Federal multi-target scheduling method and system for industrial control crowdsourcing test

Abstract

The embodiment of the invention relates to the technical field of federated learning and information security, in particular to a federated multi-target scheduling method and system for industrial control crowdsourcing testing, which is characterized in that sensitive data such as ability, experience and the like of a tester are only reserved at a local terminal by constructing a framework combining local feature coding and federated model training, so that the test efficiency is improved, and the test efficiency is improved. Outward transmission is not carried out; only the model update quantity after the privacy enhancement processing such as gradient cutting and differential privacy noise injection is uploaded to the federated matching learning server, and only the model update quantity after the privacy enhancement processing is uploaded to the federated learning server, so that the leakage risk of concentrated data storage is avoided from the source, the requirements of privacy regulations such as GDPR and the like are completely met, and the reliability of the system is improved. Compared with traditional privacy technologies such as homomorphic encryption, the method has the advantages that the calculation and communication overhead is greatly reduced, the design of local matching degree calculation is matched, the task matching response speed is greatly increased, and the real-time requirement of industrial control system testing is met.

Description

Technical Field

[0001] This application relates to the fields of federated learning and information security technology, and in particular to a federated multi-objective scheduling method and system for industrial control crowdsourcing testing. Background Technology

[0002] Industrial Control Systems (ICS) are a critical global infrastructure, widely used in energy, power, transportation, manufacturing, and other fields. The security of ICS directly impacts national security and economic stability; therefore, improving their security and mitigating potential attacks is paramount. In recent years, with the development of information technology, crowdsourced testing platforms have emerged as an effective means to enhance ICS security. These platforms bring together globally distributed security testers to conduct large-scale collaborative security testing of ICS devices, protocols, and software. However, as crowdsourced testing platforms expand and application scenarios become more complex, traditional task allocation methods face challenges in areas such as privacy protection, multi-objective optimization, and system robustness.

[0003] Traditional crowdsourced testing platforms typically require the centralized collection of testers' personal information, skill data, device configurations, and historical records for task allocation and matching, posing significant privacy risks. This is especially true when handling sensitive information, potentially leading to data breaches or violations of relevant privacy regulations. To improve privacy, some research has employed techniques such as homomorphic encryption and secure multi-party computation; however, these methods often suffer from high computational overhead and communication costs, hindering their widespread adoption in practical applications.

[0004] The crowdsourced testing task allocation problem is characterized by multi-objective optimization. Task scheduling needs to consider not only task completion time but also factors such as vulnerability discovery utility, resource cost, and system robustness. Traditional single-objective optimization methods cannot effectively balance these complex objective relationships; therefore, employing multi-objective optimization algorithms becomes crucial for solving this problem. Multi-objective evolutionary algorithms can effectively handle trade-offs between multiple objectives, but in high-dimensional discrete spaces, these algorithms have slow convergence speeds and high computational costs in large-scale task spaces. Furthermore, while swarm intelligence algorithms such as particle swarm optimization have fast convergence, they are prone to early convergence when directly applied to discrete decision spaces, and their solution diversity is insufficient.

[0005] In open crowdsourced testing platforms, participants may include malicious testers who can affect the platform's effectiveness and credibility by submitting false reports, manipulating reputation scores, or launching data poisoning attacks. Existing static reputation mechanisms or rule-based detection methods often only address known attack patterns, but their adaptability is poor as attack strategies change, making them easily circumvented. Furthermore, in federated learning frameworks, malicious clients may launch model poisoning attacks, affecting model aggregation results and thus compromising task matching accuracy and overall system stability. Therefore, ensuring platform robustness in malicious attack environments has become a significant technological challenge.

[0006] In the crowdsourced testing task allocation of industrial control systems, privacy protection, multi-objective optimization, and adversarial robustness are the three core technical challenges. By proposing a privacy-preserving federated matching learning method, a hybrid metaheuristic multi-objective scheduling algorithm, and a dynamic reputation and adversarial detection subsystem, this invention provides a new technical approach to solving these problems. These innovative methods can not only effectively protect the privacy of testers and improve the quality of task matching, but also balance the needs of multi-objective optimization, enhance the robustness of the system in adversarial environments, and further improve the performance and credibility of the crowdsourced testing platform for industrial control systems. Summary of the Invention

[0007] To address this, the present invention provides a federated multi-objective scheduling method and system for crowdsourced testing of industrial control systems. This overcomes the significant privacy risks inherent in traditional crowdsourced testing platforms that typically require centralized collection of testers' personal information, skill data, equipment configurations, and historical records for task allocation and matching. In particular, the process of handling sensitive information may lead to data leakage.

[0008] To achieve the above objectives, in a first aspect, the present invention provides a federated multi-objective scheduling method for industrial control crowdsourcing testing, comprising:

[0009] Step S1: The task publisher uploads the industrial control system test task description information. The scheduling system constructs a knowledge graph based on the domain knowledge of the industrial control system. The task description information is converted into a task feature vector through the encoding module. The local agent deployed on the tester's terminal is converted into a tester feature vector through the local encoding module.

[0010] In step S2, each local agent trains a task-tester matching model using local data, generates model update information, and uploads it to the federated learning server after privacy enhancement processing. The federated learning server processes the model update information through an aggregation strategy, generates a global model, and distributes it to each local agent. Each local agent, based on the global model, task feature vector, and tester feature vector, outputs a matching score in the local terminal environment through similarity calculation or prediction model and reports it to the scheduling center.

[0011] Step S3: The scheduling center uses the matching score as the core input to construct a multi-objective scheduling model, uses a multi-objective optimization algorithm to solve for the optimal solution set, determines dynamic weights based on task urgency and system risk, and selects and outputs task allocation schemes from the optimal solution set.

[0012] Step S4: The tester performs the test task according to the task allocation scheme and uploads a behavior log containing operation data, environmental information and test results.

[0013] In step S5, the adversarial perception module evaluates the tester's performance based on the behavior log, dynamically updates the reputation information, identifies malicious testers through an anomaly detection algorithm and triggers a punishment mechanism, and the reputation information is synchronously fed back to the model aggregation in step S2 and the scheduling process in step S3.

[0014] Step S6: Use the test results, feedback data, and updated reputation information as training data for the next round of federated learning, and iteratively optimize the global model and the local model.

[0015] Furthermore, in step S1, the encoding module includes a global task encoder and a local encoder;

[0016] The knowledge graph G=(V,R) contains nodes V of industrial control system equipment type, communication protocol, vulnerability category, attack mode and test action, and edges R contain semantic associations of effects, requirements and is-variant-of.

[0017] The task feature vector is a task embedding vector, satisfying... For global task encoder, For task t j Structured descriptive information;

[0018] The tester feature vector is the tester embedding vector, which satisfies... ,in, For local encoders, For testers A set of capability characteristics and experience data.

[0019] Furthermore, in step S2, the task-tester matching model is a task-tester affinity model, and local training is performed by minimizing the objective function. To achieve, among which, The loss function is either logistic or similarity regression. For local interactive datasets, This is the reward value for local observations;

[0020] The privacy enhancement process includes: gradient pruning and differential privacy noise injection;

[0021] The federated learning server is a federated matching learning (FML) server.

[0022] Furthermore, in step S2, the aggregation strategy is a robust aggregation operator, specifically a coordinate trimming mean operator or the Krum algorithm, and the aggregation process satisfies:

[0023]

[0024] in, This represents the aggregated result of the global model in round t+1; w i Weights for testers; This refers to the amount of local model updates. Indicates robust aggregation operation; It is a set containing N elements, representing weighted model updates from N clients; the tester weight w i The system is dynamically adjusted based on the reputation information from step S5.

[0025] Furthermore, in step S3, the multi-objective scheduling model includes four objective functions, specifically:

[0026] Total utility maximization function ,in, The expected utility value for the task and the tester;

[0027] Total cost / time minimization function ,in, The matching score, i.e. the affinity score, is for step S2;

[0028] Coverage and diversity maximization function ,in, For expected execution costs;

[0029] System robustness maximization function Where γ∈[0,1] is the weighting coefficient. It is the tester's credit information Decayed expected utility, r i For tester reputation information, xi,j Assign decision variables to the task, x i,j x ∈{0,1} i,j =1 indicates task t j Assigned to tester p i x i,j =0 indicates that the task has not been assigned.

[0030] Furthermore, in step S3, the multi-objective optimization algorithm is a hybrid algorithm that combines particle swarm optimization and non-dominated sorting genetic algorithm, and its execution flow includes:

[0031] Step S31, the task allocation scheme for particles is explored using the following formula:

[0032] ;

[0033] in, Let be the velocity of particle p at time step t+1, w be the inertia weight, c1 and c2 be the acceleration factors, and r1 and r2 be ~U(0,1). This represents the historical optimal position of particle p. The optimal position for the entire population. It is the current position of particle P at time step t;

[0034] Step S32: Map the continuous vector to assigned probabilities using the Sigmoid or Softmax function, as shown in the following formula:

[0035] ,

[0036] in, Indicates task Assigned to testers The probability of allocation, Indicates task With the tester The affinity between them, where γ is a temperature parameter, C j For task t j Set of candidate test subjects This indicates that the particle is related to the task in the current iteration. Assigned to testers This choice represents a preference value / propensity level. Indicates task With the tester The affinity between them is an exponentially weighted value. This indicates that normalization is performed on all possible testers;

[0037] Step S33: Execute a non-dominated sorting genetic algorithm to obtain the Pareto front, and maintain diversity through crowding distance, as shown in the following formula:

[0038]

[0039] In the formula, , , These represent the new value, old value, global maximum value, and global minimum value of the k-th objective function, respectively.

[0040] Step S34: Perform a task exchange operation on the excellent solution and accept the new solution with non-dominant improvement.

[0041] Furthermore, in step S3, the dynamic weights satisfy... ,in, To assess the urgency of the mission, System risk level;

[0042] pass Select the final solution from Pareto solution set A. , It is the estimated value of the k-th objective function.

[0043] Furthermore, in step S5, the reputation information is dynamically updated using one of the following methods:

[0044] The exponential smoothing update method is formulated as follows: ,

[0045] Where α∈(0,1) is the smoothing coefficient. For the tester p i The credit score in round t. For the tester p i Overall performance score in round t;

[0046] The Bayesian smooth update method has the following formula:

[0047]

[0048] Where κ is the prior strength coefficient, Si is the number of times tester pi successfully completes a task within a fixed time window, and Ni is the total number of task attempts within that window.

[0049] Furthermore, in step S5, the anomaly detection algorithm is the robust Z-score algorithm, and the formula for calculating the anomaly statistics is as follows:

[0050]

[0051] Where bi is the behavioral feature vector. For robust location and scale estimation based on median, To prevent division by zero constant; if a i >τ aIf it is, then it is marked as suspicious, τ a As a suspicious threshold, if a i >τ black or >π adv This will trigger penalties, including scheduling permission restrictions, aggregation weight decay, or system blacklisting, τ black For malicious threshold, π adv As the policy threshold, Represents the behavioral characteristics of a given tester The probability that testers will exhibit adversarial or malicious behavior.

[0052] On the other hand, the present invention provides a federated multi-objective scheduling system for industrial control crowdsourcing testing, comprising:

[0053] The federated learning server, local test agent, global scheduler, and adversarial awareness and reputation management module form a collaborative closed loop through targeted data interaction links. Specifically, the federated learning server establishes a bidirectional data link with the local test agent and a reputation data interaction link with the adversarial awareness and reputation management module; the local test agent establishes a matching score reporting link with the global scheduler; and the global scheduler establishes a reputation feedback and scheduling optimization link with the adversarial awareness and reputation management module.

[0054] The federated learning server receives the privacy-enhanced model update uploaded by the local test agent, distributes the aggregated global model to the local test agent, receives the tester reputation score output by the adversarial awareness and reputation management module to dynamically adjust the aggregation weight, and synchronizes the task feature basic data required for global encoding to the global scheduler.

[0055] The local test agent receives the global model distributed by the federated learning server, performs encoding and model training based on local features, uploads the privacy-processed model update to the federated learning server, and reports the locally calculated task-tester affinity score to the global scheduler.

[0056] The global scheduler receives the affinity score reported by the local test agent, receives the tester reputation information fed back by the adversarial perception and reputation management module, executes a multi-objective optimization algorithm based on the affinity score and the tester reputation information, and synchronizes the generated optimal task allocation scheme to the tester terminal and the adversarial perception and reputation management module for subsequent behavior evaluation.

[0057] The adversarial awareness and reputation management module receives task execution behavior logs uploaded by the tester's terminal, receives the task allocation scheme synchronized by the global scheduler to associate with the evaluation object, feeds back the updated reputation score to the federated learning server and the global scheduler respectively, and synchronizes the malicious tester judgment result to the federated learning server and the global scheduler to trigger punishment.

[0058] Furthermore, the federated learning server also includes: a robust aggregation unit, a global encoder unit, and a model distribution unit;

[0059] The robust aggregation unit receives the reputation score from the adversarial awareness and reputation management module via a data link, and dynamically adjusts the tester weight w based on the reputation score. i The coordinate trimming mean operator or Krum algorithm is used to process the model update volume uploaded by the local test agent;

[0060] The global encoder unit constructs a task encoding model based on the knowledge graph of the industrial control system domain, and the generated task embedding vector is synchronized to the local test agent through the model distribution unit.

[0061] The model distribution unit standardizes the robustly aggregated global model and distributes it to each local test agent through a bidirectional link.

[0062] Furthermore, the local test agent also includes: a local encoder unit, a model training unit, and a privacy processing unit;

[0063] The local encoder unit receives the task feature basic data synchronized by the federated learning server and maps the tester's ability features and experience data into a low-dimensional embedding vector.

[0064] The model training unit trains the affinity model based on local interaction data, and the generated model update is sent to the privacy processing unit.

[0065] The privacy processing unit performs gradient pruning and differential privacy noise injection on the model update amount, and the processed update amount is uploaded to the robust aggregation unit of the federated learning server through a directional link.

[0066] Furthermore, the global scheduler also includes: a multi-objective optimization unit, a dynamic weight unit, and a scheme output unit;

[0067] The multi-objective optimization unit receives the affinity score reported by the local test agent through the link, receives the reputation information fed back by the adversarial perception and reputation management module, and executes a hybrid algorithm that combines particle swarm optimization and NSGA-II.

[0068] The dynamic weighting unit calculates the dynamic weight ω based on the task urgency u and the system risk level ρ. k The weight parameters are then fed to the multi-objective optimization unit.

[0069] The scheme output unit synchronizes the optimal task allocation scheme to the tester's terminal, the adversarial perception and reputation management module, and the federated learning server.

[0070] Furthermore, the adversarial perception and reputation management module also includes: a reputation update unit, an anomaly detection unit, and a result feedback unit;

[0071] The reputation update unit receives the tester's behavior log and the task allocation scheme synchronized by the global scheduler, and updates the reputation score using exponential smoothing or Bayesian smoothing algorithms.

[0072] The anomaly detection unit extracts behavioral feature vectors, identifies abnormal behaviors using a robust Z-score algorithm, and generates a malicious judgment result.

[0073] The result feedback unit distributes the reputation score to the robust aggregation unit of the federated learning server and the multi-objective optimization unit of the global scheduler, and synchronizes the malicious judgment result to the above two modules to trigger the punishment mechanism.

[0074] Furthermore, the model distribution unit is a bidirectional data interaction unit, used to distribute the global model output by the robust aggregation unit to the model training units of each local test agent, and to receive the model training status data fed back by the local test agents and synchronize the model training status data to the federated learning server to assist in subsequent aggregation strategy adjustments.

[0075] Furthermore, the multi-objective optimization unit also includes a Pareto solution processing unit, used to receive ω output by the dynamic weight unit. k The congestion distance is calculated on the Pareto front obtained by non-dominated sorting, and the optimal task allocation scheme is selected by combining weights. The scheme is then sent to the scheme output unit of the global scheduler. At the same time, the matching degree correlation data in the solution processing process is fed back to the adversarial perception and reputation management module to assist in the optimization of reputation assessment accuracy.

[0076] Compared with existing technologies, the beneficial effects of this invention are that, through the architecture of local feature encoding and federated model training, this invention achieves local retention of sensitive data such as tester capabilities and experience, and only uploads the model update after privacy enhancement processing (i.e., gradient pruning and differential privacy noise injection) to the federated learning server, thus avoiding the risk of leakage from centralized data storage from the source. It fully complies with privacy regulations such as GDPR. Compared with traditional privacy technologies such as homomorphic encryption, this invention significantly reduces computation and communication overhead. Combined with the design of local matching degree calculation, it greatly improves the task matching response speed and adapts to the real-time requirements of industrial control system testing.

[0077] Furthermore, this invention addresses the multi-objective conflict problem in industrial control system testing by constructing a four-objective model encompassing utility, cost, coverage, and robustness. By integrating a hybrid algorithm combining PSO global search and NSGA-II non-dominated sorting, it effectively compensates for the shortcomings of single algorithms. PSO accelerates algorithm convergence, while NSGA-II guarantees the diversity of solutions, improving the search efficiency for Pareto optimal solutions in high-dimensional discrete spaces. Combined with a dynamic weighting mechanism for task urgency and system risk, the scheduling scheme can be flexibly adjusted according to actual needs such as prioritizing time compression for urgent tasks and enhancing robustness in high-risk scenarios, thereby improving its adaptability.

[0078] Furthermore, the dynamic reputation and adversarial perception system designed in this invention forms a dual defense. The robust Z-score algorithm can accurately identify malicious behaviors such as false reports and high-frequency abnormal submissions, improving the recognition accuracy. The reputation update mechanism adopts exponential smoothing or Bayesian smoothing, so that the reputation score can reflect the tester's performance in real time. The score is synchronously fed back to the robust aggregation and global scheduling of the federated learning server. It resists attacks from both the model training and task allocation stages, improving the vulnerability discovery rate and system robustness.

[0079] Furthermore, this invention maps tester features and task features to the same semantic space by minimizing the objective function and model learning, and predicts their matching affinity. This upgrades the matching process from simple rule judgment to data-driven intelligent prediction, significantly improving the accuracy and personalization of task allocation. All training data is stored locally on the terminal, with only model update data uploaded. Combined with gradient pruning and differential privacy noise injection, strict privacy protection is achieved, effectively preventing the reverse engineering of original data from model updates. This meets the stringent requirements for tester privacy and data security in industrial control crowdsourcing scenarios.

[0080] Furthermore, the four-objective model constructed in this invention systematically covers the commercial, efficiency, quality, and safety dimensions of industrial control crowdsourcing testing, avoiding the one-sidedness caused by traditional single-objective optimization, ensuring the optimal comprehensive performance of the scheduling scheme. PSO provides efficient global search and fast convergence capabilities, while NSGA-II ensures the diversity of solutions and the wide distribution of Pareto fronts through non-dominated sorting and crowding distance calculation. The hybrid algorithm that integrates PSO and NSGA-II effectively combines the advantages of both. PSO makes up for the slow convergence of NSGA-II, while NSGA-II overcomes the shortcomings of PSO, such as being prone to getting trapped in local optima and having difficulty handling multiple objectives.

[0081] Furthermore, the exponential smoothing update in this invention uses a smoothing coefficient to weight and fuse historical reputation with current performance. This ensures that the reputation score will not change abruptly due to accidental fluctuations in performance of a single task, giving the reputation system good inertia and stability. The Bayesian smoothing update introduces a solid Bayesian statistical foundation. For newly added testers, the prior terms in the formula play a role in smoothing the start-up, avoiding extreme reputation scores due to a small number of initial tasks and a high degree of randomness in the success rate, thus ensuring a fair starting opportunity for new testers. Attached Figure Description

[0082] Figure 1 This is a flowchart of a federated multi-objective scheduling method used for industrial control crowdsourcing testing;

[0083] Figure 2 This is a block diagram of a federated multi-objective scheduling system used for industrial control crowdsourcing testing. Detailed Implementation

[0084] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0085] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0086] like Figure 1 As shown, Figure 1 This is a flowchart of a federated multi-objective scheduling method for industrial control crowdsourcing testing.

[0087] The technical solution provided in this application includes the following steps:

[0088] Step S1: The task publisher uploads the industrial control system test task description information. The scheduling system constructs a knowledge graph based on the domain knowledge of the industrial control system. The task description information is converted into a task feature vector through the encoding module. The local agent deployed on the tester's terminal is converted into a tester feature vector through the local encoding module.

[0089] In step S2, each local agent trains a task-tester matching model using local data, generates model update information, and uploads it to the federated learning server after privacy enhancement processing. The federated learning server processes the model update information through an aggregation strategy, generates a global model, and distributes it to each local agent. Each local agent, based on the global model, task feature vector, and tester feature vector, outputs a matching score in the local terminal environment through similarity calculation or prediction model and reports it to the scheduling center.

[0090] Step S3: The scheduling center uses the matching score as the core input to construct a multi-objective scheduling model, uses a multi-objective optimization algorithm to solve for the optimal solution set, determines dynamic weights based on task urgency and system risk, and selects and outputs task allocation schemes from the optimal solution set.

[0091] Step S4: The tester performs the test task according to the task allocation scheme and uploads a behavior log containing operation data, environmental information and test results.

[0092] In step S5, the adversarial perception module evaluates the tester's performance based on the behavior log, dynamically updates the reputation information, identifies malicious testers through an anomaly detection algorithm and triggers a punishment mechanism, and the reputation information is synchronously fed back to the model aggregation in step S2 and the scheduling process in step S3.

[0093] Step S6: Use the test results, feedback data, and updated reputation information as training data for the next round of federated learning, and iteratively optimize the global model and the local model.

[0094] Specifically, in this embodiment, the industrial control system is a general term for various systems used to monitor and control industrial production processes (such as power, water conservancy, petrochemical, and manufacturing industries), including monitoring and data acquisition systems, distributed control systems, and programmable logic controllers. It achieves automated management of industrial processes by collecting field data and executing control logic. It is a core component of modern industrial critical infrastructure, and its security is directly related to the stability of industrial production and public safety.

[0095] Specifically, this embodiment employs crowdsourcing testing, a testing model that distributes software or system testing tasks to a large number of volunteers (testers) via the internet. Testers come from diverse backgrounds and possess a variety of equipment, environments, and skills, enabling them to conduct multi-dimensional testing of products in real-world scenarios. This allows them to discover vulnerabilities and compatibility issues that are difficult for traditional testing teams to cover. In the field of industrial control systems, crowdsourcing testing can integrate global resources to conduct extensive security and functional testing on various devices, protocols, and applications of industrial control systems, improving the efficiency and comprehensiveness of vulnerability discovery.

[0096] Specifically, in this embodiment, federated learning is a distributed machine learning paradigm that allows multiple participants to jointly train a machine learning model without sharing the original data. Each participant trains the model locally on the data and only uploads the model updates (such as gradients and parameters) to the server for aggregation, thereby generating a global model. This approach effectively resolves the contradiction between data privacy protection and cross-device and cross-organizational joint modeling. In crowdsourced testing of industrial control systems, it can be used to achieve efficient matching of tasks and testers for model training while protecting the privacy of testers (such as personal skill data and test history).

[0097] Specifically, in this embodiment, multi-objective optimization refers to the process of finding Pareto optimal solutions when multiple conflicting or non-substitutable objective functions exist simultaneously. In the allocation of crowdsourced testing tasks for industrial control systems, multiple objectives such as task completion time, vulnerability discovery rate, resource cost, and system robustness need to be considered simultaneously. These objectives often cannot be optimally achieved at the same time. Multi-objective optimization algorithms (such as NSGA-II, MOEA-D, etc.) can weigh these objectives and find a series of optimal task allocation schemes.

[0098] Specifically, in this embodiment, privacy protection refers to preventing the unauthorized access, use, disclosure, or tampering of sensitive information of individuals or organizations (such as personal identity, data, behavioral patterns, etc.) through technical, legal, or management means. In the scenario of crowdsourcing testing of industrial control systems, the personal information, skill data, testing history, etc. of testers are sensitive information. Privacy protection technology can ensure that this data is not disclosed during task allocation, model training, etc., while meeting the requirements of data privacy regulations (such as GDPR).

[0099] Specifically, in this embodiment, adversarial robustness refers to the ability of a system, model, or algorithm to maintain stable performance and normal function when faced with malicious attacks. In crowdsourced testing of industrial control systems, there may be malicious testers submitting false reports, manipulating reputation scores, or launching data poisoning attacks. Systems with adversarial robustness can effectively detect and resist these attacks, ensuring the reliability of test results and the security of the system.

[0100] This invention utilizes a local feature encoding and federated model training architecture to achieve local retention of sensitive data such as tester capabilities and experience. Only the model updates after privacy enhancement processing (i.e., gradient pruning and differential privacy noise injection) are uploaded to the federated learning server, thus avoiding the risk of data leakage from centralized data storage at the source. This fully complies with privacy regulations such as GDPR. Compared with traditional privacy technologies such as homomorphic encryption, this invention significantly reduces computational and communication overhead. Combined with the design of local matching degree calculation, it greatly improves the task matching response speed and adapts to the real-time requirements of industrial control system testing.

[0101] This invention addresses the multi-objective conflict problem in industrial control system testing by constructing a four-objective model encompassing utility, cost, coverage, and robustness. By integrating a hybrid algorithm combining PSO global search and NSGA-II non-dominated sorting, it effectively compensates for the shortcomings of single algorithms. PSO accelerates algorithm convergence, while NSGA-II guarantees solution diversity, improving the search efficiency for Pareto optimal solutions in high-dimensional discrete spaces. Furthermore, by incorporating a dynamic weighting mechanism based on task urgency and system risk, the algorithm can be flexibly adjusted according to actual needs such as prioritizing urgent tasks to compress time and enhancing robustness in high-risk scenarios, thereby improving the adaptability of the scheduling scheme.

[0102] The dynamic reputation and adversarial perception system designed in this invention forms a dual defense. The robust Z-score algorithm can accurately identify malicious behaviors such as false reports and high-frequency abnormal submissions, improving the recognition accuracy. The reputation update mechanism adopts exponential smoothing or Bayesian smoothing so that the reputation score can reflect the tester's performance in real time. The score is synchronously fed back to the robust aggregation and global scheduling of the federated learning server. It resists attacks from both the model training and task allocation stages, improving the vulnerability discovery rate and system robustness.

[0103] Specifically, in step S1, the encoding module includes a global task encoder and a local encoder;

[0104] The knowledge graph G=(V,R) contains nodes V of industrial control system equipment type, communication protocol, vulnerability category, attack mode and test action, and edges R contain semantic associations of effects, requirements and is-variant-of.

[0105] The task feature vector is a task embedding vector, satisfying... For global task encoder, For task t j Structured descriptive information;

[0106] The tester feature vector is the tester embedding vector, which satisfies... ,in, For local encoders, For testers A set of capability characteristics and experience data.

[0107] Specifically, in step S2, the task-tester matching model is a task-tester affinity model, and local training is performed by minimizing the objective function. To achieve, among which, The loss function is either logistic or similarity regression. For local interactive datasets, This is the reward value for local observations;

[0108] The privacy enhancement process includes: gradient pruning and differential privacy noise injection;

[0109] The federated learning server is a federated matching learning (FML) server.

[0110] Specifically, in step S2, the aggregation strategy is a robust aggregation operator, specifically a coordinate trimming mean operator or the Krum algorithm, and the aggregation process satisfies:

[0111] ;

[0112] in, This represents the aggregated result of the global model in round t+1; w i Weights for testers; This refers to the amount of local model updates. Indicates robust aggregation operation; It is a set containing N elements, representing weighted model updates from N clients; the tester weight w i The system is dynamically adjusted based on the reputation information from step S5.

[0113] This invention maps tester features and task features to the same semantic space by minimizing the objective function and model learning, and predicts their matching affinity. This upgrades the matching process from simple rule judgment to data-driven intelligent prediction, significantly improving the accuracy and personalization of task allocation. All training data is stored locally on the terminal, and only the model update is uploaded. Combined with gradient pruning and differential privacy noise injection, strict privacy protection is achieved, effectively preventing the original data from being inferred from the model update. This meets the stringent requirements for tester privacy and data security in industrial control crowdsourcing scenarios.

[0114] Specifically, in step S3, the multi-objective scheduling model includes four objective functions, specifically:

[0115] Total utility maximization function ,in, The expected utility value for the task and the tester;

[0116] Total cost / time minimization function ,in, The matching score, i.e. the affinity score, is for step S2;

[0117] Coverage and diversity maximization function ,in, For expected execution costs;

[0118] System robustness maximization function Where γ∈[0,1] is the weighting coefficient. It is the tester's credit information Decayed expected utility, r i For tester reputation information, x i,j Assign decision variables to the task, x i,j x ∈{0,1} i,j =1 indicates task t j Assigned to tester p i x i,j =0 indicates that the task has not been assigned.

[0119] Specifically, in this embodiment, the affinity score represents the expected utility of the tester in handling the task. The optimization direction is essentially to maximize it. Mathematically, minimizing the negative affinity score is equivalent to maximizing the affinity score.

[0120] Specifically, in step S3, the multi-objective optimization algorithm is a hybrid algorithm combining particle swarm optimization (PSO) and non-dominated sorting genetic algorithm (NSGA-II), and the execution flow includes:

[0121] Step S31, the task allocation scheme for particles is explored using the following formula:

[0122] ;

[0123] in, Let be the velocity of particle p at time step t+1, w be the inertia weight, c1 and c2 be the acceleration factors, and r1 and r2 be ~U(0,1). This represents the historical optimal position of particle p. The optimal position for the entire population. It is the current position of particle P at time step t;

[0124] Step S32: Map the continuous vector to assigned probabilities using the Sigmoid or Softmax function, as shown in the following formula:

[0125] ,

[0126] in, Indicates task Assigned to testers The probability of allocation, Indicates task With the tester The affinity between them, where γ is a temperature parameter, C j For task t j Set of candidate test subjects This indicates that the particle is related to the task in the current iteration. Assigned to testers This choice's preference value / propensity magnitude Indicates task With the tester The affinity between them is an exponentially weighted value. This indicates that normalization is performed on all possible testers;

[0127] Step S33: Execute a non-dominated sorting genetic algorithm to obtain the Pareto front, and maintain diversity through crowding distance, as shown in the following formula:

[0128]

[0129] In the formula, , , These represent the new value, old value, global maximum value, and global minimum value of the k-th objective function, respectively.

[0130] Step S34: Perform a task exchange operation on the excellent solution and accept the new solution with non-dominant improvement.

[0131] Specifically, in step S3, the dynamic weights satisfy... ,in, To assess the urgency of the mission, System risk level;

[0132] pass Select the final solution from Pareto solution set A. , It is the estimated value of the k-th objective function.

[0133] The four-objective model constructed in this invention systematically covers the commercial, efficiency, quality, and safety dimensions of industrial control crowdsourcing testing, avoiding the one-sidedness caused by traditional single-objective optimization, ensuring the optimal comprehensive performance of the scheduling scheme. PSO provides efficient global search and fast convergence capabilities, while NSGA-II ensures the diversity of solutions and the wide distribution of Pareto fronts through non-dominated sorting and crowding distance calculation. The hybrid algorithm that integrates PSO and NSGA-II effectively combines the advantages of both. PSO makes up for the slow convergence of NSGA-II, while NSGA-II overcomes the shortcomings of PSO, such as being prone to getting trapped in local optima and having difficulty handling multiple objectives.

[0134] Specifically, in step S5, the reputation information is dynamically updated using one of the following methods:

[0135] The exponential smoothing update method is formulated as follows: ,

[0136] Where α∈(0,1) is the smoothing coefficient. For the tester p i The credit score in round t. For the tester pi Overall performance score in round t;

[0137] The Bayesian smooth update method has the following formula:

[0138]

[0139] Where κ is the prior strength coefficient, Si is the number of times tester pi successfully completes a task within a fixed time window, and Ni is the total number of task attempts within that window.

[0140] Specifically, in step S5, the anomaly detection algorithm is the robust Z-score algorithm, and the formula for calculating the anomaly statistics is as follows:

[0141]

[0142] Where bi is the behavioral feature vector. For robust location and scale estimation based on median, To prevent division by zero constant; if a i >τ a If it is, then it is marked as suspicious, τ a As a suspicious threshold, if a i >τ black or >π adv This will trigger penalties, including scheduling permission restrictions, aggregation weight decay, or system blacklisting, τ black For malicious threshold, π adv As the policy threshold, Represents the behavioral characteristics of a given tester The probability that testers will exhibit adversarial or malicious behavior.

[0143] Specifically, in this embodiment, a simulation environment is built on the MATLAB R2021a platform, and the scenario includes... A distributed tester and Each industrial control task has tester features including: multi-dimensional skill vector, dynamic reputation score, and processing speed factor. Each task also has: task requirement vector, task urgency, and environment type. Experimental data is generated based on a Gaussian distribution in a normalized latent feature space. The feature vectors of tasks and testers are randomly generated using a Gaussian distribution and mapped to a shared semantic embedding space. The true matching degree between tasks and testers is obtained by adding Gaussian noise to the inner product similarity and then normalizing it using the Sigmoid function. The true matching degree between tasks and testers is also obtained by linearly combining skill similarity, reputation score, and noise value, then transforming it using the Sigmoid function. The experiment compares three typical scheduling methods: Greedy (greedy algorithm): assigns tasks to the tester with the highest current matching degree, emphasizing immediate optimality; Random (random allocation): serves as a baseline method, simulating a "non-intelligent strategy"; and FMS-AA (the method proposed in this paper): the federated multi-objective scheduling framework proposed in this invention, which integrates federated matching learning and adversarial awareness. (with Adversarial-Awareness) In the multi-objective scheduling phase, a hybrid MOPSO+NSGA-II algorithm is adopted to comprehensively optimize three indicators: task completion time (corresponding to the time and cost objectives in the technical solution), vulnerability discovery rate (corresponding to the utility objective), and system robustness (corresponding to the reputation and robustness objectives). Repeated tests were conducted at different proportions of malicious testers (0.05, 0.10, 0.20, 0.30), and the average performance of each method on the three indicators of task completion time, vulnerability discovery rate, and system robustness was recorded.

[0144]

[0145] Table 1 shows the experimental results and analysis: when the proportion of malicious testers is 5%, the three methods perform relatively stably. The method of this invention has the shortest task completion time (63.077), but it is slightly lower than the Greedy method (0.791, 0.674) in vulnerability discovery rate (0.781) and system robustness (0.669). This indicates that under low malicious interference, the traditional greedy strategy still has certain advantages in performance indicators, while the present invention has a significant improvement in time efficiency.

[0146] When the malicious proportion rises to 10%, the task completion time of Greedy and Random increases, while the task completion time of the present invention remains low (63.171). In terms of vulnerability discovery rate, the present invention (0.781) is slightly lower than Greedy (0.798) and Random (0.804). In terms of robustness, the present invention (0.708) is close to Random (0.714), indicating that under moderate malicious interference, the present invention maintains stable time efficiency, while being comparable to traditional methods in terms of discovery rate and robustness.

[0147] When the malicious proportion was further increased to 20%, Greedy's task completion time increased significantly to 117.537, and Random's completion time also fluctuated (86.173). The present invention is superior to or close to the comparison method in terms of task completion time (65.742), vulnerability discovery rate (0.808), and system robustness (0.721), showing better overall stability.

[0148] At an extreme malicious ratio (30%), the method of this invention performs best in terms of task completion time (60.151), vulnerability discovery rate (0.815), and system robustness (0.781). In particular, it is significantly better than Greedy and Random methods in terms of time efficiency and robustness, demonstrating its adaptability to high-intensity malicious interference.

[0149] The experimental results show that as the proportion of malicious testers increases, the traditional Greedy and Random methods fluctuate significantly in task completion time, while the method of this invention maintains a consistently low and stable time efficiency. In terms of vulnerability discovery rate and system robustness, this invention outperforms or approaches traditional methods under medium to high malicious proportions, and its comprehensive advantages are even more obvious when the malicious proportion is high. This indicates that the federated multi-objective scheduling framework with fusion of adversarial awareness proposed in this paper can effectively adapt to the crowdsourcing testing environment of industrial control systems with large-scale, heterogeneous characteristics and malicious interference, and is suitable for actual industrial systems with high privacy protection requirements.

[0150] In this invention, exponential smoothing update uses a smoothing coefficient to weight and fuse historical reputation with current performance. This ensures that the reputation score will not change abruptly due to accidental fluctuations in performance of a single task, giving the reputation system good inertia and stability. Bayesian smoothing update introduces a solid Bayesian statistical foundation. For newly added testers, the prior terms in the formula play a role in smoothing the start, avoiding extreme reputation scores due to a small number of initial tasks and a high degree of randomness in the success rate, thus ensuring a fair starting opportunity for new testers.

[0151] Please see Figure 2 As shown, it is a structural block diagram of a federated multi-objective scheduling system used for industrial control crowdsourcing testing.

[0152] The federated multi-objective scheduling system for industrial control crowdsourcing testing provided in this embodiment includes:

[0153] The Federated Learning Server (FML Server), local test agent, global scheduler, and adversarial awareness and reputation management module form a collaborative closed loop through targeted data interaction links. Specifically, the Federated Learning Server establishes a bidirectional data link with the local test agent and a reputation data interaction link with the adversarial awareness and reputation management module; the local test agent establishes a matching score reporting link with the global scheduler; and the global scheduler establishes a reputation feedback and scheduling optimization link with the adversarial awareness and reputation management module.

[0154] The federated learning server receives the privacy-enhanced model update uploaded by the local test agent, distributes the aggregated global model to the local test agent, receives the tester reputation score output by the adversarial awareness and reputation management module to dynamically adjust the aggregation weight, and synchronizes the task feature basic data required for global encoding to the global scheduler.

[0155] The local test agent receives the global model distributed by the federated learning server, performs encoding and model training based on local features, uploads the privacy-processed model update to the federated learning server, and reports the locally calculated task-tester affinity score to the global scheduler.

[0156] The global scheduler receives the affinity score reported by the local test agent, receives the tester reputation information fed back by the adversarial perception and reputation management module, executes a multi-objective optimization algorithm based on the affinity score and the tester reputation information, and synchronizes the generated optimal task allocation scheme to the tester terminal and the adversarial perception and reputation management module for subsequent behavior evaluation.

[0157] The adversarial awareness and reputation management module receives task execution behavior logs uploaded by the tester's terminal, receives the task allocation scheme synchronized by the global scheduler to associate with the evaluation object, feeds back the updated reputation score to the federated learning server and the global scheduler respectively, and synchronizes the malicious tester judgment result to the federated learning server and the global scheduler to trigger punishment.

[0158] Specifically, the federated learning server also includes: a robust aggregation unit, a global encoder unit, and a model distribution unit;

[0159] The robust aggregation unit receives the reputation score from the adversarial awareness and reputation management module via a data link, and dynamically adjusts the tester weight w based on the reputation score. i The coordinate trimming mean operator or Krum algorithm is used to process the model update volume uploaded by the local test agent;

[0160] The global encoder unit constructs a task encoding model based on the knowledge graph of the industrial control system domain, and the generated task embedding vector is synchronized to the local test agent through the model distribution unit.

[0161] The model distribution unit standardizes the robustly aggregated global model and distributes it to each local test agent through a bidirectional link.

[0162] Specifically, the local test agent also includes: a local encoder unit, a model training unit, and a privacy processing unit;

[0163] The local encoder unit receives the task feature basic data synchronized by the federated learning server and maps the tester's ability features and experience data into a low-dimensional embedding vector.

[0164] The model training unit trains the affinity model based on local interaction data, and the generated model update is sent to the privacy processing unit.

[0165] The privacy processing unit performs gradient pruning and differential privacy noise injection on the model update amount, and the processed update amount is uploaded to the robust aggregation unit of the federated learning server through a directional link.

[0166] Specifically, the global scheduler further includes: a multi-objective optimization unit, a dynamic weight unit, and a scheme output unit;

[0167] The multi-objective optimization unit receives the affinity score reported by the local test agent through the link, receives the reputation information fed back by the adversarial perception and reputation management module, and executes a hybrid algorithm that combines particle swarm optimization and NSGA-II.

[0168] The dynamic weighting unit calculates the dynamic weight ω based on the task urgency u and the system risk level ρ. k The weight parameters are then fed to the multi-objective optimization unit.

[0169] The scheme output unit synchronizes the optimal task allocation scheme to the tester's terminal, the adversarial perception and reputation management module, and the federated learning server.

[0170] Specifically, the adversarial perception and reputation management module also includes: a reputation update unit, an anomaly detection unit, and a result feedback unit;

[0171] The reputation update unit receives the tester's behavior log and the task allocation scheme synchronized by the global scheduler, and updates the reputation score using exponential smoothing or Bayesian smoothing algorithms.

[0172] The anomaly detection unit extracts behavioral feature vectors, identifies abnormal behaviors using a robust Z-score algorithm, and generates a malicious judgment result.

[0173] The result feedback unit distributes the reputation score to the robust aggregation unit of the federated learning server and the multi-objective optimization unit of the global scheduler, and synchronizes the malicious judgment result to the above two modules to trigger the punishment mechanism.

[0174] Specifically, the model distribution unit is a bidirectional data interaction unit, used to distribute the global model output by the robust aggregation unit to the model training units of each local test agent, and to receive the model training status data fed back by the local test agents and synchronize the model training status data to the federated learning server to assist in subsequent aggregation strategy adjustments.

[0175] Specifically, the multi-objective optimization unit further includes a Pareto solution processing unit, used to receive ω output by the dynamic weight unit. k The congestion distance is calculated on the Pareto front obtained by non-dominated sorting, and the optimal task allocation scheme is selected by combining weights. The scheme is then sent to the scheme output unit of the global scheduler. At the same time, the matching degree correlation data in the solution processing process is fed back to the adversarial perception and reputation management module to assist in the optimization of reputation assessment accuracy.

[0176] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this application.

Claims

1. A federated multi-objective scheduling method for industrial control crowdsourcing testing, characterized in that, include: Step S1: The task publisher uploads the industrial control system test task description information. The scheduling system constructs a knowledge graph based on the domain knowledge of the industrial control system. The task description information is converted into a task feature vector through the encoding module. The local agent deployed on the tester's terminal is converted into a tester feature vector through the local encoding module. In step S2, each local agent trains a task-tester matching model using local data, generates model update information, and uploads it to the federated learning server after privacy enhancement processing. The federated learning server processes the model update information through an aggregation strategy, generates a global model, and distributes it to each local agent. Each local agent, based on the global model, task feature vector, and tester feature vector, outputs a matching score in the local terminal environment through similarity calculation or prediction model and reports it to the scheduling center. Step S3: The scheduling center uses the matching score as the core input to construct a multi-objective scheduling model, uses a multi-objective optimization algorithm to solve for the optimal solution set, determines dynamic weights based on task urgency and system risk, and selects and outputs task allocation schemes from the optimal solution set. Step S4: The tester performs the test task according to the task allocation scheme and uploads a behavior log containing operation data, environmental information and test results. In step S5, the adversarial perception module evaluates the tester's performance based on the behavior log, dynamically updates the reputation information, identifies malicious testers through an anomaly detection algorithm and triggers a punishment mechanism, and the reputation information is synchronously fed back to the model aggregation in step S2 and the scheduling process in step S3. Step S6: Use the test results, feedback data, and updated reputation information as training data for the next round of federated learning, and iteratively optimize the global model and the local model.

2. The method according to claim 1, characterized in that, In step S1, the encoding module includes a global task encoder and a local encoder; The knowledge graph G=(V,R) contains nodes V of industrial control system equipment type, communication protocol, vulnerability category, attack mode and test action, and edges R contain semantic associations of effects, requirements and is-variant-of. The task feature vector is a task embedding vector, satisfying... For global task encoder, For task t j Structured descriptive information; The tester feature vector is the tester embedding vector, which satisfies... ,in, For local encoders, For testers A set of capability characteristics and experience data.

3. The method according to claim 1, characterized in that, In step S2, the task-tester matching model is a task-tester affinity model, and local training is performed by minimizing the objective function. To achieve, among which, The loss function is either logistic or similarity regression. For local interactive datasets, This is the reward value for local observations; The privacy enhancement process includes: gradient pruning and differential privacy noise injection; The federated learning server is a federated matching learning (FML) server.

4. The method according to claim 3, characterized in that, In step S2, the aggregation strategy is a robust aggregation operator, specifically a coordinate trimming mean operator or the Krum algorithm, and the aggregation process satisfies: ； in, This represents the aggregated result of the global model in round t+1; w i Weights for testers; This refers to the amount of local model updates. Indicates robust aggregation operation; It is a set containing N elements, representing weighted model updates from N clients; the tester weight w i The system is dynamically adjusted based on the reputation information from step S5.

5. The method according to claim 3, characterized in that, In step S3, the multi-objective scheduling model includes four objective functions, specifically: Total utility maximization function ,in, The expected utility value for the task and the tester; Total cost / time minimization function ,in, The matching score, i.e. the affinity score, is for step S2; Coverage and diversity maximization function ,in, For expected execution costs; System robustness maximization function Where γ∈[0,1] is the weighting coefficient. It is the tester's credit information Decayed expected utility, r i For tester reputation information, x i,j Assign decision variables to the task, x i,j x ∈{0,1} i,j =1 indicates task t j Assigned to tester p i x i,j =0 indicates that the task has not been assigned.

6. The method according to claim 1, characterized in that, In step S3, the multi-objective optimization algorithm is a hybrid algorithm that combines particle swarm optimization and non-dominated sorting genetic algorithm. The execution process includes: Step S31, the task allocation scheme for particles is explored using the following formula: ； in, Let be the velocity of particle p at time step t+1, w be the inertia weight, c1 and c2 be the acceleration factors, and r1 and r2 be ~U(0,1). This represents the historical optimal position of particle p. The optimal position for the global population. It is the current position of particle P at time step t; Step S32: Map the continuous vector to assigned probabilities using the Sigmoid or Softmax function, as shown in the following formula: ， in, Indicates task Assigned to testers The probability of allocation, Indicates task With the tester The affinity vector between them, where γ is the temperature parameter, C j For task t j Set of candidate test subjects This indicates that the particle is related to the task in the current iteration. Assigned to testers This choice represents a preference value / propensity level. Indicates task With the tester The affinity between them is an exponentially weighted value. This indicates that normalization is performed on all possible testers; Step S33: Execute a non-dominated sorting genetic algorithm to obtain the Pareto front, and maintain diversity through crowding distance, as shown in the following formula: ； In the formula, , , These represent the new value, old value, global maximum value, and global minimum value of the k-th objective function, respectively. Step S34: Perform a task exchange operation on the excellent solution and accept the new solution with non-dominant improvement.

7. The method according to claim 1, characterized in that, In step S3, the dynamic weights satisfy... ,in, To assess the urgency of the mission, System risk level; pass Select the final solution X* from the Pareto solution set A. It is the estimated value of the k-th objective function.

8. The method according to claim 1, characterized in that, In step S5, the reputation information is dynamically updated using one of the following methods: The exponential smoothing update method is expressed by the following formula: , Where α∈(0,1) is the smoothing coefficient. For the tester p i The credit score in round t. For the tester p i Overall performance score in round t; The Bayesian smooth update method has the following formula: ； Where κ is the prior strength coefficient, Si is the number of times tester pi successfully completes a task within a fixed time window, and Ni is the total number of task attempts within that window.

9. The method according to claim 1, characterized in that, In step S5, the anomaly detection algorithm is the robust Z-score algorithm, and the formula for calculating the anomaly statistics is as follows: ； Where bi is the behavioral feature vector. For robust location and scale estimation based on median, To prevent division by zero constant; if a i >τ a If it is, then it is marked as suspicious, τ a As a suspicious threshold, if a i >τ black or >π adv This will trigger penalties, including scheduling permission restrictions, aggregation weight decay, or system blacklisting, τ black For malicious threshold, π adv As the policy threshold, Represents the behavioral characteristics of a given tester The probability that testers will exhibit adversarial or malicious behavior.

10. A federated multi-objective scheduling system for industrial control crowdsourcing testing, characterized in that, include: The federated learning server, local test agent, global scheduler, and adversarial awareness and reputation management module form a collaborative closed loop through targeted data interaction links. Specifically, the federated learning server establishes a bidirectional data link with the local test agent and a reputation data interaction link with the adversarial awareness and reputation management module; the local test agent establishes a matching score reporting link with the global scheduler; and the global scheduler establishes a reputation feedback and scheduling optimization link with the adversarial awareness and reputation management module. The federated learning server receives the privacy-enhanced model update uploaded by the local test agent, distributes the aggregated global model to the local test agent, receives the tester reputation score output by the adversarial awareness and reputation management module to dynamically adjust the aggregation weight, and synchronizes the task feature basic data required for global encoding to the global scheduler. The local test agent receives the global model distributed by the federated learning server, performs encoding and model training based on local features, uploads the privacy-processed model update to the federated learning server, and reports the locally calculated task-tester affinity score to the global scheduler. The global scheduler receives the affinity score reported by the local test agent, receives the tester reputation information fed back by the adversarial perception and reputation management module, executes a multi-objective optimization algorithm based on the affinity score and the tester reputation information, and synchronizes the generated optimal task allocation scheme to the tester terminal and the adversarial perception and reputation management module for subsequent behavior evaluation. The adversarial awareness and reputation management module receives task execution behavior logs uploaded by the tester's terminal, receives the task allocation scheme synchronized by the global scheduler to associate with the evaluation object, feeds back the updated reputation score to the federated learning server and the global scheduler respectively, and synchronizes the malicious tester judgment result to the federated learning server and the global scheduler to trigger punishment.

11. The system according to claim 10, characterized in that, The federated learning server also includes: a robust aggregation unit, a global encoder unit, and a model distribution unit; The robust aggregation unit receives the reputation score from the adversarial awareness and reputation management module via a data link, and dynamically adjusts the tester weight w based on the reputation score. i The coordinate trimming mean operator or Krum algorithm is used to process the model update volume uploaded by the local test agent; The global encoder unit constructs a task encoding model based on the knowledge graph of the industrial control system domain, and the generated task embedding vector is synchronized to the local test agent through the model distribution unit. The model distribution unit standardizes the robustly aggregated global model and distributes it to each local test agent through a bidirectional link.

12. The system according to claim 10, characterized in that, The local test agent also includes: a local encoder unit, a model training unit, and a privacy processing unit; The local encoder unit receives the task feature basic data synchronized by the federated learning server and maps the tester's ability features and experience data into a low-dimensional embedding vector. The model training unit trains the affinity model based on local interaction data, and the generated model update is sent to the privacy processing unit. The privacy processing unit performs gradient pruning and differential privacy noise injection on the model update amount, and the processed update amount is uploaded to the robust aggregation unit of the federated learning server through a directional link.

13. The system according to claim 10, characterized in that, The global scheduler also includes: a multi-objective optimization unit, a dynamic weight unit, and a scheme output unit; The multi-objective optimization unit receives the affinity score reported by the local test agent through the link, receives the reputation information fed back by the adversarial perception and reputation management module, and executes a hybrid algorithm that combines particle swarm optimization and NSGA-II. The dynamic weighting unit calculates the dynamic weight ω based on the task urgency u and the system risk level ρ. k The weight parameters are then fed to the multi-objective optimization unit. The scheme output unit synchronizes the optimal task allocation scheme to the tester's terminal, the adversarial perception and reputation management module, and the federated learning server.

14. The system according to claim 10, characterized in that, The adversarial perception and reputation management module also includes: a reputation update unit, an anomaly detection unit, and a result feedback unit; The reputation update unit receives the tester's behavior log and the task allocation scheme synchronized by the global scheduler, and updates the reputation score using exponential smoothing or Bayesian smoothing algorithms. The anomaly detection unit extracts behavioral feature vectors, identifies abnormal behaviors using a robust Z-score algorithm, and generates a malicious judgment result. The result feedback unit distributes the reputation score to the robust aggregation unit of the federated learning server and the multi-objective optimization unit of the global scheduler, and synchronizes the malicious judgment result to the above two modules to trigger the punishment mechanism.

15. The system according to claim 11, characterized in that, The model distribution unit is a bidirectional data interaction unit, used to distribute the global model output by the robust aggregation unit to the model training units of each local test agent, and to receive the model training status data fed back by the local test agents and synchronize the model training status data to the federated learning server to assist in subsequent aggregation strategy adjustments.

16. The system according to claim 13, characterized in that, The multi-objective optimization unit further includes a Pareto solution processing unit for receiving ω output by the dynamic weight unit. k The congestion distance is calculated on the Pareto front obtained by non-dominated sorting, and the optimal task allocation scheme is selected by combining weights. The scheme is then sent to the scheme output unit of the global scheduler. At the same time, the matching degree correlation data in the solution processing process is fed back to the adversarial perception and reputation management module to assist in the optimization of reputation assessment accuracy.

Images