmission constraints between industrial execution objects in the industrial business process.

[0063] The industrial cloud computing platform, based on the execution order defined in the process structure dependencies, identifies the decision transmission relationships between industrial execution objects corresponding to the decision-making behavior of AI model instances level by level, forming the transmission order of AI model instance decision-making behavior in the industrial business process. Through this transmission order, the relationships between industrial execution objects are recorded, forming the influence extension path of AI model instance decision-making behavior in the industrial business process. The influence extension path describes the process of AI model instance decision-making behavior extending level by level from the initial industrial execution object to subsequent industrial execution objects in the industrial business process.

[0064] S2.2: The industrial cloud computing platform sequentially reads the industrial business process locations contained in the influence expansion path along the path of influence expansion, and records the industrial execution object corresponding to each industrial business process location; the industrial cloud computing platform obtains the hierarchical number corresponding to the initial industrial business process location in the influence expansion path and the hierarchical number corresponding to the final industrial business process location in the influence expansion path, and calculates the hierarchical span of the AI ​​model instance decision behavior in the industrial business process based on the hierarchical number. The expression for calculating the hierarchical span is:

[0065] ;

[0066] in, This indicates the hierarchical span of AI model instance decision-making behavior within industrial business processes. This indicates the hierarchical number within the industrial business process that affects the termination position of the extended path. This indicates the hierarchical number in the industrial business process that affects the starting position of the extended path.

[0067] To further clarify, the hierarchical numbering of industrial business processes is predefined when the industrial cloud computing platform configures the industrial business processes. During the industrial business process modeling phase, the industrial cloud computing platform hierarchically divides the industrial business processes based on the control flow dependencies between various industrial execution objects within the process. Industrial business processes at the same control stage are grouped into the same process level, and a unique hierarchical number is assigned to each level. The process hierarchical number is incremented according to the execution order of the control flow within the industrial business process, representing the hierarchical relationship of progressive control from upstream to downstream. When determining the impact propagation path, the industrial cloud computing platform reads the starting and ending positions of the industrial business process from the impact propagation path and searches for the corresponding process hierarchical numbers in the industrial business process configuration, using these numbers as the basis for its determination. and Based on process hierarchy numbering, the industrial cloud computing platform calculates the hierarchical span of AI model instance decision-making behavior in industrial business processes, reflecting the range of process hierarchy traversed by the AI ​​model instance decision-making behavior in industrial business processes.

[0068] The industrial cloud computing platform determines the scope of process hierarchy covered by the decision-making behavior of AI model instances in industrial business processes based on the hierarchical span. Based on the positions of industrial business processes included within the process hierarchy, it determines the specific role of each industrial business process position in the business process, thereby determining the scope of influence of the decision-making behavior of AI model instances in industrial business processes. The scope of influence includes all business process positions directly or indirectly affected by the decision-making behavior. These positions are jointly determined by the decision output content and the hierarchical relationship in the business process.

[0069] The industrial cloud computing platform reads the process level number corresponding to each industrial business process location in the industrial business process configuration based on the location of the industrial business process covered in the influence extension path, and determines the process level range that the influence extension path crosses in the industrial business process according to the minimum and maximum values ​​of the process level number; based on the industrial business process locations included in the process level range, the industrial cloud computing platform obtains the scope of the AI ​​model instance decision behavior in the industrial business process.

[0070] After determining the scope of influence, the industrial cloud computing platform divides the industrial business process locations within the scope of influence into hierarchical levels according to the predefined process hierarchy rules in the industrial business process. The hierarchical division includes a first influence level, a second influence level, and a third influence level. The first influence level corresponds to the industrial business process locations directly affected by the decision-making behavior of the AI ​​model instance. The second influence level corresponds to the industrial business process locations indirectly affected by the decision-making behavior of the AI ​​model instance through process transmission. The third influence level corresponds to the affected industrial business process locations at the end of the process chain.

[0071] The industrial cloud computing platform generates a corresponding industrial impact level identifier for each industrial business process location based on its impact level, and records the industrial impact level identifier in relation to the decision-making behavior of the AI ​​model instance to form an industrial impact level result.

[0072] S3: Based on the results of the industrial impact hierarchy and the current operating status of the industrial cloud computing platform, determine the amplification state of AI decision-making and generate the decision amplification state results.

[0073] S3.1: In an industrial cloud computing environment, the industrial cloud computing platform continuously acquires its current operational status information. This operational status information is provided by the platform's scheduling management, resource isolation management, and task execution management functions. From this operational status information, the platform extracts computing power scheduling rhythm, resource isolation boundary status, execution buffer queue depth, and task queuing sequence characteristics directly related to the execution of AI model instance decision-making behaviors. Specifically, the computing power scheduling rhythm represents the frequency and interval at which the industrial cloud computing platform allocates computing resources to AI model instance decision-making behaviors at different points in time; the resource isolation boundary status represents the resource isolation center boundary defined by the industrial cloud computing platform for AI model instance decision-making behaviors, determined by the computing power resource quota, storage resource quota, and network access domain range allocated to the AI ​​model instance decision-making behaviors, thus limiting the resource range available to the AI ​​model instance decision-making behaviors; the execution buffer queue depth represents the number of tasks queued for execution when the industrial cloud computing platform executes AI model instance decision-making behaviors; and the task queuing sequence characteristics represent the entry order and waiting order of tasks corresponding to AI model instance decision-making behaviors in the execution queue.

[0074] The industrial cloud computing platform organizes the computing power scheduling rhythm, resource isolation boundary status, execution buffer queue depth, and task queuing timing characteristics in chronological order. It combines these characteristics within the same time window to form a set of operational status features. This set of operational status features represents the resource capacity of the industrial cloud computing platform for AI model instance decision-making behavior over a continuous time period.

[0075] S3.2: After forming a set of operational status characteristics, the industrial cloud computing platform compares the computing power scheduling rhythm, resource isolation boundary status, execution buffer queue depth, and task queuing sequence characteristics within adjacent time windows in the set, according to time sequence. The comparison method is as follows: the computing power scheduling rhythm, resource isolation boundary status, execution buffer queue depth, and task queuing sequence characteristics in the later time window are compared with those in the previous time window, i.e., the differences in the values ​​of the corresponding characteristics. Based on the direction and whether the value differences have changed, the industrial cloud computing platform obtains the change results of the operational status characteristics between adjacent time windows. The change results indicate whether the characteristics have increased, decreased, or remained unchanged in the time dimension.

[0076] Based on the comparison results, the feedback change characteristics of the decision-making behavior of AI model instances on the operational status of the industrial cloud computing platform are identified. The identification process is as follows: the computing power scheduling rhythm, resource isolation boundary status, execution buffer queue depth, and task queuing sequence characteristics within three consecutive time windows in the operational status feature set are compared item by item in chronological order. The three consecutive time windows are used to form the comparison results of two adjacent changes, thereby confirming that the changes in operational status characteristics are continuous rather than single fluctuations. Using fewer than three time windows cannot distinguish between instantaneous changes and continuous feedback, while using more than three time windows does not affect the judgment result but increases the judgment delay. Therefore, three consecutive time windows are the minimum effective time scale for identifying feedback change characteristics. When the computing power scheduling rhythm, resource isolation boundary state, execution buffer queue depth, or task queuing sequence characteristics exhibit continuous changes in the same direction within three time windows, it is determined that the corresponding operational status characteristic has feedback change characteristics. Continuous changes in the same direction include continuous increases in scheduling frequency, continuous expansion of isolation range, continuous increases in queue depth, or continuous extension of queuing sequence. When the direction of change of the same operational status characteristic from the first time window to the second time window is inconsistent with that from the second time window to the third time window, it is determined that the corresponding operational status characteristic does not form feedback change characteristics. When the change result of the same operational status characteristic remains unchanged between any adjacent time windows, it is determined that the corresponding operational status characteristic does not form feedback change characteristics. The identified computing power scheduling rhythm feedback change characteristics, resource isolation boundary state feedback change characteristics, execution buffer queue depth feedback change characteristics, and task queuing sequence feedback change characteristics are organized in chronological order to form a decision-making status feedback characteristic set.

[0077] S3.3: The industrial cloud computing platform reads the industrial impact level identifier corresponding to the decision behavior of the AI ​​model instance from the industrial impact level result, and determines the industrial business process level range corresponding to the industrial impact level result based on the industrial impact level identifier, and generates an industrial business process location identifier set containing multiple industrial business process location identifiers. Each industrial business process location identifier in the industrial business process location identifier set corresponds to an industrial execution object within the industrial impact level range.

[0078] The system searches for records in the operational status feature set that match the location identifiers of each industrial business process in the industrial business process location identifier set. It then extracts the corresponding computing power scheduling rhythm, resource isolation boundary status, execution buffer queue depth, and task queuing sequence features. The industrial cloud computing platform then aggregates these extracted features according to the industrial business process location identifiers, forming a subset of operational status features that corresponds one-to-one with the industrial business process location identifier set.

[0079] Find records from the decision-making situation feedback feature set that match the location identifiers of each industrial business process in the industrial business process location identifier set, extract the feedback change features corresponding to the records, and similarly group the extracted feedback change features according to the industrial business process location identifiers to form a decision-making situation feedback feature subset that corresponds one-to-one with the industrial business process location identifier set.

[0080] For each industrial business process location identifier in the set of industrial business process location identifiers, the industrial cloud computing platform reads the corresponding subset of operational status features and subset of decision-making status feedback features. It then combines the computing power scheduling rhythm, resource isolation boundary status, execution buffer queue depth, and task queuing sequence features in the subset of operational status features with the feedback change features in the subset of decision-making status feedback features to generate a description of the carrying status of the corresponding industrial business process location.

[0081] Preferably, the present invention analyzes the correlation between the decision-making behavior of AI model instances and the operational status of industrial cloud computing platforms. Based on the continuous changes in operational status characteristics within a defined time window, it identifies the continuous operational status feedback caused by the decision-making behavior of AI model instances. Combined with the results of industrial impact hierarchy, the operational status analysis is limited to the corresponding industrial business process level. This allows the capacity of the industrial cloud computing platform to bear the decision-making behavior of AI model instances to be accurately characterized, thereby providing a direct basis for determining the decision amplification state.

[0082] S3.4: The industrial cloud computing platform parses and processes the carrying capacity status results, reads the values ​​of the carrying capacity status results at multiple consecutive time points in chronological order, compares the carrying capacity status results at adjacent time points, records the changes in the carrying capacity status results within a continuous time period, and forms a carrying capacity change sequence.

[0083] The industrial cloud computing platform identifies carrying capacity change trends based on a carrying capacity change sequence. When the carrying capacity value in the carrying capacity change sequence decreases sequentially at three consecutive time points, a state of decreasing carrying capacity is identified within the corresponding industrial impact level. When the carrying capacity value in the carrying capacity change sequence shows alternating increases and decreases between three consecutive time points, and the difference between values ​​at adjacent time points is not zero, a state of intensified carrying capacity fluctuation is identified within the corresponding industrial impact level. The three consecutive time points are used to form two adjacent change intervals, thereby simultaneously covering the change occurrence stage and the change reversal stage, in order to distinguish between continuous change trends and single-point instantaneous fluctuations. If fewer than three time points are used, a complete change interval cannot be formed; if more than three time points are used, it will not increase the judgment accuracy and will lead to unnecessary judgment delays. Therefore, three consecutive time points are selected as the minimum effective time scale for carrying capacity trend identification. When the carrying capacity value in the carrying capacity change sequence has decreased and has not recovered to the carrying capacity value at the corresponding time point before the decrease after three consecutive time points, a state of delayed carrying capacity recovery is identified within the corresponding industrial impact level.

[0084] The industrial cloud computing platform aligns the carrying capacity change sequence with the feedback change features in the decision-making situation feedback feature set in chronological order. When the change nodes in the carrying capacity change sequence and the feedback change features recorded in the decision-making situation feedback feature set appear repeatedly in time, the AI ​​model instance decision behavior is identified as forming a repeatedly triggered execution state in the industrial cloud computing environment. When multiple (at least two) feedback change features continuously correspond to the same carrying capacity change trend in adjacent time periods, the AI ​​model instance decision behavior is identified as forming a continuously superimposed execution state in the industrial cloud computing environment. When the change result in the carrying capacity change sequence acts in reverse on the subsequent decision-making situation feedback features and triggers at least one new carrying capacity change again, the AI ​​model instance decision behavior is identified as forming a loop amplification execution state in the industrial cloud computing environment.

[0085] The industrial cloud computing platform identifies the amplification state of AI model instance decision-making behavior based on a combination of states: decreased capacity, increased capacity fluctuation, delayed capacity recovery, and loop amplification. The identification rules are as follows: when only a decreased capacity state is detected, the AI ​​model instance decision-making behavior is identified as Level 1 amplification; when both decreased capacity and increased capacity fluctuation are detected simultaneously, it is identified as Level 2 amplification; when all three states—decreased capacity, increased capacity fluctuation, and delayed capacity recovery—are detected simultaneously, it is identified as Level 3 amplification; and when any of these states is detected along with loop amplification, it is identified as Level 4 amplification. The industrial cloud computing platform records the amplification state identification along with the AI ​​model instance decision-making behavior, generating a decision amplification state result.

[0086] Preferably, this invention performs serial analysis on the changes in carrying capacity status results over a defined time scale, and combines this with a set of decision-making situation feedback features to uniformly characterize the changes in carrying capacity, execution triggering patterns, and loop amplification phenomena caused by the decision-making behavior of AI model instances in an industrial cloud computing environment. This allows the decision amplification state to be graded, identified, and recorded. By combining the judgment of carrying capacity decline, increased carrying capacity fluctuation, carrying capacity recovery delay, and loop amplification execution states, it avoids the problem of risk judgment based solely on a single carrying capacity indicator or instantaneous state. This provides clear judgment rules and traceable results for the amplification degree of AI model instance decision-making behavior, offering a direct and reliable decision-making basis for subsequent safety risk classification and risk response management.

[0087] S4: Based on the results of the decision amplification state, the results of the comprehensive industrial impact level, and the controllability of the current cloud environment, the degree of security risk of AI decision-making is classified and determined, and a risk level result is generated.

[0088] The industrial cloud computing platform analyzes the decision amplification status results, reads the amplification status identifier corresponding to the decision behavior of the AI ​​model instance, and determines the amplification status level corresponding to the decision behavior of the AI ​​model instance. At the same time, it reads the industrial impact level identifier corresponding to the decision behavior of the AI ​​model instance from the industrial impact level results, searches for the industrial business process level interval corresponding to the industrial impact level identifier in the industrial business process configuration based on the industrial impact level identifier, extracts the industrial business process location identifiers located within the industrial business process level interval, and forms a set of industrial business process location identifiers, which is used to determine the impact level range of the decision behavior of the AI ​​model instance in the industrial business process.

[0089] The industrial cloud computing platform quantifies the controllability of AI model instance decision-making behavior under the current operating conditions of the industrial cloud computing environment by combining resource scheduling margin status, execution buffer availability status, execution continuity status, and execution interruptibility status. The quantification method is as follows: the resource scheduling margin status, execution buffer availability status, execution continuity status, and execution interruptibility status are mapped to discrete state values, and the state values ​​corresponding to the same point in time are combined into a state vector. The corresponding intervention capability status is determined according to the position of the state vector in the preset state space. The intervention capability status represents the state result of the controllability of the current cloud environment under specific operating conditions.

[0090] To further explain, the industrial cloud computing platform predefines the judgment rules for each state during the operation strategy configuration phase: the resource scheduling reserve state is determined based on the quantitative relationship between the allocable computing resources of the industrial cloud computing platform within the current time window and the computing resources required for the decision-making behavior of the AI ​​model instance. The industrial cloud computing platform pre-sets high reserve thresholds and low reserve thresholds during the operation strategy configuration phase. When the allocable computing resources are greater than the computing resources required for the decision-making behavior of the AI ​​model instance, and the proportion of remaining computing power to the required computing power is not less than the high reserve threshold, the resource scheduling reserve state is determined to be a high reserve state. When the allocable computing resources are only slightly higher than the computing resources required for the decision-making behavior of the AI ​​model instance, and the proportion of remaining computing power to the required computing power is between the high reserve threshold and the low reserve threshold, the resource scheduling reserve state is determined to be a limited reserve state. When the allocable computing resources are less than or equal to the computing resources required for the decision-making behavior of the AI ​​model instance, the resource scheduling reserve state is determined to be a low reserve state. The execution buffer availability status is determined based on the relationship between the available buffer capacity in the execution buffer queue and the number of tasks currently awaiting execution. When the available buffer capacity is greater than the number of tasks, it is considered sufficient; when the available buffer capacity equals the number of tasks, it is considered strained; and when the available buffer capacity is less than the number of tasks, it is considered exhausted. The execution continuity status is determined based on whether the AI ​​model instance's decision-making behavior in the industrial business process can be split or delayed. When the decision-making behavior must be completed continuously during execution, it is considered continuous execution; when the decision-making behavior can be executed in segments or delayed, it is considered discontinuous execution. The execution interruptibility status is determined based on whether the AI ​​model instance's decision-making behavior can be interrupted during execution. When the decision-making behavior can be paused or terminated without affecting the consistency of the industrial business process, it is considered interruptible; when the decision-making behavior cannot be interrupted, it is considered uninterruptible.

[0091] It should also be noted that the high and low margin thresholds are set based on the resource scheduling security redundancy requirements of the industrial cloud computing platform to ensure controllable scheduling capabilities even under conditions of computing power fluctuations and sudden task disruptions. Specifically, the high margin threshold is set at no less than 30% of the remaining allocable computing power required for the current AI model instance's decision-making behavior. This 30% high margin threshold ensures that the platform has sufficient flexibility to adjust scheduling and handle sudden tasks when computing power demand fluctuates, without affecting the stable execution of existing tasks. The low margin threshold is set at no more than 10% of the required computing power. This is used to define a state where computing power is close to saturation and scheduling flexibility is significantly limited. The 10% low margin threshold allows the platform to issue warnings and take emergency measures when computing power is close to saturation, preventing excessive consumption of computing power from causing crashes or scheduling failures.

[0092] The industrial cloud computing platform pre-configures intervention capability mapping rules during the operation strategy configuration phase to clarify the intervention capability status corresponding to different resource state combinations. These rules include at least the following: A strong intervention capability status is determined when the resource scheduling margin is high, the execution buffer availability is sufficient, and the execution interruptibility status is interruptible; a medium intervention capability status is determined when the resource scheduling margin is limited, the execution buffer availability is strained, and the execution continuity status is continuous; and a weak intervention capability status is determined when the resource scheduling margin is low, the execution buffer availability is exhausted, the execution continuity status is continuous, and the execution interruptibility status is uninterruptible. During the determination process, the industrial cloud computing platform uses the resource states acquired at the same time point as a set of inputs and determines the intervention capability status of the industrial cloud computing environment on the decision-making behavior of AI model instances according to the intervention capability mapping rules.

[0093] The industrial cloud computing platform combines and processes resource scheduling reserve status, execution buffer availability status, execution continuity status, and execution interruptibility status according to pre-configured state correspondences. These pre-configured state correspondences are configured in the form of rule tables during the industrial cloud computing platform's operational strategy configuration phase. These rule tables pre-assign a corresponding intervention capability status for each combination of values ​​for resource scheduling reserve status, execution buffer availability status, execution continuity status, and execution interruptibility status. During the combined processing, the industrial cloud computing platform treats the resource scheduling reserve status, execution buffer availability status, execution continuity status, and execution interruptibility status acquired simultaneously as a set of state inputs. It then searches the rule table for records matching these state inputs and reads the corresponding intervention capability status from these records to determine the industrial cloud computing environment's intervention capability status for the AI ​​model instance's decision-making behavior.

[0094] The industrial cloud computing platform uses amplified state level, scope of impact, and intervention capability status as risk assessment parameters, and combines them for assessment according to a pre-configured risk classification mapping relationship. This risk classification mapping relationship is configured during the industrial cloud computing platform's risk strategy configuration phase. The security risk level of AI model instance decision-making behavior is divided into at least three risk levels, from low to high, representing controllable risk, limited risk, and uncontrollable risk. This is used to define the risk level range corresponding to different combinations of amplified state level, scope of impact, and intervention capability status. The risk classification mapping relationship includes at least the following judgment rules: when the amplified state level is level one or two, the scope of impact only includes the first impact level, and the intervention capability status is strong, the risk level is determined to be controllable; when the amplified state level is level three, the scope of impact includes the second impact level, and the intervention capability status is medium, the risk level is determined to be limited; when the amplified state level is level four, or the scope of impact includes the third impact level, or the intervention capability status is weak, the risk level is determined to be uncontrollable.

[0095] During the judgment process, the industrial cloud computing platform uses the current amplification level, the scope of impact, and the status of intervention capability as a set of judgment inputs. It then searches for a mapping record that matches the current set of inputs in the risk classification mapping relationship, reads the corresponding risk level in the mapping record, classifies and identifies the degree of security risk of the AI ​​model instance's decision-making behavior, and generates a risk level result.

[0096] S5: Based on the risk level results, generate corresponding risk response management strategies to control the security risks of AI decision-making in the industrial cloud environment.

[0097] The industrial cloud computing platform reads the risk level identifier corresponding to the decision-making behavior of the AI ​​model instance from the risk level results, and simultaneously reads the intervention capability status corresponding to the decision-making behavior of the AI ​​model instance. The industrial cloud computing platform uses the risk level identifier and the intervention capability status as joint judgment parameters, and determines the risk response control level according to the pre-configured risk response mapping relationship. The risk response mapping relationship is configured during the risk response strategy configuration stage of the industrial cloud computing platform, and includes at least the following mapping rules: when the risk level identifier is controllable and the intervention capability status is strong, the risk response control level is determined to be low risk control level; when the risk level identifier is limited and the intervention capability status is medium, the risk response control level is determined to be medium risk control level; when the risk level identifier is out of control or the intervention capability status is weak, the risk response control level is determined to be high risk control level.

[0098] Based on the risk response control level, the execution frequency, execution order, and execution triggering conditions of AI model instance decision-making behavior are limited. The limitation method is as follows: When the risk response control level is low, the original execution frequency, execution order, and execution triggering conditions of AI model instance decision-making behavior are maintained. When the risk response control level is medium, the industrial cloud computing platform calculates the maximum allowed execution frequency of AI model instance decision-making behavior under the current operating state based on the resource scheduling margin status and execution buffer availability status corresponding to the carrying capacity status results. The maximum execution frequency is limited to the highest number of executions allowed per unit time without causing a continuous increase in the execution buffer queue depth or triggering a decrease in carrying capacity. The execution order of AI model instance decision-making behavior in the industrial business process is also delayed. When the risk response control level is high, the automatic triggering of AI model instance decision-making behavior is suspended, and the execution triggering condition for AI model instance decision-making behavior is limited to manual confirmation.

[0099] The industrial cloud computing platform combines execution frequency limitation rules, execution order limitation rules, and execution trigger condition limitation rules to generate risk response management strategies corresponding to the decision-making behavior of AI model instances, and applies these risk response management strategies to the execution control of AI model instance decision-making behavior in the industrial cloud computing environment.

[0100] This embodiment also provides a computer device applicable to the AI ​​security risk assessment method based on industrial cloud computing, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the AI ​​security risk assessment method based on industrial cloud computing as proposed in the above embodiment.

[0101] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0102] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the AI ​​security risk assessment method based on industrial cloud computing as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0103] In summary, this invention, by analyzing deployed AI model instances and associating their decision-making behavior with specific industrial execution objects, clarifies the actual execution path and scope of AI decisions within industrial business processes. This solves the problem in existing technologies where AI decisions are evaluated as isolated events, making it difficult to pinpoint their business impact. Furthermore, by combining the industrial impact hierarchy results with the operational status of the industrial cloud computing platform, it identifies the repetitive triggering, continuous superposition, and loop amplification states of AI model instance decision-making behavior under the influence of resource scheduling, task queuing, and execution feedback. This enables effective determination of the amplification effect of AI decisions, preventing security risks from being gradually amplified during operation.

[0104] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for assessing AI security risks based on industrial cloud computing, characterized in that: include: In an industrial cloud computing environment, deployed AI model instances are parsed to identify the actual AI model instance decision-making behaviors triggered by the AI ​​model instances, and associated with the corresponding industrial execution objects to generate decision execution association results. Based on the results of decision execution correlation, determine the impact extension path of AI model instance decision behavior, determine the scope of AI model instance decision behavior, and generate industrial impact hierarchy results; Based on the results of the industrial impact hierarchy and the current operating status of the industrial cloud computing platform, the amplification status of the decision-making behavior of the AI ​​model instance is determined, and the decision amplification status result is generated. The determination of the amplified state of the decision-making behavior of the AI ​​model instance includes: Obtain the current operational status information of the industrial cloud computing platform, analyze the operational status information, and extract the computing power scheduling rhythm, resource isolation boundary status, execution buffer queue depth, and task queuing sequence features directly triggered by the decision-making behavior of AI model instances. The computing power scheduling rhythm, resource isolation boundary status, execution buffer queue depth, and task queuing sequence characteristics are organized in chronological order to form a set of operational status characteristics; Based on the set of operational status features, the feedback change characteristics of AI model instance decision-making behavior on the operational status of industrial cloud computing platform are identified, forming a set of decision-making status feedback features. The carrying capacity status result is generated by combining the set of operational status characteristics, the set of decision-making status feedback characteristics, and the results of industrial impact levels. The generated decision amplification state result includes: The results of the carrying capacity status are analyzed to extract the change information of the carrying capacity status results over a continuous time period and generate a carrying capacity change sequence. Based on the carrying capacity change sequence, identify the states of carrying capacity decline, increased carrying capacity fluctuation and delayed carrying capacity recovery within the corresponding industrial impact level; By using the sequence of changes in carrying capacity and the set of feedback features of decision-making status, we can identify the execution state of repeated triggering, continuous superposition, and loop amplification of AI model instance decision-making behavior in the industrial cloud computing environment. Based on the states of decreased carrying capacity, increased carrying capacity fluctuation, delayed carrying capacity recovery, and the execution state of loop amplification, the amplification state corresponding to the decision-making behavior of AI model instances is identified, and decision amplification state results are generated. Based on the results of the decision amplification state, the results of the industrial impact hierarchy, and the controllability of the current cloud environment, the degree of security risk of the decision-making behavior of AI model instances is classified and determined, and a risk level result is generated. Based on the risk level results, corresponding risk response management strategies are generated to control the security risks of AI model instance decision-making behavior in the industrial cloud environment.

2. The AI ​​security risk assessment method based on industrial cloud computing as described in claim 1, characterized in that: The generated decision execution association results include: The runtime status information and decision output records of deployed AI model instances are analyzed to extract the decision output content generated by the AI ​​model instances during operation, and the decision behavior and decision triggering conditions of the AI ​​model instances are determined based on the decision output content. The decision-making behavior of AI model instances is associated with predefined industrial execution objects in industrial business processes to generate decision execution association results between AI model instance decision-making behavior and industrial execution objects.

3. The AI ​​security risk assessment method based on industrial cloud computing as described in claim 1, characterized in that: The scope of determining the decision-making behavior of AI model instances includes: Analyze the results of decision execution to extract the industrial execution objects and industrial business process locations corresponding to the decision-making behavior of AI model instances; Based on the location of industrial business processes, the transmission order and correlation of AI model instance decision-making behavior in industrial business processes are analyzed to determine the impact extension path of AI model instance decision-making behavior in industrial business processes. Based on the location of industrial business processes in the influence expansion path, the covered business process levels are identified layer by layer, and the identified business process levels are used as the scope of the AI ​​model instance decision-making behavior.

4. The AI ​​security risk assessment method based on industrial cloud computing as described in claim 1, characterized in that: The process of generating industrial impact hierarchy results includes: hierarchically dividing the scope of the decision-making behavior of AI model instances to form corresponding industrial impact hierarchy identifiers, recording the correspondence between the industrial impact hierarchy identifiers and the decision-making behavior of AI model instances, and generating industrial impact hierarchy results.

5. The AI ​​security risk assessment method based on industrial cloud computing as described in claim 1, characterized in that: The classification and determination of the security risk level of AI model instance decision-making behavior includes: By analyzing the execution status of AI model instance decision-making behavior in the industrial cloud computing environment through the decision amplification status results, the amplification status level corresponding to the AI ​​model instance decision-making behavior is obtained, and the influence level range of AI model instance decision-making behavior in industrial business processes is determined based on the industrial influence level results. Based on the results of the carrying capacity status and the decision amplification status, the controllability of the industrial cloud computing environment to the decision-making behavior of AI model instances under the current operating conditions is quantitatively evaluated, and the intervention capability status of the industrial cloud computing environment to the decision-making behavior of AI model instances is extracted. The risk assessment criteria are formed by combining the amplified status level, the scope of impact, and the intervention capability status.

6. The AI ​​security risk assessment method based on industrial cloud computing as described in claim 1, characterized in that: The risk level results are obtained by classifying and labeling the degree of security risk of AI model instance decision-making behavior based on risk assessment criteria.

7. The AI ​​security risk assessment method based on industrial cloud computing as described in claim 5, characterized in that: The extraction of the industrial cloud computing environment's ability to intervene in the decision-making behavior of AI model instances specifically includes: By analyzing the carrying capacity status results, we can obtain the resource scheduling margin status and execution buffer availability status of the industrial cloud computing environment within the corresponding industrial impact level. At the same time, by analyzing the decision amplification status results, we can obtain the execution continuity status and execution interruptibility status of the AI ​​model instance decision behavior in the industrial cloud computing environment. According to the preset state correspondence, the resource scheduling margin state, execution buffer available state, execution continuity state, and execution interruptibility state are combined and mapped to the intervention capability state.

8. The AI ​​security risk assessment method based on industrial cloud computing as described in claim 1, characterized in that: The risk response management strategy is generated by limiting the execution frequency, execution order, and execution triggering conditions of AI model instance decision-making behavior in the industrial cloud computing environment through risk level results and intervention capability status.

Images