Distributed equipment cluster security evaluation system and method based on adaptive target selection, edge computing device and medium

Abstract

The invention discloses a distributed equipment cluster security assessment method and system based on adaptive target selection, an edge computing device and a medium, and the method comprises the steps: collecting the multi-dimensional state data of a distributed equipment cluster in real time, and calculating the threat index of each terminal through an attention mechanism and reinforcement learning; determining a target terminal based on the threat index, and determining a malicious action; simulating the operation logic of the target terminal through the lightweight GAN model to obtain a simulation prediction result, generating a disturbance signal based on the simulation prediction result, and injecting the disturbance signal into the target terminal to induce the target terminal to execute a malicious action; and obtaining a system performance difference before and after the distributed device cluster attacks, generating a reinforcement learning reward signal, and updating model parameters. According to the method, efficient, accurate and dynamic safety assessment of the distributed equipment cluster is realized through a closed-loop process of state acquisition, risk assessment, target determination, action generation, simulation prediction, disturbance injection and feedback optimization.

Description

A security assessment system, method, edge computing device, and medium for distributed device clusters based on adaptive target selection. Technical Field

[0001] This invention relates to the field of distributed device cluster security assessment technology, and in particular to a distributed device cluster security assessment method, system, edge computing device and medium based on adaptive target selection. Background Technology

[0002] With the rapid evolution of technologies such as the Internet of Things (IoT) and smart cities, distributed device clusters have become core infrastructure in production and daily life. In the industrial sector, distributed device clusters, consisting of industrial robots in automobile manufacturing lines, SCADA system controllers in wind power plants, and AGV navigation modules in smart factories, support the efficient operation of core production processes. In the smart city sector, device clusters such as intersection traffic signal control networks and building HVAC (heating, ventilation, and air conditioning) clusters directly impact urban operational efficiency and residents' quality of life. In emerging fields, new types of distributed device clusters, such as OpenHarmony ecosystem devices and OBU (On-Board Unit) devices in vehicle-road cooperative systems, are gradually becoming key carriers for technological innovation and industrial upgrading.

[0003] The secure and stable operation of distributed device clusters is a core prerequisite for ensuring production continuity, urban security, and service availability. Therefore, security assessment technologies for distributed device clusters have become a key focus of industry research and application. Their core objective is to identify high-risk terminals within the cluster in real time (such as those that have been compromised or have firmware vulnerabilities), verify the attack resistance capabilities of the terminals and the cluster, and proactively identify security vulnerabilities and disrupt attack chains.

[0004] Although existing technologies have made some progress in the field of security assessment of distributed device clusters, there are still some technical defects in practical application scenarios, which seriously restrict the assessment effect and industrial implementation. These defects are mainly manifested in: security assessment delays, inability to quickly block attacks; poor compatibility of black-box devices; insufficient adaptability, high false alarm rate and inability to cope with dynamic scenarios.

[0005] Therefore, existing technologies still need improvement. Summary of the Invention

[0006] To address the aforementioned deficiencies in existing technologies, this invention provides a method, system, edge computing device, and medium for security assessment of distributed device clusters based on adaptive target selection. The technical solution adopted by this invention is as follows: Firstly, this invention provides a method for security assessment of distributed device clusters based on adaptive target selection. The method includes: real-time acquisition of multi-dimensional state data of all terminals in the distributed device cluster; calculating a threat index for each terminal based on the multi-dimensional state data using an attention mechanism and reinforcement learning; filtering high-risk terminals based on the threat index to identify target terminals; searching for malicious actions that would cause the target terminal to produce the greatest negative chain reaction; imitating the operating logic of the target terminal using a lightweight GAN model to obtain simulation prediction results; generating a perturbation signal based on the simulation prediction results; injecting the perturbation signal into the target terminal to induce the target terminal to execute the malicious action; obtaining the system performance difference of the distributed device cluster before and after the attack; generating a reinforcement learning reward signal; and updating the calculation model parameters based on the reinforcement learning reward signal.

[0007] In one implementation, multi-dimensional status data of all terminals in a distributed device cluster are collected in real time, including: configuring a load sensor, a network latency monitoring module, and an energy consumption metering chip for each terminal in the distributed device cluster, and deploying a data aggregation gateway on the edge computing node; collecting multi-dimensional status data of each terminal in real time based on the load sensor, network latency monitoring module, and energy consumption metering chip, and summarizing the collected multi-dimensional status data based on the data aggregation gateway.

[0008] In one implementation, based on the multi-dimensional state data, a threat index for each terminal is calculated using an attention mechanism and reinforcement learning, including: constructing a global state matrix based on the multi-dimensional state data; inputting the global state matrix into a dynamic risk assessment engine; evaluating the global risk level using the threat value function in the dynamic risk assessment engine based on the attention mechanism; and calculating the threat index for each terminal using a softmax function.

[0009] In one implementation, the process of filtering high-risk terminals based on the threat index and determining target terminals includes: comparing the threat index of each terminal with a preset threshold, filtering out threat indices greater than the preset threshold; determining the high-risk terminals based on the terminals corresponding to threat indices greater than the preset threshold; and determining the target terminals from the high-risk terminals through Top-k filtering.

[0010] In one implementation, determining the high-risk terminals based on the terminals corresponding to threat indices greater than the preset threshold further includes: selecting terminals corresponding to threat indices greater than the preset threshold as candidate terminals and obtaining importance weights for the candidate terminals; and selecting a number of top-ranked terminals from the candidate terminals as high-risk terminals based on the importance weights.

[0011] In one implementation, for a locked target terminal, searching for malicious actions that cause the target terminal to produce the greatest negative chain reaction includes: pre-defining an action space, the action space including action types and action constraints, the action types including hardware actions, software actions and combined actions; based on the target terminal, combined with the global state matrix, searching for malicious actions that cause the target terminal to produce the greatest negative chain reaction in the preset action space.

[0012] In one implementation, based on the target terminal and the global state matrix, a malicious action that causes the target terminal to produce the greatest negative chain reaction is searched in a preset action space. This includes: training an action evaluation network based on a Q-Learning model; inputting the target terminal, the global state matrix, and the action type into the action evaluation network, and outputting an action damage effect score; and obtaining the malicious action that causes the target terminal to produce the greatest negative chain reaction based on the action damage effect score.

[0013] In one implementation, a simulation prediction result is obtained by mimicking the operating logic of the target terminal using a lightweight GAN model, including: inputting the target terminal and the malicious action into the lightweight GAN model; learning the mapping relationship between the input and output of the target terminal based on the generator in the lightweight GAN model; and determining the simulation prediction result after the target terminal executes the malicious action based on the mapping relationship.

[0014] In one implementation, generating a disturbance signal based on the simulation prediction results includes: determining an optimization objective, wherein the optimization objective is a balance between the minimum disturbance amplitude and the action execution effect; solving the optimization objective using an L-BFGS constraint solver based on the simulation prediction results, and generating a disturbance signal under the condition of satisfying the hidden constraints and behavior induction requirements.

[0015] In one implementation, injecting the disturbance signal into the target terminal to induce the target terminal to perform the malicious action further includes: if the target terminal does not respond to the malicious action, increasing the disturbance step size of the disturbance signal.

[0016] In one implementation, the process of obtaining the system performance difference before and after the attack on the distributed device cluster, generating a reinforcement learning reward signal, and updating model parameters based on the reinforcement learning reward signal includes: obtaining the baseline performance of the distributed device cluster before the attack and the post-attack performance after the injection of the perturbation signal; determining the system performance difference based on the baseline performance and the post-attack performance; generating a reinforcement learning reward signal based on the reward function based on the baseline performance and the post-attack performance; and updating the parameters in the dynamic risk assessment engine based on the reinforcement learning reward signal using a policy gradient algorithm.

[0017] In one implementation, the method involves obtaining the system performance difference before and after the attack on the distributed device cluster, generating a reinforcement learning reward signal, and updating the model parameters based on the reinforcement learning reward signal. The method also includes updating the parameters in the action evaluation network and the parameters in the lightweight GAN model based on the reinforcement learning reward signal.

[0018] Secondly, embodiments of the present invention also provide a distributed device cluster security assessment system based on adaptive target selection. The system implements the steps of the distributed device cluster security assessment method based on adaptive target selection as described in any of the above schemes. The system includes: a risk assessment module, used to collect multi-dimensional state data of all terminals in the distributed device cluster in real time, and calculate the threat index of each terminal based on the multi-dimensional state data through an attention mechanism and reinforcement learning; a malicious action determination module, used to filter high-risk terminals based on the threat index, determine target terminals, and search for malicious actions that cause the target terminal to produce the largest negative chain reaction; a simulation and induction module, used to simulate the operating logic of the target terminal through a lightweight GAN model, obtain simulation prediction results, generate a perturbation signal based on the simulation prediction results, and inject the perturbation signal into the target terminal to induce the target terminal to execute the malicious action; and a result analysis module, used to obtain the system performance difference before and after the distributed device cluster attack, generate a reinforcement learning reward signal, and update the model parameters based on the reinforcement learning reward signal.

[0019] In one implementation, the risk assessment module includes: a state matrix construction unit, used to construct a global state matrix based on the multi-dimensional state data; a risk level assessment unit, used to input the global state matrix into a dynamic risk assessment engine, and assess the global risk level based on an attention mechanism and using the threat value function in the dynamic risk assessment engine; and a threat index calculation unit, used to calculate the threat index of each terminal using a softmax function.

[0020] In one implementation, the malicious action determination module includes: a threat index filtering unit, used to compare the threat index of each terminal with a preset threshold and filter out threat indices greater than the preset threshold; a high-risk terminal determination unit, used to determine the high-risk terminals based on the terminals corresponding to threat indices greater than the preset threshold; and a target terminal determination unit, used to determine the target terminal from the high-risk terminals through Top-k filtering.

[0021] In one implementation, the malicious action determination module further includes: an action space definition unit, used to predefine an action space, the action space including action types and action constraints, the action types including hardware actions, software actions and combined actions; and a malicious action determination unit, used to search for malicious actions that cause the target terminal to produce the largest negative chain reaction in the preset action space based on the target terminal and in combination with the global state matrix.

[0022] In one implementation, the simulation and induction module includes: a data input unit for inputting the target terminal and the malicious action into the lightweight GAN model; a mapping relationship learning unit for learning the mapping relationship between the input and output of the target terminal based on the generator in the lightweight GAN model; a simulation prediction unit for determining the simulation prediction result after the target terminal executes the malicious action based on the mapping relationship; an optimization target determination unit for determining the optimization target, wherein the optimization target is a balance between the minimum perturbation amplitude and the action execution effect; and a perturbation signal generation unit for solving the optimization target through an L-BFGS constraint solver based on the simulation prediction result, and generating a perturbation signal under the condition of satisfying the concealment constraint and behavior induction requirements.

[0023] In one implementation, the result analysis module includes: a system performance difference determination unit, used to obtain the baseline performance of the distributed device cluster before the attack and the post-attack performance after the injection of the disturbance signal, and to determine the system performance difference based on the baseline performance and the post-attack performance; a reward signal generation unit, used to generate a reinforcement learning reward signal based on the baseline performance and the post-attack performance and a reward function; and a parameter update unit, used to update the parameters in the dynamic risk assessment engine based on the reinforcement learning reward signal using a policy gradient algorithm.

[0024] Thirdly, embodiments of the present invention also provide an edge computing device, wherein the edge computing device includes a memory, a processor, and a distributed device cluster security assessment program based on adaptive target selection stored in the memory and executable on the processor. When the processor executes the distributed device cluster security assessment program based on adaptive target selection, it implements the steps of the distributed device cluster security assessment method based on adaptive target selection in any of the above schemes.

[0025] Fourthly, embodiments of the present invention also provide a computer-readable storage medium, wherein the computer-readable storage medium stores a distributed device cluster security assessment program based on adaptive target selection, the distributed device cluster security assessment program based on adaptive target selection implementing the steps of the distributed device cluster security assessment method based on adaptive target selection as described in any of the above schemes on the computer-readable storage medium.

[0026] Beneficial Effects: Compared with existing technologies, this invention provides a distributed device cluster security assessment method based on adaptive target selection. First, this invention collects multi-dimensional state data of all terminals in the distributed device cluster in real time. Based on this multi-dimensional state data, it calculates the threat index of each terminal through an attention mechanism and reinforcement learning. Then, based on the threat index, high-risk terminals are screened to determine target terminals. For the locked target terminals, malicious actions that cause the largest negative chain reaction are searched. Next, a lightweight GAN model is used to simulate the operating logic of the target terminal to obtain simulation prediction results. Based on the simulation prediction results, a perturbation signal is generated and injected into the target terminal to induce the target terminal to execute the malicious action. Finally, the system performance difference before and after the attack on the distributed device cluster is obtained, a reinforcement learning reward signal is generated, and the model parameters are updated based on the reinforcement learning reward signal. This invention achieves efficient, accurate, and dynamic security assessment of distributed device clusters through a closed-loop process of state acquisition, risk assessment, target determination, action generation, simulation prediction, perturbation injection, and feedback optimization. Attached Figure Description

[0027] Figure 1 shows the application environment of the distributed device cluster security assessment method based on adaptive target selection provided in the embodiment of the present invention.

[0028] Figure 2 is a flowchart illustrating a preferred embodiment of the distributed device cluster security assessment method based on adaptive target selection provided by the present invention.

[0029] Figure 3 is a schematic diagram of the technical route of the distributed device cluster security assessment method based on adaptive target selection provided in the embodiment of the present invention.

[0030] Figure 4 is a schematic diagram of the execution flow of the dynamic risk assessment engine in the distributed device cluster security assessment method based on adaptive target selection provided in the embodiment of the present invention.

[0031] Figure 5 is a schematic diagram of the process for selecting target terminals in the distributed device cluster security assessment method based on adaptive target selection provided in the embodiment of the present invention.

[0032] Figure 6 is a flowchart illustrating the malicious action in the distributed device cluster security assessment method based on adaptive target selection provided in an embodiment of the present invention.

[0033] Figure 7 is a schematic diagram of the simulation prediction process in the distributed device cluster security assessment method based on adaptive target selection provided in the embodiment of the present invention.

[0034] Figure 8 is a system block diagram of the distributed device cluster security assessment system based on adaptive target selection provided in an embodiment of the present invention.

[0035] Figure 9 is a schematic block diagram of the edge computing device provided in an embodiment of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0037] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content, operations, or steps, nor does it require execution in the described order. For example, some operations or steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0038] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0039] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. For example, the first control information and the second control information are only used to distinguish different control information and do not limit their order.

[0040] Those skilled in the art will understand that the words "first" and "second" do not limit the quantity or the order of execution, and that the words "first" and "second" do not necessarily imply that they are different.

[0041] It should also be understood that the terms "and / or" as used in this specification and the appended claims refer to any combination of one or more of the associated listed items and all possible combinations, and include such combinations.

[0042] Currently, mainstream distributed device cluster security assessment technologies are mainly divided into three categories: 1. Cloud-based centralized assessment scheme: After the terminal collects full status data (such as device load, network latency, and energy consumption), it uploads the data to the cloud server via the network. The cloud calculates the terminal's risk level based on fixed rules or simple machine learning models, and then issues response instructions. The core advantage of this scheme is centralized management and sufficient computing power, but it relies on the upload of full data and centralized cloud computing, which has an inherent latency defect.

[0043] 2. Edge Distributed Evaluation Scheme: This scheme migrates some computing tasks to edge nodes (such as industrial gateways and edge servers) to reduce data upload volume and improve response speed. However, existing edge solutions do not optimize the target selection mechanism and still require evaluating all terminals in the cluster one by one, resulting in high computational load and insufficient compatibility with black-box devices and new devices.

[0044] Static rule assessment scheme: Security assessment is based on fixed rules preset by humans (such as judging high risk if CPU utilization is >90%). The scheme is simple to implement and has low deployment cost, but it cannot adapt to dynamic behavior of the cluster (such as production line reorganization and changes in equipment collaboration relationships), and the false alarm rate remains high.

[0045] Despite some progress in the field of distributed device cluster security assessment, three major pain points remain in practical applications, severely hindering assessment effectiveness and industrial implementation. Specifically: 1. The contradiction between assessment latency and real-time performance: the inability to block the full data upload and centralized computing mode of cloud-based centralized solutions for rapid attacks results in persistently high response latency. Taking an automotive manufacturing production line as an example, research found that 200 industrial robots each collect 8-dimensional status data per second, with a single data upload volume of up to 500MB, a network transmission latency of approximately 2 seconds, a cloud computing latency of approximately 3 seconds, and a total response latency exceeding 5 seconds. Furthermore, the attack chain trigger time for APT (Advanced Persistent Threat) attacks is typically within 3 seconds. By the time the cloud issues defense commands, the attack has already caused welding robots to misoperate, resulting in welding defects in the car body, causing significant losses to each production line in a single instance.

[0046] Even with existing edge-distributed solutions, the lack of optimized target selection mechanisms still requires evaluating each terminal individually. The evaluation latency for a 200-node cluster remains at 2.3 seconds, failing to meet the real-time requirements of industrial scenarios. This latency issue leads to a "post-event remediation" rather than "pre-event defense" approach to security assessments, failing to effectively block rapid attack chains.

[0047] II. Black-box device compatibility defects and violations of industrial safety standards: Many black-box devices exist in industrial settings (such as pitch system controllers in wind farms and compressor controllers in building HVAC systems). These devices cannot have agent programs installed or their internal configurations modified, otherwise, they would violate industrial safety standards such as IEC 62443. Traditional assessment methods rely on pre-installed agents or specific protocol libraries to collect data, resulting in near-zero assessment coverage for black-box devices.

[0048] Meanwhile, with the popularization of new operating systems such as OpenHarmony, HarmonyOS devices using HDF (Hardware Driver Foundation) (such as AGV navigation modules and smart sensors) are increasingly being used in industrial scenarios. Traditional solutions cannot adapt to the HDF driver architecture, have limited protocol coverage for HarmonyOS devices, and cannot collect core status data (such as the running status of path planning algorithms and the communication status of the driver layer), making these new devices a "blind spot" for security assessment.

[0049] Existing compatibility solutions attempt to adapt black-box devices and HarmonyOS devices through protocol conversion gateways, but the gateway introduces additional latency and has a high data loss rate during protocol conversion, which seriously affects the accuracy of the assessment. Some solutions attempt to modify the device configuration to install a proxy, which not only violates industrial safety regulations but may also cause abnormal device operation and lead to production accidents.

[0050] Third, static strategies lack adaptability, have a high false alarm rate, and cannot cope with dynamic scenarios. The core evaluation logic of existing evaluation schemes is mostly based on fixed rules, and such static strategies cannot adapt to the dynamic behavior of distributed device clusters. In dynamic scenarios such as flexible production line reorganization, changes in equipment collaboration relationships, and fluctuations in business load, the normal operating status of equipment will change significantly, and the original static rules will become invalid, leading to a surge in false alarm rates.

[0051] Taking the restructuring of an automobile manufacturing production line as an example, research has found that the collaboration mode between welding robots and handling robots has changed after the restructuring. The CPU utilization rate of welding robots is normally maintained at 85% (the original rule set 80% as the high-risk threshold), causing the false alarm rate of traditional solutions to soar. Security personnel are overwhelmed by a large number of invalid alarms, which not only increases labor costs but may also cause them to miss real attacks.

[0052] Existing dynamic adjustment solutions attempt to adapt to dynamic scenarios by manually updating rules, but the update cycle is as long as 24 hours. Threat detection accuracy decreases within 24 hours after production line reorganization, making it unable to cope with rapid dynamic changes in the cluster. Static strategies lack adaptability, resulting in a significant reduction in the reliability of security assessments and making it difficult to meet the accuracy requirements of security assessments in industrial production and urban operations.

[0053] To address this, this embodiment provides a distributed device cluster security assessment method based on adaptive target selection. In specific application, this embodiment first collects multi-dimensional state data of all terminals in the distributed device cluster in real time. Based on this multi-dimensional state data, an attention mechanism and reinforcement learning are used to calculate the threat index of each terminal. Then, high-risk terminals are screened based on the threat index to determine target terminals. For the locked target terminals, malicious actions that would cause the largest negative chain reaction are searched. Next, a lightweight GAN model is used to simulate the operating logic of the target terminal to obtain simulation prediction results. Based on these simulation prediction results, a perturbation signal is generated and injected into the target terminal to induce it to execute the malicious action. Finally, the system performance difference before and after the attack on the distributed device cluster is obtained, a reinforcement learning reward signal is generated, and the model parameters are updated based on the reinforcement learning reward signal.

[0054] The distributed device cluster security assessment method based on adaptive target selection in this embodiment can be applied to the application environment shown in Figure 1. Terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated on server 104, or it can be located in the cloud or on other network servers. The distributed device cluster security assessment method based on adaptive target selection can be executed by terminal 102 or server 104, or it can be executed collaboratively by terminal 102 and server 104.

[0055] The terminal 102 can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, IoT device, or portable wearable device. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices, etc.

[0056] Server 104 can be an independent physical server, a cloud server, or a service node in a blockchain system, where the service nodes in the blockchain system form a peer-to-peer network.

[0057] In addition, server 104 can also be a server cluster consisting of multiple physical servers, which can be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0058] Terminal 102 and server 104 can be connected via Bluetooth, USB (Universal Serial Bus) or network, etc., and this application does not impose any restrictions.

[0059] In some embodiments, as shown in Figure 2, a distributed device cluster security assessment method based on adaptive target selection is provided. This method can be applied to terminals. Specifically, the method includes the following steps: Step S100: Real-time collection of multi-dimensional state data of all terminals in the distributed device cluster, and based on the multi-dimensional state data, calculation of the threat index of each terminal through attention mechanism and reinforcement learning.

[0060] In this embodiment, the edge-cloud collaborative architecture deploys core computing tasks on the edge computing device to achieve a local closed loop of detection and response. The cloud management platform is responsible for knowledge base storage, global policy coordination, and historical data statistical analysis. The edge computing device is the core execution unit of this embodiment, as shown in Figure 3.

[0061] Specifically, in this embodiment, each terminal in the distributed device cluster is configured with a load sensor, a network latency monitoring module, and an energy consumption metering chip, and a data aggregation gateway is deployed at the edge computing node. The load sensor is used to collect CPU / memory usage; the network latency monitoring module is used to collect network uplink / downlink latency and packet loss rate; and the energy consumption metering chip is used to collect device energy consumption values. Therefore, multi-dimensional status data of each terminal is collected in real time based on the load sensor, network latency monitoring module, and energy consumption metering chip, and the collected multi-dimensional status data is summarized based on the data aggregation gateway. This embodiment can also adapt black-box devices to HarmonyOS devices.

[0062] Furthermore, after collecting multi-dimensional state data, this embodiment can perform preprocessing such as denoising, normalization, and integrity verification on the multi-dimensional state data. Specifically, denoising can employ a sliding window denoising algorithm with a window size of 5 and a sliding step size of 1. Abnormal jump values ​​are replaced by the mean of the data within the window to filter sensor noise. Normalization can use a Min-Max normalization algorithm to map all dimensional data to the [0,1] interval, avoiding the influence of dimensional differences on subsequent risk calculations. Integrity verification can use a CRC32 checksum algorithm to verify the collected data. If verification fails, the data is re-collected, with a maximum of 3 retries. If it still fails, it is marked as missing data, and the ARIMA (AutoRegressive Integrated Moving Average) model is used to predict the current state data to ensure the integrity of the state matrix.

[0063] In one implementation, as shown in Figure 4, step S100 of this embodiment includes the following steps: Step S101: Construct a global state matrix based on the multi-dimensional state data; Step S102: Input the global state matrix into the dynamic risk assessment engine, and evaluate the global risk level based on the attention mechanism and the threat value function in the dynamic risk assessment engine; Step S103: Calculate the threat index of each terminal using the softmax function.

[0064] Specifically, after collecting multi-dimensional state data, this embodiment constructs a global state matrix:

[0065] in, This indicates the number of terminals in a distributed device cluster. Indicates the first individual terminals The dimension is denoted by R, which represents a real number. In practical applications, the InfluxDB time-series database can be used to store the global state matrix. The database is sharded by day and supports fast retrieval by terminal ID and timestamp, meeting the needs of subsequent risk calculation and historical data backtracking, and providing an input basis for subsequent risk calculation.

[0066] As shown in Figure 3, after acquiring the aforementioned multi-dimensional state data and constructing a global state matrix based on the edge computing device, the global state matrix can be input into the dynamic risk assessment engine, which is deployed on the GPU core of the edge computing device. Next, based on an attention mechanism, the threat value function in the dynamic risk assessment engine is used to assess the global risk level. Then, the threat index for each terminal is calculated using the softmax function.

[0067] Specifically, this embodiment employs a 4-head attention mechanism (head_dim=64, model_dim=256) to dynamically assign weights to the aforementioned multi-dimensional state data. The threat value function in this embodiment is expressed as:

[0068] in, The output is the global risk level. To quantify the destructive potential of a single terminal in the current state, the value range is [0,1]. Here, M represents the global anomaly probability, M is the cluster size coefficient, and W is the dynamic weight of the attention mechanism. The global risk level calculated in this embodiment can be pushed to the security posture dashboard shown in Figure 3 in real time.

[0069] The softmax function is expressed as: Q represents the number of dynamic weights, and the parameter... The model can be updated via online reinforcement learning. The update frequency is every 10 sampling periods, with a learning rate η=0.01, balancing real-time performance and model stability. Parameter updates are triggered when the global risk level is ≥0.5 or the standard deviation of the terminal threat index is >0.2, ensuring timely model iteration as the cluster state changes. In this embodiment, calculating the threat index of each terminal helps identify high-risk terminals in subsequent steps.

[0070] Step S200: Based on the threat index, screen high-risk terminals, determine target terminals, and search for malicious actions that would cause the target terminal to have the greatest negative chain reaction.

[0071] In one implementation, as shown in Figure 5, this embodiment includes the following steps when determining the target terminal: Step S201: Compare the threat index of each terminal with a preset threshold and filter out the threat indices that are greater than the preset threshold; Step S202: Based on the terminals corresponding to the threat indices that are greater than the preset threshold, determine the high-risk terminals; Step S203: Determine the target terminal from the high-risk terminals through Top-k filtering.

[0072] In practical applications, the logic for selecting the target terminal in this embodiment is as follows:

[0073]

[0074] For the selected target terminal. This is a preset threshold.

[0075] Specifically, this embodiment compares the threat index of each terminal with a preset threshold and selects terminals with a threat index greater than the preset threshold as candidate terminals. Next, this embodiment assigns importance weights to the candidate terminals. The importance weights are set based on the terminal's role in the cluster. For example, core business terminals (such as welding robots and traffic signal controllers) have an importance weight of 0.6, ordinary terminals (such as auxiliary sensors) have an importance weight of 0.3, and backup terminals have an importance weight of 0.1. The importance weights in this embodiment can be manually configured through a cloud management platform or automatically calculated based on the terminal's historical traffic volume. Then, based on the importance weights, this embodiment selects several high-risk terminals from the candidate terminals. In specific implementation, this embodiment calculates the priority score of the candidate terminals using the formula: Priority = Threat Index × Importance Weight. The terminals are sorted in descending order of priority, and several high-risk terminals are selected. If the number of candidate terminals is less than 3, all are considered high-risk terminals to ensure the stability of the target terminal quantity. Further, a Top-k approach is used. The filtering algorithm (k=3) selects the three highest-priority terminals from high-risk terminals as target terminals, reducing the computational overhead of subsequent malicious action generation and simulation. Furthermore, this embodiment can also verify the reachability of the selected target terminals by using ping commands and port scanning to confirm that the terminals are online and network-connected; if a target terminal is unreachable, it is replaced sequentially from the high-risk terminals to ensure the effectiveness of the target terminals.

[0076] Furthermore, this embodiment predefines an action space, which includes action types and action constraints. The action types include hardware actions, software actions, and combined actions. Specifically, hardware actions include forced hibernation (to trigger a terminal hibernation command), CPU overclocking, network disconnection (to shut down the terminal's network interface), and power depletion simulation (to modify the energy consumption metering chip reading). Software actions include tampering with system configuration (to modify the terminal's core configuration file), occupying core memory (to create a process with excessively large memory), terminating critical processes (to kill terminal service processes), and injecting junk data (to transmit invalid data packets to the terminal). Combined actions include overclocking + injecting junk data, network disconnection + terminating critical processes, and forced hibernation + tampering with system configuration. The action constraints in this embodiment exclude actions that directly lead to hardware damage or permanent data loss (such as physical power-off decoupling or disk formatting) to ensure test security; one-click recovery is supported after the actions are executed, without affecting the normal operation of the terminal.

[0077] Next, based on the target terminal and the global state matrix, this embodiment searches for malicious actions that cause the target terminal to produce the largest negative chain reaction in a preset action space. In practical applications, as shown in Figure 6, this embodiment includes the following steps when determining malicious actions: Step S21, train an action evaluation network based on a Q-Learning model; Step S22, input the target terminal, the global state matrix, and the action type into the action evaluation network, and output an action damage effect score; Step S23, based on the action damage effect score, obtain the malicious action that causes the target terminal to produce the largest negative chain reaction.

[0078] Specifically, during model training, the training environment is a simulation environment built on Docker containers, simulating the operating logic of over 100 common terminal models (such as industrial robots, traffic signal controllers, and HarmonyOS AGVs). The training data can contain the correspondence between 100,000 real terminal actions and responses. The Q-Learning model is configured as follows: State space: containing the target terminal state vector, global state matrix, and action type a, with a state space dimension of d+N+d×|A| (d is the acquisition dimension, N is the number of terminals, and |A| is the number of actions); Action value function: (s, a)= w concat (s, a)+b, where w is the weight matrix (dimension 128×(d+N+d×|A|)), b is the bias vector (dimension 128×1), and the activation function is ReLU; training parameters: learning rate η=0.01, discount factor γ=0.9, exploration rate ε=0.1 (decays linearly with the number of iterations, decay rate 0.001 / iteration), number of iterations ≥ 100,000, and convergence condition is Q-value change ≤ 0.001.

[0079] This embodiment constructs an action evaluation network based on a trained Q-Learning model. The input is a concatenated vector of "target terminal + global state matrix + action type," and the output is an action damage effect score. The score ranges from -1 to 1, where 1 represents optimal damage and -1 represents no damage. Therefore, this embodiment can use the target terminal state... Global state matrix Each action 'a' in action space A is sequentially input into the action evaluation network to obtain a damage effect score for each action; the action with the highest score is selected as the malicious action. The search logic for the action evaluation network is as follows: .

[0080] The action evaluation network in this embodiment supports online incremental updates. When the actual effect of a malicious action deviates from the score by more than 10%, the model is updated. The learning rate η is dynamically adjusted. For example, when the attack success rate is ≥80%, η=0.005; when the success rate is ≤60%, η=0.02, to ensure that the model adapts to changes in terminal status.

[0081] Step S300: Simulate the operating logic of the target terminal using a lightweight GAN model to obtain simulation prediction results. Based on the simulation prediction results, generate a perturbation signal and inject the perturbation signal into the target terminal to induce the target terminal to perform the malicious action.

[0082] Specifically, after selecting the target terminal and determining the malicious action, the behavior simulation module in Figure 3 can be used for simulation prediction. As shown in Figure 7, the simulation prediction in this embodiment includes the following steps: Step S301, inputting the target terminal and the malicious action into the lightweight GAN model; Step S302, learning the mapping relationship between the input and output of the target terminal based on the generator in the lightweight GAN model; Step S303, determining the simulation prediction result after the target terminal executes the malicious action based on the mapping relationship.

[0083] The lightweight GAN model in this embodiment includes a generator and a discriminator, as detailed below: Generator Employing a depthwise separable convolutional architecture, it includes 3 convolutional layers (3×3 kernel size, stride 1), 2 pooling layers (max pooling, 2×2 kernel size), and 1 fully connected layer (output dimension d). It has <2MB of parameters, inference latency ≤5ms, and supports deployment on edge computing devices. The generator's input is the target terminal state. With malicious actions The output is the terminal response prediction value (such as changes in CPU utilization and the increase in network latency).

[0084] Discriminator A two-layer fully connected network is used, with the input being "real terminal response + generator response" and the output being the discrimination score (value range [0,1], 1 for real response, 0 for generator response), with a parameter size of <500KB and an inference latency of ≤2ms.

[0085] This embodiment can utilize a supervised learning pre-trained generator. The training data consists of mappings between 100,000 real terminal actions and responses, and the loss function is MSE (mean squared error). The training objective of adversarial training is:

[0086] Where 'o' represents the state of the target terminal, and 'a' represents the malicious action. Based on this, this embodiment utilizes a generator in a lightweight GAN model to learn the mapping relationship between the input and output of the target terminal. Then, based on this mapping relationship, this embodiment can determine the simulation prediction result after the target terminal executes the malicious action; this simulation prediction result is the predicted response value of the target terminal to the malicious action. This embodiment utilizes a generator... It replaces the actual terminal response and enables prediction of black-box environmental disturbances.

[0087] In this embodiment, data input: the data input unit inputs the target terminal status. malicious actions The input is a lightweight GAN model, and the input data is normalized (value range [0,1]) to ensure the consistency of the model input.

[0088] Mapping relation learning: The mapping relation learning unit is based on a generator. The input-output mapping relationship of the target terminal is learned, that is, the mapping of "state + action → response". The mapping relationship is stored in the cache of the edge computing device. The cache is valid for 30 minutes to avoid repeated learning.

[0089] Simulation prediction output: Based on the mapping relationship, the simulation prediction unit outputs the malicious actions executed by the target terminal. The subsequent simulation prediction results include changes in terminal hardware performance (e.g., CPU utilization spikes to 95%), network performance changes (e.g., network latency increases by 200ms), and changes in service execution status (e.g., task execution failure rate ≥80%), while also labeling the "response credibility" (based on a discriminator). The discrimination score is ≥0.8 for high confidence and <0.8 for low confidence.

[0090] Furthermore, this embodiment defines an optimization objective, which is a balance between minimizing the perturbation amplitude and the effect of action execution. Then, based on the simulation prediction results, the optimization objective is solved using an L-BFGS constraint solver, and a perturbation signal is generated under the conditions of satisfying the hidden constraints and behavior induction requirements.

[0091] The optimization objective is:

[0092] Among them, The action similarity loss function uses MSE loss to measure the feature differences between malicious and normal actions, ensuring that the perturbation signal is not recognized by security software. δ represents the perturbation signal (dimension consistent with the terminal state vector), and ||δ||² is the L2 norm of the perturbation amplitude, ensuring that the perturbation amplitude is minimized. Concealment constraint: the perturbation's... (Human-machine imperceptible threshold) means that changes in terminal state caused by disturbances ≤5%, which are imperceptible to the user. Behavioral inducement constraints and malicious actions. The expected response deviation is ≤10%, ensuring that the disturbance signal can effectively induce the terminal to perform malicious actions.

[0093] This embodiment employs an L-BFGS (Limited-memory Broyden-Fletcher-Goldfarb-Shanno) constraint solver to solve the optimization objective. The solver performs at least 50 iterations, has a convergence threshold of ≤0.001, and a solution time of ≤20ms, meeting real-time requirements. Upon receiving a disturbance signal, this embodiment injects the disturbance signal into the target terminal to induce it to perform the malicious action. The disturbance signal injection can be performed during off-peak hours (such as the standby time of industrial robots or the green light switching period of traffic signal controllers) to avoid affecting normal operations. The injection timing is automatically determined by the edge computing device based on terminal status data. This embodiment can transmit the disturbance signal based on the TC protocol, selecting an idle port on the terminal (determined through port scanning). Transmission encryption uses the AES-128 algorithm, and the key is pre-negotiated between the edge computing device and the terminal to ensure transmission security. If the injection of the disturbance signal fails (e.g., due to network interruption), it will be retried twice, with a retry interval of 50ms. If it still fails, it will be marked as an injection anomaly, an alarm will be pushed to the security posture dashboard, and the terminal will be removed from the target terminal list. If the target terminal does not respond to the malicious action, the disturbance step size of the disturbance signal will be increased.

[0094] Step S400: Obtain the system performance difference before and after the attack on the distributed device cluster, generate a reinforcement learning reward signal, and update the model parameters based on the reinforcement learning reward signal.

[0095] This embodiment can obtain the baseline performance of the distributed device cluster before the attack and the post-attack performance after the injection of the disturbance signal. Based on the baseline performance and the post-attack performance, the system performance difference is determined. Specifically, this embodiment can monitor four types of core performance indicators before and after the attack, covering both local terminal performance and global cluster performance. These include: local terminal performance: changes in CPU / memory utilization, task completion rate, and response time; global cluster performance: cluster collaborative communication latency, associated terminal task failure rate, and overall cluster availability (node ​​disconnection rate).

[0096] This embodiment can collect baseline performance data recorded before the attack (1 minute before the perturbation signal is injected). Record post-attack performance 30 seconds after the attack (after perturbation signal injection). The performance data was collected at a frequency of 10ms / time, and the average value was taken as the final performance value to ensure data accuracy. Next, based on the baseline performance and the post-attack performance, a reinforcement learning reward signal was generated based on the reward function, expressed as: Finally, the policy gradient algorithm is used to update the parameters in the dynamic risk assessment engine based on the reinforcement learning reward signal. Similarly, this embodiment can also update the parameters in the action evaluation network and the parameters in the lightweight GAN model based on the reinforcement learning reward signal, requiring only an improvement to the policy gradient algorithm described above.

[0097] Compared with the prior art, the present invention has the following significant advantages: 1. Extremely low evaluation latency, meeting real-time requirements: Through the Top-k filtering mechanism of the key terminal selector, only the relevant data of 3 target terminals are processed, the evaluation latency of the cluster is greatly reduced, and the APT attack chain can be effectively blocked.

[0098] 2. High compatibility, adaptable to black-box and new devices: Adopting non-intrusive black-box testing technology, it uses a lightweight GAN model to clone terminals without the need for pre-installing agents or modifying device configurations, thereby improving the coverage of black-box device evaluation and solving compatibility defects.

[0099] 3. Dynamic adaptation and low false alarm rate: A closed-loop feedback architecture is constructed, and the evaluation strategy is updated in real time through reinforcement learning to meet the accurate evaluation needs of dynamic clusters.

[0100] 4. Low resource consumption and support for large-scale clusters: The optimized dynamic risk assessment engine and lightweight GAN model reduce the memory footprint of a single node; the edge-cloud collaborative architecture distributes a large amount of computing load to edge devices, supporting parallel assessment of multi-device clusters.

[0101] 5. Closed-loop optimization to continuously improve evaluation results: Reinforcement learning reward signals are generated by the difference in system performance before and after the attack, driving the update of model parameters across the entire chain. Combined with the reuse of historical data in the knowledge base, the evaluation accuracy of new scenarios is improved. Malicious actions are generated more specifically, and the attack effect is quantified more accurately, providing data support for cluster security protection.

[0102] Based on the above embodiments, the present invention also provides a distributed device cluster security assessment system based on adaptive target selection. The system in this embodiment is used to implement the steps of the above method embodiments. As shown in Figure 8, the system includes: a risk assessment module 10, a malicious action determination module 20, a simulation and induction module 30, and a result analysis module 40. Specifically, the risk assessment module 10 is used to collect multi-dimensional state data of all terminals in the distributed device cluster in real time, and calculate the threat index of each terminal based on the multi-dimensional state data through attention mechanisms and reinforcement learning. The malicious action determination module 20 is used to filter high-risk terminals based on the threat index, determine target terminals, and search for malicious actions that cause the target terminal to produce the largest negative chain reaction. The simulation and induction module 30 is used to simulate the operating logic of the target terminal through a lightweight GAN model, obtain simulation prediction results, generate a perturbation signal based on the simulation prediction results, and inject the perturbation signal into the target terminal to induce the target terminal to execute the malicious action. The result analysis module 40 is used to obtain the system performance difference before and after the distributed device cluster attack, generate a reinforcement learning reward signal, and update the model parameters based on the reinforcement learning reward signal.

[0103] In one implementation, the risk assessment module 10 includes: a state matrix construction unit, used to construct a global state matrix based on the multi-dimensional state data; a risk level assessment unit, used to input the global state matrix into a dynamic risk assessment engine, and assess the global risk level based on an attention mechanism and using the threat value function in the dynamic risk assessment engine; and a threat index calculation unit, used to calculate the threat index of each terminal using a softmax function.

[0104] In one implementation, the malicious action determination module 20 includes: a threat index filtering unit, used to compare the threat index of each terminal with a preset threshold and filter out threat indices greater than the preset threshold; a high-risk terminal determination unit, used to determine the high-risk terminals based on the terminals corresponding to threat indices greater than the preset threshold; and a target terminal determination unit, used to determine the target terminal from the high-risk terminals through Top-k filtering.

[0105] In one implementation, the malicious action determination module 20 further includes: an action space definition unit, used to predefine an action space, the action space including action types and action constraints, the action types including hardware actions, software actions and combined actions; and a malicious action determination unit, used to search for malicious actions that cause the target terminal to produce the largest negative chain reaction in the preset action space based on the target terminal and in combination with the global state matrix.

[0106] In one implementation, the simulation and induction module 30 includes: a data input unit for inputting the target terminal and the malicious action into the lightweight GAN model; a mapping relationship learning unit for learning the mapping relationship between the input and output of the target terminal based on the generator in the lightweight GAN model; a simulation prediction unit for determining the simulation prediction result after the target terminal executes the malicious action based on the mapping relationship; an optimization target determination unit for determining the optimization target, wherein the optimization target is a balance between the minimum perturbation amplitude and the action execution effect; and a perturbation signal generation unit for solving the optimization target through an L-BFGS constraint solver based on the simulation prediction result, and generating a perturbation signal under the condition of satisfying the concealment constraint and behavior induction requirements.

[0107] In one implementation, the result analysis module 40 includes: a system performance difference determination unit, used to obtain the baseline performance of the distributed device cluster before the attack and the post-attack performance after the injection of the disturbance signal, and to determine the system performance difference based on the baseline performance and the post-attack performance; a reward signal generation unit, used to generate a reinforcement learning reward signal based on the baseline performance and the post-attack performance and based on a reward function; and a parameter update unit, used to update the parameters in the dynamic risk assessment engine based on the reinforcement learning reward signal using a policy gradient algorithm.

[0108] In some embodiments, an edge computing device is provided, which can be a terminal, and its internal structure can be as shown in Figure 9. The edge computing device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor of the edge computing device provides computing and control capabilities. The memory of the edge computing device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface of the edge computing device is used for exchanging information between the processor and external devices. The communication interface of the edge computing device is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a distributed device cluster security assessment method based on adaptive target selection. The display unit of the edge computing device is used to form a visually visible image. It can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the edge computing device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the outer shell of the edge computing device, or external keyboards, touchpads, or mice, etc.

[0109] Those skilled in the art will understand that the structure shown in Figure 9 is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the edge computing device to which the present application is applied. A specific edge computing device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0110] In some embodiments, an edge computing device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the following steps: real-time acquisition of multi-dimensional state data of all terminals in a distributed device cluster, and calculation of a threat index for each terminal based on the multi-dimensional state data through an attention mechanism and reinforcement learning; screening high-risk terminals based on the threat index, identifying target terminals, and searching for malicious actions that would cause the target terminal to produce the greatest negative chain reaction; imitating the operating logic of the target terminal through a lightweight GAN model to obtain simulation prediction results, generating a perturbation signal based on the simulation prediction results, and injecting the perturbation signal into the target terminal to induce the target terminal to execute the malicious action; obtaining the system performance difference of the distributed device cluster before and after the attack, generating a reinforcement learning reward signal, and updating the calculation model parameters based on the reinforcement learning reward signal.

[0111] It should be noted that the user information (including but not limited to user device function information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0112] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0113] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0114] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for security assessment of distributed device clusters based on adaptive target selection, characterized in that, The method includes: real-time collection of multi-dimensional state data of all terminals in a distributed device cluster; calculation of a threat index for each terminal based on the multi-dimensional state data using an attention mechanism and reinforcement learning; screening high-risk terminals based on the threat index to identify target terminals; searching for malicious actions that would cause the target terminal to produce the greatest negative chain reaction; obtaining simulation prediction results by imitating the operating logic of the target terminal using a lightweight GAN model; generating a perturbation signal based on the simulation prediction results; injecting the perturbation signal into the target terminal to induce the target terminal to execute the malicious action; obtaining the system performance difference before and after the attack on the distributed device cluster; generating a reinforcement learning reward signal; and updating the model parameters based on the reinforcement learning reward signal.

2. The distributed device cluster security assessment method based on adaptive target selection according to claim 1, characterized in that, Real-time collection of multi-dimensional status data of all terminals in a distributed device cluster includes: configuring a load sensor, a network latency monitoring module, and an energy consumption metering chip for each terminal in the distributed device cluster, and deploying a data aggregation gateway on the edge computing node; collecting multi-dimensional status data of each terminal in real time based on the load sensor, network latency monitoring module, and energy consumption metering chip, and summarizing the collected multi-dimensional status data based on the data aggregation gateway.

3. The distributed device cluster security assessment method based on adaptive target selection according to claim 2, characterized in that, Based on the multi-dimensional state data, the threat index of each terminal is calculated through attention mechanism and reinforcement learning, including: constructing a global state matrix based on the multi-dimensional state data; inputting the global state matrix into a dynamic risk assessment engine; evaluating the global risk level using the threat value function in the dynamic risk assessment engine based on the attention mechanism; and calculating the threat index of each terminal through the softmax function.

4. The distributed device cluster security assessment method based on adaptive target selection according to claim 3, characterized in that, The process of filtering high-risk terminals based on the threat index and determining target terminals includes: comparing the threat index of each terminal with a preset threshold, filtering out threat indices greater than the preset threshold; determining the high-risk terminals based on the terminals corresponding to threat indices greater than the preset threshold; and determining the target terminals from the high-risk terminals through Top-k filtering.

5. The distributed device cluster security assessment method based on adaptive target selection according to claim 4, characterized in that, Determining high-risk terminals based on the terminals corresponding to threat indices greater than the preset threshold further includes: selecting terminals corresponding to threat indices greater than the preset threshold as candidate terminals and obtaining importance weights for the candidate terminals; and selecting a number of top-ranked terminals from the candidate terminals as high-risk terminals based on the importance weights.

6. The distributed device cluster security assessment method based on adaptive target selection according to claim 5, characterized in that, For a locked target terminal, search for malicious actions that cause the target terminal to have the greatest negative chain reaction, including: pre-defining an action space, the action space including action types and action constraints, the action types including hardware actions, software actions and combined actions; based on the target terminal, combined with the global state matrix, search for malicious actions that cause the target terminal to have the greatest negative chain reaction in the preset action space.

7. The distributed device cluster security assessment method based on adaptive target selection according to claim 6, characterized in that, Based on the target terminal and the global state matrix, a malicious action that causes the target terminal to have the greatest negative chain reaction is searched in a preset action space. This includes: training an action evaluation network based on a Q-Learning model; inputting the target terminal, the global state matrix, and the action type into the action evaluation network and outputting an action damage effect score; and obtaining the malicious action that causes the target terminal to have the greatest negative chain reaction based on the action damage effect score.

8. The distributed device cluster security assessment method based on adaptive target selection according to claim 7, characterized in that, The simulation prediction results are obtained by mimicking the operating logic of the target terminal using a lightweight GAN model, including: inputting the target terminal and the malicious action into the lightweight GAN model; learning the mapping relationship between the input and output of the target terminal based on the generator in the lightweight GAN model; and determining the simulation prediction results after the target terminal executes the malicious action based on the mapping relationship.

9. The distributed device cluster security assessment method based on adaptive target selection according to claim 8, characterized in that, Based on the simulation prediction results, a disturbance signal is generated, including: determining an optimization objective, wherein the optimization objective is a balance between the minimum disturbance amplitude and the action execution effect; based on the simulation prediction results, solving the optimization objective using an L-BFGS constraint solver, and generating a disturbance signal under the condition of satisfying the hidden constraints and behavior induction requirements.

10. The distributed device cluster security assessment method based on adaptive target selection according to claim 9, characterized in that, Injecting the disturbance signal into the target terminal to induce the target terminal to perform the malicious action further includes: if the target terminal does not respond to the malicious action, increasing the disturbance step size of the disturbance signal.

11. The distributed device cluster security assessment method based on adaptive target selection according to claim 9, characterized in that, The process involves obtaining the system performance difference before and after the attack on the distributed device cluster, generating a reinforcement learning reward signal, and updating the parameters for calculating the threat index based on the reinforcement learning reward signal. This includes: obtaining the baseline performance of the distributed device cluster before the attack and the post-attack performance after the injection of the perturbation signal; determining the system performance difference based on the baseline performance and the post-attack performance; generating a reinforcement learning reward signal based on the reward function and the baseline performance and the post-attack performance; and updating the model parameters based on the reinforcement learning reward signal using a policy gradient algorithm.

12. The distributed device cluster security assessment method based on adaptive target selection according to claim 11, characterized in that, The method involves obtaining the system performance difference before and after the attack on the distributed device cluster, generating a reinforcement learning reward signal, and updating the model parameters based on the reinforcement learning reward signal. It also includes updating the parameters in the action evaluation network and the parameters in the lightweight GAN model based on the reinforcement learning reward signal.

13. A distributed device cluster security assessment system based on adaptive target selection, characterized in that, The system is used to implement the steps of the distributed device cluster security assessment method based on adaptive target selection as described in any one of claims 1-12. The system includes: a risk assessment module, used to collect multi-dimensional state data of all terminals in the distributed device cluster in real time, and calculate the threat index of each terminal based on the multi-dimensional state data through attention mechanism and reinforcement learning; a malicious action determination module, used to filter high-risk terminals based on the threat index, determine target terminals, and search for malicious actions that cause the target terminal to produce the largest negative chain reaction; a simulation and induction module, used to simulate the operating logic of the target terminal through a lightweight GAN model, obtain simulation prediction results, generate a perturbation signal based on the simulation prediction results, and inject the perturbation signal into the target terminal to induce the target terminal to execute the malicious action; and a result analysis module, used to obtain the system performance difference before and after the attack on the distributed device cluster, generate a reinforcement learning reward signal, and update the model parameters based on the reinforcement learning reward signal.

14. The distributed device cluster security assessment system based on adaptive target selection according to claim 13, characterized in that, The risk assessment module includes: a state matrix construction unit, used to construct a global state matrix based on the multi-dimensional state data; a risk level assessment unit, used to input the global state matrix into a dynamic risk assessment engine, and assess the global risk level based on an attention mechanism and the threat value function in the dynamic risk assessment engine; and a threat index calculation unit, used to calculate the threat index of each terminal using the softmax function.

15. The distributed device cluster security assessment system based on adaptive target selection according to claim 14, characterized in that, The malicious action determination module includes: a threat index filtering unit, used to compare the threat index of each terminal with a preset threshold and filter out threat indices greater than the preset threshold; a high-risk terminal determination unit, used to determine the high-risk terminals based on the terminals corresponding to threat indices greater than the preset threshold; and a target terminal determination unit, used to determine the target terminal from the high-risk terminals through Top-k filtering.

16. The distributed device cluster security assessment system based on adaptive target selection according to claim 15, characterized in that, The malicious action determination module further includes: an action space definition unit, used to predefine an action space, the action space including action types and action constraints, the action types including hardware actions, software actions and combined actions; and a malicious action determination unit, used to search for malicious actions that cause the target terminal to produce the largest negative chain reaction in the preset action space based on the target terminal and in combination with the global state matrix.

17. The distributed device cluster security assessment system based on adaptive target selection according to claim 16, characterized in that, The simulation and induction module includes: a data input unit for inputting the target terminal and the malicious action into the lightweight GAN model; a mapping relationship learning unit for learning the mapping relationship between the input and output of the target terminal based on the generator in the lightweight GAN model; a simulation prediction unit for determining the simulation prediction result after the target terminal executes the malicious action based on the mapping relationship; an optimization target determination unit for determining the optimization target, which is a balance between the minimum perturbation amplitude and the action execution effect; and a perturbation signal generation unit for solving the optimization target through an L-BFGS constraint solver based on the simulation prediction result, and generating a perturbation signal under the condition of satisfying the concealment constraint and behavior induction requirements.

18. The distributed device cluster security assessment system based on adaptive target selection according to claim 16, characterized in that, The result analysis module includes: a system performance difference determination unit, used to obtain the baseline performance of the distributed device cluster before the attack and the post-attack performance after the injection of the disturbance signal, and to determine the system performance difference based on the baseline performance and the post-attack performance; a reward signal generation unit, used to generate a reinforcement learning reward signal based on the baseline performance and the post-attack performance and a reward function; and a parameter update unit, used to update the parameters in the dynamic risk assessment engine based on the reinforcement learning reward signal using a policy gradient algorithm.

19. An edge computing device, characterized in that, The edge computing device includes a memory, a processor, and a distributed device cluster security assessment program based on adaptive target selection stored in the memory and executable on the processor. When the processor executes the distributed device cluster security assessment program based on adaptive target selection, it implements the steps of the distributed device cluster security assessment method based on adaptive target selection as described in any one of claims 1-12.

20. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a distributed device cluster security assessment program based on adaptive target selection, the distributed device cluster security assessment program based on adaptive target selection implementing the steps of the distributed device cluster security assessment method based on adaptive target selection as described in any one of claims 1-12 on the computer-readable storage medium.

Images