Urban communication system facing resilience evaluation and attack strategy evaluation method and system

Abstract

The application discloses a kind of urban communication system flexibility evaluation and attack strategy evaluation method and system.This method constructs current graph state and joint action, generates condition vector and joint action feasible region mask, inputs condition graph diffusion model with feasible region projection to sample future topology sample set;Based on sample set, calculate structure evolution uncertainty, service confidence index, risk lower bound and robust flexibility value, and update joint action by constraint Markov game;When meeting convergence condition or stopping condition, output node flexibility, link flexibility and attack, recovery strategy sorting.This method can reduce service capability overestimation, improve strategy evaluation reliability and recovery decision pertinence.

Description

Technical Field

[0001] This invention relates to the fields of communication network resilience assessment, network attack and defense strategy evaluation, graph data processing, and intelligent optimization control, and particularly to a method and system for assessing the resilience of urban communication systems and evaluating attack strategies. Background Technology

[0002] Urban communication systems typically include base stations, aggregation nodes, core equipment rooms, transmission links, access links, and service pairs carrying critical business operations such as government affairs, emergency response, transportation, finance, and public services. Under conditions such as natural disasters, equipment failures, partial link outages, node capacity reduction, deliberate attacks, and resource-constrained recovery, the node availability, link capacity, path reachability, and continuity of critical services in urban communication systems will dynamically change. Therefore, it is necessary to accurately assess the resilience level of urban communication systems and further unify the evaluation of attack and recovery strategies.

[0003] In existing technologies, methods for assessing the resilience of communication networks or complex networks often employ static topology analysis, single-point fault injection, average connectivity metrics, average shortest path length, or deterministic recovery simulation. While these methods can reflect the network structure state at a specific moment or in a specific scenario, they typically struggle to simultaneously express the coupling relationship between attack actions, recovery actions, and future topology evolution.

[0004] Specifically, existing methods typically treat the attack and recovery processes separately. Attack strategies focus on maximizing damage gains, while recovery strategies focus on minimizing recovery gains or costs. These two processes are not uniformly encoded into a single joint action, nor are they used to influence the generation of future topology samples through unified action constraints. Therefore, existing methods struggle to accurately characterize the future network state distribution under the combined influence of attack and recovery actions.

[0005] Meanwhile, most existing resilience assessment methods rely on deterministic indicators such as average connectivity, average path length, or local centrality, lacking explicit characterization of future topological sample distribution fluctuations and confidence calibration based on key business services with minimum path cost. This can easily overestimate the actual service capacity of urban communication systems in high-risk scenarios.

[0006] Furthermore, existing strategy evaluation methods typically rank strategies based solely on single attack damage, single recovery benefit, or simple cost ratios, failing to incorporate risk lower bounds, service confidence levels, and robustness values ​​into the attack and defense strategy update process. This results in strategy evaluations that struggle to balance service continuity, tail risks, and resource constraints.

[0007] In addition, while existing recovery solutions may involve path reconfiguration or temporary relay deployment, they are usually not limited to the affected subgraph induced by the subset of affected services. They also do not incorporate temporary relay deployment constraints, robust resilience increments, and changes in the risk lower bound into the recovery action selection process. This results in problems such as a large search scope, insufficient recovery targeting, and weak engineering feasibility.

[0008] Therefore, how to provide a technical solution that can unify joint actions, feasible domain constraints, future topology distribution generation, service confidence calibration, risk lower bound calculation, robust resilience assessment, constraint game update, and counterfactual ranking into the same closed-loop data processing chain has become an urgent technical problem to be solved in this field. Summary of the Invention

[0009] The purpose of this invention is to provide a method and system for assessing the resilience of urban communication systems and evaluating attack strategies, in order to solve the problems in the prior art such as the separation of attack actions and recovery actions, the decoupling of future topology evolution and action constraints, the tendency of average indicators to overestimate service capabilities, the failure of strategy evaluation to fully consider risk lower bounds and resource constraints, and the lack of affected subgraph constraints for recovery actions.

[0010] This invention generates a future topology sample set that is consistent with the joint actions and satisfies the feasible region constraints by encoding joint actions into conditional vectors and a feasible region mask, and inputting them into a conditional graph diffusion model with feasible region projection. Furthermore, based on the future topology sample set, it calculates structural evolution uncertainty, service confidence index, risk lower bound, and robustness value, and feeds the robustness value and risk lower bound back into a constrained Markov game composed of attacking and recovering agents to update the joint actions, thus forming a closed-loop coupled data processing chain. This improves the reliability, conservatism, interpretability, and engineering adaptability of urban communication system resilience assessment and attack / recovery strategy evaluation.

[0011] To achieve the above objectives, this invention provides a method for assessing the resilience of urban communication systems and evaluating attack strategies, comprising: Build Time Current graph state and generate time joint action The current graph state This includes a city communication map, a set of node attributes, a set of link attributes, and a set of key service pairs; the joint action... It includes attack sub-actions and recovery sub-actions, wherein the recovery sub-actions include local bypass reconstruction and temporary relay deployment; By the aforementioned joint action Generation time condition vector and time Joint action feasible domain mask ; The current graph state Conditional vector and joint action feasible domain mask The input conditional graph diffusion model with feasible region projection is used to make the inverse denoising output subject to the joint action feasible region mask. Projection constraints are used to obtain the joint action. A consistent set of future topological samples; The time is calculated from the future topology sample set. Structural evolution uncertainty Service confidence index Risk lower bound and robustness value ; The robustness value and risk lower bound Input a constrained Markov game consisting of an attacking agent and a recovering agent to update the joint action. and the updated joint action As the joint action input for the next closed-loop round, it is used to regenerate the condition vector for the next closed-loop round. and joint action feasible domain mask ; When the closed-loop update meets the preset convergence condition and the number of closed-loop update rounds reaches the maximum number of closed-loop rounds. Alternatively, if a preset stopping condition is met, the data transmission update is stopped, and the robustness value obtained when the data transmission update is stopped is used as the basis for the update. The counterfactual marginal contribution outputs node resilience, link resilience, attack strategy ranking, and recovery strategy ranking.

[0012] This invention also provides a system for assessing the resilience of urban communication systems and evaluating attack strategies, comprising: The graph state and constraint encoding module is used to construct the current graph state. and joint actions And by the joint action Generate condition vector and joint action feasible domain mask ; The constraint diffusion sampling module is used to sample the current graph state. Conditional vector and joint action feasible domain mask The input conditional graph diffusion model with feasible region projection is used to make the inverse denoising output subject to the joint action feasible region mask. Projection constraints are applied to obtain a future topology sample set. The lower confidence resilience calculation module is used to calculate the structural evolution uncertainty from the future topological sample set. Service confidence index Risk lower bound and robustness value ; The game-theoretic closed-loop optimization module is used to optimize the robustness value. and risk lower bound Inputting a constrained Markov game between the attacking agent and the recovering agent, updating the joint action. and the updated joint action The data is fed back to the graph state and constraint encoding module, enabling the graph state and constraint encoding module to base its actions on the updated joint actions. Regenerate condition vector and joint action feasible domain mask Then regenerate the condition vector and joint action feasible domain mask Transmitted to the constrained diffusion sampling module; The evaluation output module is used to evaluate the robustness value obtained after meeting the convergence or stopping conditions. The counterfactual robustness value calculated from the future topology sample set is used to perform counterfactual marginal contribution calculation, and output node resilience, link resilience, attack strategy ranking, and recovery strategy ranking.

[0013] Compared with the prior art, the present invention has at least the following beneficial effects:

[0014] First, this invention combines attack sub-actions and recovery sub-actions into a unified action. and further through joint actions Generate condition vector and joint action feasible domain mask This allows the attack location, recovery location, local bypass reconstruction index, and temporary relay deployment index to be expressed under the same data structure, avoiding the separation of the attack process and the recovery process.

[0015] Second, the present invention will determine the current diagram state. Conditional vector and joint action feasible domain mask The common input is a conditional graph diffusion model with feasible region projection, and the joint action feasible region mask is used in the inverse denoising process. Projection constraints are applied to the locations of nodes, links, path switching, and temporary relay deployments to ensure that the future topology sample set is consistent with the joint actions and engineering feasible domain, thus avoiding the decoupling problem of generating the topology first and then patching the constraints in the past, as is the case in existing technologies.

[0016] Third, this invention does not evaluate resilience solely based on average connectivity or average path length, but rather calculates structural evolution uncertainty based on a future topological sample set. Furthermore, a service confidence index is constructed by combining key business parameters with the minimum path cost, the mean of the sample service metrics, and the standard deviation of the sample service metrics. This can reduce the overestimation of actual service capabilities in high-risk scenarios.

[0017] Fourth, this invention will provide a confidence index for the service. Further transformed into the lower bound of risk and lower risk bound With robust toughness value The common input consists of a constrained Markov game composed of an attacking agent and a recovering agent, which enables the policy update process to simultaneously consider service continuity, tail risk, attack budget and recovery resource constraints, thereby improving the stability and risk sensitivity of the evaluation results of attack and recovery policies.

[0018] Fifth, this invention limits local detour reconstruction to be performed within the affected subgraph induced by the subset of affected services, and performs temporary relay deployment when local detour reconstruction cannot meet the business continuity constraints, so that the recovery action is focused on the affected area, reducing the large-scale search of irrelevant areas, improving the efficiency of recovery action selection and the targeted nature of engineering implementation.

[0019] Sixth, after the convergence or stopping conditions are met, this invention simultaneously outputs node resilience, link resilience, attack strategy ranking, and recovery strategy ranking based on counterfactual marginal contributions. It can provide interpretable quantitative evaluation results at the node, link, and strategy levels, providing direct basis for urban communication system planning, risk contingency plan formulation, attack strategy evaluation, and emergency recovery decision-making.

[0020] Seventh, the method steps and system modules of the present invention are constructed around the same closed-loop coupling chain, and there are clear data input and output relationships between each step and each module. It is not a simple parallel combination of several algorithm modules, but forms a complete technical chain from joint action generation, constraint diffusion sampling, lower confidence resilience calculation, constraint game update to counterfactual judgment output. Therefore, it has strong integrity, unity and feasibility. Attached Figure Description

[0021] Figure 1 The flowchart illustrates the overall process of the urban communication system resilience assessment and attack strategy evaluation method provided in this embodiment of the invention.

[0022] Figure 2 This is a block diagram of a system for assessing the resilience of urban communication systems and evaluating attack strategies, provided in an embodiment of the present invention. Detailed Implementation

[0023] The present invention will be further described below with reference to specific embodiments. This embodiment provides a method and system for assessing the resilience of urban communication systems and evaluating attack strategies. It can be applied to urban base station networks, transmission networks, emergency communication networks, aggregation node networks, core data center interconnection networks, and communication systems that carry critical services such as government affairs, emergency response, transportation, finance, and public services.

[0024] In this embodiment, the same basic letter, under different superscripts, subscripts, or parentheses, represents different parameters after being defined by the corresponding superscript, subscript, or parentheses. Subscript Indicates time, Indicates attack attribute. Indicates the restoration attribute. Represents node dimension, Indicates the link dimension. Indicates the future topological sample number.

[0025] In one implementation, this embodiment first constructs the current graph state. It generates a combined action that includes attack sub-actions and recovery sub-actions. Then, by the aforementioned combined action Generate condition vector and joint action feasible domain mask Then the current graph state Conditional vector and joint action feasible domain mask Input a conditional graph diffusion model with feasible region projection to obtain the joint action. A consistent set of future topological samples; then, the structural evolution uncertainty is calculated based on the set of future topological samples. Service confidence index Risk lower bound and robustness value The robustness value and risk lower bound Input a constrained Markov game consisting of an attacking agent and a recovering agent to update the joint action. and the updated joint action As the joint action input for the next closed-loop round, it is used to regenerate the condition vector for the next closed-loop round. and joint action feasible domain mask When the closed-loop update meets the preset convergence condition and the number of closed-loop update rounds reaches the maximum number of closed-loop rounds. Alternatively, if a preset stopping condition is met, the data transmission update is stopped, and the robustness value obtained when the data transmission update is stopped is used as the basis for the update. The counterfactual marginal contribution outputs node resilience, link resilience, attack strategy ranking, and recovery strategy ranking.

[0026] After adopting the above processing, the attack process, recovery process, future topology generation, service capability assessment, risk lower bound calculation and policy update form a unified data coupling closed loop, which can reduce the overestimation of service capability caused by relying solely on average topology indicators, improve the conservatism of urban communication system resilience assessment, the risk sensitivity of attack strategy evaluation, the targeting of recovery strategy selection and the interpretability of the final ranking results.

[0027] Example 1

[0028] like Figure 1 As shown, the method for assessing the resilience of urban communication systems and evaluating attack strategies provided in Embodiment 1 includes steps S1 to S7.

[0029] Step S1: Construct the current graph state and generate joint actions

[0030] In one implementation, step S1 includes acquiring node data, link data, critical service bearer data, node operation status data, link operation status data, historical fault data, attack disturbance data, and recovery and handling data from the urban communication system, and performing node identifier unification, link identifier unification, timestamp unification, unit unification, and outlier cleaning processing on the above data to obtain the time. Current graph state .

[0031] Current diagram status satisfy: , in, For a moment The current state of the graph. For a moment City communication map, For a moment The set of node attributes For a moment The set of link attributes, For a moment The key business pairs.

[0032] In one implementation, a city communication map satisfy: , in, For a moment The set of nodes, For a moment A set of links. A set of nodes. This includes macro base stations, micro base stations, aggregation nodes, core equipment rooms, access nodes, vehicle-mounted relay nodes, portable relay nodes, or satellite emergency access nodes; link aggregation. This includes fiber optic links, microwave links, wireless backhaul links, access links, or temporary emergency links.

[0033] In one implementation, a set of node attributes Link attribute set and key business sets It can be represented as: , , in, For nodes At any moment The node attribute vector, For a set of nodes Any node in; For link At any moment Link attribute vector, For link set Any link in; The source node for the business pair. For the target node of the business pair, For business At any moment The business weight, For business At any moment The acceptable path cost threshold.

[0034] In one implementation, after obtaining the current graph state... Then, generate joint actions. ,satisfy: , in, For a moment Joint actions, For a moment attack sub-actions, For a moment The recovery sub-action. The attack sub-action. Including node attacks, link attacks, or node capacity degradation; recovery sub-actions. This includes node recovery, link recovery, partial detour reconstruction, and temporary relay deployment.

[0035] Preferably, in this embodiment, the generated joint action can also be... Pre-build state fingerprint The state fingerprint Used to indicate the current graph state Compression encoding of node availability, link reachability, critical service load, and historical recovery failure locations. Status fingerprint. This method can be used to determine whether the current state is similar to an inefficient state that has appeared in the historical loop closure. When the similarity exceeds a preset threshold, the priority of the action positions that previously caused inefficient loop closure or infeasible sampling is reduced during subsequent joint action generation. This approach can reduce repeated and invalid strategy searches and improve loop closure update efficiency.

[0036] In this step, the current graph state This provides a unified data foundation for subsequent condition vector generation, joint action feasible region mask generation, future topology sample sampling, service-based confidence index calculation, and constrained Markov game updates; joint actions As a subsequent generation condition vector and joint action feasible domain mask Direct input allows attack and recovery sub-actions to enter the same processing chain from the starting point; state fingerprint As an optional auxiliary quantity, it is used to reduce the recurrence of historically inefficient actions.

[0037] This step unifies the structural data, attribute data, and key business data of the urban communication system into the current graph state. Furthermore, the attack sub-action and the recovery sub-action are uniformly expressed as a combined action. This avoids separating the attack and recovery processes, ensuring that subsequent assessments and policy updates are based on the same state and action semantics.

[0038] Step S2: Generate condition vector and joint action feasible region mask

[0039] In one implementation, step S2 includes, according to the combined action Generate condition vector and joint action feasible domain mask Condition vector satisfy: , in, For a moment The condition vector, For a moment The node attack mask vector, For a moment Link attack mask vector, For a moment The vector of the percentage decrease in node capacity. For a moment The node recovery mask vector, For a moment The link recovery mask vector, For a moment The local bypass reconstruction index vector, For a moment The temporary relay deployment index vector. For the first closed-loop round, and Generate based on the initial recovery strategy, historical recovery handling data, or preset recovery rules; for non-first closed-loop rounds... and This is generated based on the update results of local detour reconstruction and temporary relay deployment in the previous closed-loop round.

[0040] In one implementation, a joint action-feasibility domain mask is used. This is used to identify permitted action locations that satisfy attack budget constraints, recovery resource constraints, deployment location reachability constraints, and business continuity constraints. For candidate node attack locations, candidate link attack locations, candidate node recovery locations, candidate link recovery locations, candidate path switching locations, and candidate temporary relay deployment locations, it is determined whether they satisfy the above constraints respectively.

[0041] Joint action feasible domain mask satisfy: , in, For a moment The joint action feasible domain mask in the first The value at each candidate action position This is the index of the candidate action position; a value of 1 indicates that the corresponding candidate action position meets the constraints and is allowed to be selected, while a value of 0 indicates that the corresponding candidate action position does not meet the constraints and is prohibited from being selected.

[0042] Preferably, in this embodiment, a joint action feasible domain mask can also be used. Introducing boundary buffer identifiers The boundary buffer identifier Used to mark candidate action locations approaching attack budget, resource recovery, or deployment reachability boundaries; superscript This indicates the boundary buffer property. For However, for candidate action positions near the constraint boundary, the boundary buffer identifier... It can be set to 1, and the sampling probability of this candidate action position can be reduced in subsequent diffusion sampling or game update. This can prevent the system from repeatedly generating critical actions near the constraint boundary and reduce closed-loop oscillations.

[0043] The data coupling relationship in this step is: joint action. Decomposed into conditional vectors and joint action feasible domain mask Among them, the condition vector The action semantics of attack and recovery sub-actions are passed to the conditional graph diffusion model; joint action feasibility region mask. Constraints on attack budget, recovery resources, deployment reachability, and business continuity are passed to the conditional graph diffusion model; boundary buffer identifiers. As an optional auxiliary quantity, actions near the feasible region boundary are weighted less. The above data collectively determine the sampling space of the future topological sample set in step S3.

[0044] This step uses condition vectors. and joint action feasible domain mask The dual encoding ensures that subsequent topology generation processes are both aware of the meaning of attack and recovery actions and constrained by the engineering feasible domain, avoiding the generation of future topology samples that are inconsistent with the meaning of the actions or conflict with the engineering constraints; this is achieved through boundary buffering. This can reduce the repeated oscillations in the closed loop caused by critical feasible actions.

[0045] Step S3: Conditional graph diffusion sampling with feasible region projection and model training

[0046] In one implementation, step S3 includes setting the current graph state. Conditional vector and joint action feasible domain mask Input a conditional graph diffusion model with feasible region projection, and sample to obtain a set of future topological samples.

[0047] Conditional graph diffusion sampling satisfies: , in, For a moment The A future topological sample, For a moment The target future topology, To predict the step size, The sample number. The total number of future topological samples and to , For parameters The conditional diffusion distribution, where Dist is an abbreviation for Distribution. These are the parameters of the conditional graph diffusion model.

[0048] In one implementation, during the reverse denoising process, a feasible region projection constraint is applied to the output of each reverse denoising step, satisfying: , in, For the first Step-by-step reverse denoising of the input image. For the first Step-by-step reverse denoising output image This is the sequence number of the reverse denoising step. For parameters The inverse denoising operator, Based on the feasible domain mask of joint action The constraint operator that performs the projection of the feasible region; Proj is an abbreviation for Projection.

[0049] In one implementation, the feasible region projection constraint operator Constraints are applied to the node state, link state, path switching location, and temporary relay deployment location in the reverse denoising output to make the feasible domain mask of the combined action. Locations marked as prohibited are not sampled as valid action locations, thus preventing future topological samples from being included in joint actions. The action semantics and engineering constraints are consistent.

[0050] Preferably, in this embodiment, a shadow topology sample set can also be generated after the conditional graph diffusion sampling. Shadow topology sample set Using the same current graph state and condition vector However, without changing the joint action feasible domain mask Based on the prohibited positions, slight perturbations are applied to the allowed positions for sampling. Then, the consistency of the future topology sample set and the shadow topology sample set is checked to obtain the action consistency residual. : , in, For a moment motion consistency residuals For the first shadow topology sample set A shadow topology sample, This is a topological difference metric function. This is an abbreviation for Distance. If the motion consistency residual... If the value exceeds a preset consistency threshold, the future topology samples in the corresponding batch are resampled or their weights are reduced. This approach can suppress unstable topology samples generated near the feasible region boundary in the conditional graph diffusion model, thereby improving the stability of subsequent resilience calculations.

[0051] The training method for the conditional graph diffusion model is as follows.

[0052] During the training phase, a diffusion training sample set is constructed. Each training sample includes the current graph state obtained from history or simulation. Conditional vector Joint action feasible domain mask and the real future topology .in, This is the training sample set for the conditional graph diffusion model. Indicates diffusion properties.

[0053] Topology of the real future Adding noise, we get the first Step-by-step noise map Then, the conditional graph diffusion model predicts the noise and performs inverse denoising. The training loss of the conditional graph diffusion model satisfies:

[0054]

[0055] in, To mitigate training loss, For expectation operator, This is to add real noise to the actual future topology during the training phase. For parameters The noise prediction function, For the first Step-by-step noise map It is a norm 2. To constrain the loss weights using a mask, The mask is used to constrain the loss.

[0056] During training, the parameters of the conditional graph diffusion model can be updated using Adam (Adaptive Moment Estimation) or mini-batch stochastic gradient descent. Until the training loss is diffused. Convergence or reaching the preset training rounds.

[0057] In this step, the current graph state Provides the current communication architecture and service foundation, condition vectors Provides attack and recovery semantics for joint actions, and a feasible domain mask for joint actions. Provides feasible action boundaries. These three factors collectively determine the resulting topology sample set generation. Action consistency residuals. As an optional verification measure, the future topology sample set is used for stability screening. The screened future topology sample set continues to be used as input for steps S4 and S5 to calculate structural evolution uncertainty, service confidence index, risk lower bound, and robustness value.

[0058] This step uses conditional graph diffusion sampling with feasible region projection to ensure that the future topological sample set is not a freely generated result detached from action constraints, but rather a result linked to joint actions. and joint action feasible domain mask Consistent constrained topology distribution; through shadow topology consistency verification, the interference of unstable samples on resilience calculation can be further reduced.

[0059] Step S4: Calculate the structural evolution uncertainty

[0060] In one implementation, step S4 includes extracting a topological statistics vector for each future topological sample. ,satisfy: , in, For a moment The The topological statistics vector corresponding to each future topological sample. The size of the maximum connected subgraph. The average weighted shortest path length, The total betweenness of critical nodes. For key business reachability.

[0061] Structural evolution uncertainty satisfy: , in, For a moment Structural evolution uncertainty, Index for topological statistics components. topological statistics vector The One portion, For the first The weighting coefficients of each topological statistic component. For sample serial number The variance operator for calculating variance; Var is an abbreviation for variance.

[0062] Preferably, in this embodiment, the weighting coefficients of the topological statistics components can also be based on historical prediction errors. Adaptive updates are performed. For topological statistics components with significant errors in historical assessments, their corresponding weights are increased; for topological statistics components with smaller fluctuations and weaker impact on service capability, their corresponding weights are decreased. This approach enables the reduction of structural evolution uncertainty. More attention is being paid to structural fluctuations that have a greater impact on the ability to serve key business operations.

[0063] In this step, the future topological sample set output from step S3 is converted into a topological statistics vector. The structural evolution uncertainty is then calculated from the statistical fluctuations among multiple future topological samples. Structural evolution uncertainty This will be used as the service confidence indicator in step S5. The penalty term is included as part of the constraint Markov game state in step S7.

[0064] This step involves structural evolution uncertainty. Explicitly characterizing the structural fluctuations between future topology samples makes resilience assessment no longer dependent on a single topology or average topology, which helps to reduce the overestimation of the service capabilities of urban communication systems in future topologically unstable scenarios.

[0065] Step S5: Calculate the service confidence index, risk lower bound, and robustness value.

[0066] In one implementation, step S5 includes calculating the comprehensive path cost for each link in the future topology sample. ,satisfy: , in, For the first Links in a future topology sample The overall path cost For cost component index, For the first Class cost component weighting coefficient, For link The Class cost component. When hour, Represents the inverse bandwidth cost component; when hour, Represents the delay cost component; when hour, This represents the reliability loss cost component; when hour, Indicates the maintenance cost component; when hour, This indicates the risk cost component.

[0067] In one implementation, for business pairs Calculate the minimum path cost ,satisfy: , in, For business The minimum path cost, As the source node, For the destination node, For business The set of candidate paths, Candidate paths, Candidate paths The links in the network.

[0068] In one implementation, based on minimum path cost Does it meet the acceptable path cost threshold? Calculate sample service indicators ,satisfy: , in, For the first Sample service metrics corresponding to each future topology sample For business At any moment The business weight, For indicator functions, For business At any moment The acceptable path cost threshold For a moment The key business pairs.

[0069] Sample service index mean and the standard deviation of sample service indicators satisfy: , , in, The mean of the service index for the sample. The standard deviation of the service index for the sample. This represents the total number of future topological samples.

[0070] Service confidence index satisfy: , in, To serve the lower confidence index, This is the structural uncertainty penalty coefficient. The service fluctuation penalty coefficient.

[0071] Preferably, in this embodiment, the structural uncertainty penalty coefficient can be adjusted based on historical prediction deviations. Service fluctuation penalty coefficient Adaptive calibration is performed. Specifically, when the actual service capacity in historical assessments is lower than the service confidence index... When the number of times increases, it improves or When the service confidence index When the service capacity remains below actual capacity for an extended period, leading to excessive investment in recovery resources, the service level will decrease. or This method enables confidence metrics under the service. An adaptive balance is achieved between conservatism and resource utilization efficiency.

[0072] In one implementation, a set of sample service losses is calculated based on sample service metrics from a future topology sample set. (Sample service loss) satisfy: , in, For a moment The The service loss corresponding to each future topology sample As a baseline service indicator, For the first Sample service metrics corresponding to each future topology sample.

[0073] Risk lower bound satisfy: , in, For a moment The lower bound of risk, Confidence level The Conditional Value at Risk (CVaR) operator is used to define the Conditional Value at Risk. For risk confidence level, For the reason The sample service loss set consists of the service loss corresponding to each future topology sample.

[0074] Robust toughness value satisfy: , in, For a moment robustness value For baseline service indicators and , To truncate the input value to Interval cutoff function, It is a natural exponential function. This represents the tail risk penalty coefficient.

[0075] Within the evaluation window, the overall robustness value satisfy: , in, To evaluate the overall robustness value within the window, To evaluate the window length.

[0076] In this step, the future topology sample set is first used to calculate the comprehensive path cost of the links. and the minimum path cost for business Then used to calculate sample service indicators Mean of sample service indicators and the standard deviation of sample service indicators Subsequently, , and Confidence metrics under jointly generated services Sample service indicators Further generate a sample service loss set The lower bound of risk is obtained through conditional value at risk calculation. Finally, in Under the conditions, , and The robustness value calculation formula is input together and passed through the cutoff function. Limit the calculation results to Within the range, the robustness value is obtained. Then , , and Proceed to step S7.

[0077] This step uses the service confidence index. We consider average service capacity, structural volatility, and service volatility in a unified manner, and calculate the lower bound of risk using a sample service loss set. This approach avoids optimistic assessments caused by using only a single average service metric, and improves the conservatism and reliability of resilience assessment results in high-risk scenarios.

[0078] Step S6: Local detour reconstruction and temporary relay deployment within the affected subgraph

[0079] In one implementation, step S6 includes generating a subset of affected service pairs only for critical service pairs affected by the attacked sub-action or the recovery sub-action. And by the affected business to a subset Inducing the affected subgraph Then, in the affected subgraph Internal partial reconfiguration is performed.

[0080] link The average comprehensive path cost of the sample satisfy: , in, For link The average comprehensive path cost of the sample. For the first Links in a future topology sample The overall path cost This represents the total number of future topological samples.

[0081] Business Optimal local detour path satisfy: , in, For business The optimal local detour path, For business In the affected subgraph The set of candidate paths within, This is the resilience increment trade-off coefficient. This is the risk change trade-off coefficient. To select a path Post-business The corresponding robustness increment, To select a path Post-business The corresponding change in the lower bound of risk.

[0082] When partial rerouting and reconstruction cannot meet business continuity constraints, a temporary relay deployment is executed. The temporary relay deployment satisfies: , , in, For a moment temporary links capacity, It is an abbreviation for Capacity. For a moment temporary links The time delay, It is an abbreviation for Delay. For a moment The The processing capacity of a temporary relay node It is an abbreviation for Processing capacity. For temporary relay node index; and These are the lower and upper limits of temporary link capacity, respectively. and These are the lower and upper limits of temporary link latency, respectively. This represents the lower limit of the processing capacity of temporary relay nodes. Indicates a temporary attribute.

[0083] Preferably, in this embodiment, cooling markers can also be set for temporary relay deployment locations. If a temporary relay deployment location is selected in multiple consecutive closed-loop rounds but fails to improve robustness, the following applies: Then write the cooling identifier at that location. Furthermore, the selection priority of this location is reduced in subsequent rounds of closed-loop updates. This approach avoids the repeated deployment of temporary relay resources in inefficient locations, improving the efficiency of resource recovery utilization.

[0084] In this step, the link synthesis path cost output in step S5 is... Robust toughness value and risk lower bound It is further used for local detour reconstruction and temporary relay deployment selection. The results of local detour reconstruction and temporary relay deployment are used as part of the recovery sub-action to reorganize the updated joint action. This is then fed back into the condition vector as input for the next closed-loop round. and joint action feasible domain mask The generation process; therefore, the future topology sample set of the current closed-loop round is generated by the joint actions already determined in the current closed-loop round. The local detour reconstruction results and temporary relay deployment results obtained in the current closed-loop round are used to constrain and update the joint actions of the next closed-loop round. Cooling indicator As an optional auxiliary quantity, it is used to reduce the priority of duplicate and inefficient deployment locations.

[0085] This step involves applying the method only to a subset of the affected business pairs. Induced affected subgraph Performing local reconfiguration via detours can narrow the recovery search scope and improve the efficiency of recovery action selection; by introducing temporary relay deployment constraints when local detours are insufficient, the engineering feasibility of critical business continuity recovery can be improved; and by using cooling indicators... It can reduce the duplication of recovery resources to inefficient locations.

[0086] Step S7: Constrain the Markov game by updating the joint actions and performing a counterfactual evaluation.

[0087] In one implementation, step S7 includes obtaining the structural evolution uncertainty. Service confidence index Risk lower bound and robustness value Subsequently, a constrained Markov game consisting of an attacking agent and a recovering agent is constructed. Markov games are used to represent game models in which multiple agents update their policies based on their respective actions and payoffs during state transitions.

[0088] Game state satisfy: , in, For a moment The state of the game, For a moment The cumulative attack budget consumption, For a moment The cumulative recovery resource consumption.

[0089] Attack budget constraints and recovery resource constraints are satisfied: , , in, Add an index for time accumulation. For a moment The cost of attack sub-actions, Represents the cost function. To reach the attack budget limit, For a moment Resource consumption of recovery sub-actions For the Resource consumption function, To restore the resource limit.

[0090] Incremental robustness satisfy: , in, For a moment The robustness increment, This represents the robustness value at the previous moment.

[0091] Attacking agent rewards and recovery agent rewards satisfy: , , in, For a moment Rewards for attacking intelligent agents, For a moment The reward for the recovery agent, to This is the reward weighting coefficient.

[0092] The attacking agent and the recovering agent can be trained using either PPO (Proximal Policy Optimization) or Actor-Critic methods, respectively.

[0093] During the training phase, the attacking agent and the recovering agent are in a game state. As input, output the attack sub-actions respectively. and recovery sub-action Execute joint operations Then, following steps S2 to S6, the condition vector, joint action feasible region mask, future topology sample set, service lower confidence index, risk lower bound, and robustness value are regenerated, and the results are calculated based on the attacking agent's reward. and recovery agent rewards Update the strategy parameters.

[0094] The strategy parameter update satisfies: , , in, For the attack agent's policy parameters, To restore the agent's policy parameters, To attack the learning rate of the agent, To restore the agent's learning rate, To attack the agent's objective function, To recover the agent's objective function, Let be the gradient of the attack agent's objective function with respect to the attack agent's policy parameters. To recover the gradient of the agent's objective function with respect to the agent's policy parameters.

[0095] During training, if the attack budget constraint or recovery resource constraint is violated, the corresponding action is masked by the joint action feasible domain. By shielding or penalizing in the reward function, the trained policy can output feasible attack and recovery sub-actions within budget and resource constraints.

[0096] Preferably, in this embodiment, closed-loop convergence can be determined using a dual-threshold hysteresis method. Specifically, a convergence threshold is set. and convergence exit threshold and make ,in, To converge to the threshold, To converge to the exit threshold, Indicates entering the attribute. This indicates the exit attribute. When the changes in robustness value, risk lower bound, and joint action are all below the convergence threshold in multiple consecutive closed-loop updates... When the condition is met, the system is determined to enter a convergence candidate state; if any change in the convergence candidate state exceeds the convergence exit threshold again... If so, it exits the convergence candidate state and continues closed-loop updates. Because Greater than The closed-loop process only exits the convergence candidate state when the change amount increases significantly, thus avoiding repeated entry and exit from the convergence candidate state within the critical interval. This method can prevent the closed-loop process from repeatedly converging and diverging in the critical state, improving the stability of the strategy output.

[0097] Furthermore, to avoid the constrained Markov game repeatedly switching between a few joint actions, which could lead to a continuous non-convergence of the closed-loop process, this embodiment sets a maximum number of closed-loop rounds. and no improvement in round threshold When the closed-loop update rounds reach the maximum number of closed-loop rounds. When, stop continuing to transmit updates; or, when continuously Robust toughness value during closed-loop update No improvement and lower bound of risk If the risk level is not reduced, stop feeding back updates and select the joint action with higher robustness value and lower risk lower bound from the already generated feasible joint actions as the final joint action.

[0098] After satisfying the double-threshold hysteresis convergence condition and reaching the maximum number of closed-loop cycles. Or, after stopping the feedback update when the no-improvement stopping condition is met, the robustness value obtained at the time of stopping is used. Perform counterfactual marginal contribution calculations for nodes, links, candidate attack strategies, and candidate recovery strategies respectively.

[0099] Node resilience and link resilience satisfy: , , in, For nodes Node resilience, For nodes The robustness value obtained after performing counterfactual removal; For link Link resilience, For the link The robustness value obtained after performing counterfactual removal.

[0100] Counterfactual marginal damage of candidate attack strategies Counterfactual marginal recovery gain of candidate recovery strategies satisfy: , , in, For the first One candidate attack strategy Counterfactual marginal damage, For a moment The benchmark robustness value, To execute the candidate attack strategy The robustness value obtained afterwards; For the first One candidate recovery strategy The counterfactual marginal recovery gain, To execute candidate recovery strategies The robustness value obtained afterwards.

[0101] Candidate attack strategy ranking value and candidate recovery strategy ranking value satisfy: , , in, Candidate attack strategies The sorting value, Candidate attack strategies The corresponding lower bound of risk, Candidate attack strategies The strategy cost; Candidate recovery strategies The sorting value, Candidate recovery strategies Strategic resource consumption, Candidate recovery strategies The corresponding lower bound of risk; Index of candidate attack strategies For candidate recovery strategy index, to These are the ranking weight coefficients. To prevent positive numbers with a denominator of zero.

[0102] In this step, the structural evolution uncertainty output in step S5 Service confidence index Risk lower bound and robustness value The input-constrained Markov game forms the game state. It also calculates the rewards for the attacking agent and the recovering agent; the updated attack sub-actions and recovering sub-actions are then recombined into an updated joint action. This is then used as the joint action input for the next closed-loop round and fed back to step S2 to generate the condition vector for the next closed-loop round. and joint action feasible domain mask The convergence condition with double threshold hysteresis is met, and the maximum number of closed-loop cycles is reached. If the no-improvement stopping condition is met, continue the backfeeding and proceed to the counterfactual marginal contribution calculation. After stopping the backfeeding update, the robustness value obtained at the time of stopping is... It continues to serve as the benchmark for calculating counterfactual marginal contributions, used to output node resilience, link resilience, attack strategy ranking, and recovery strategy ranking. A dual-threshold hysteresis determination method is used as an optional convergence control mechanism to improve closed-loop stability.

[0103] This step involves defining the lower bound of risk. and robustness value Simultaneously inputting constrained Markov games for both the attacking and recovering agents allows the policy update process to consider attack effects, recovery rewards, tail risks, attack budgets, and recovery resource constraints. Through counterfactual marginal contribution calculation, interpretable evaluation results at the node, link, and policy levels can be generated. By using dual-threshold hysteresis convergence determination, repeated oscillations in the closed-loop critical state can be reduced.

[0104] Example 2

[0105] like Figure 2As shown in Example 2, a system for assessing the resilience of urban communication systems and evaluating attack strategies is provided. This system includes a graph state and constraint coding module, a constraint diffusion sampling module, a lower confidence resilience calculation module, a game-theoretic closed-loop optimization module, and an evaluation output module.

[0106] In one implementation, the graph state and constraint encoding module is used to construct the current graph state. and joint actions and by joint action Generate condition vector and joint action feasible domain mask This module corresponds to executing steps S1 and S2, and its output is the current graph state. Conditional vector and joint action feasible domain mask Transmitted to the constrained diffusion sampling module.

[0107] In one implementation, the constraint diffusion sampling module is used to obtain the current graph state. Conditional vector and joint action feasible domain mask The input conditional graph diffusion model with feasible region projection is used to make the inverse denoising output subject to the joint action feasible region mask. Projection constraints are applied, and a future topology sample set is obtained. This module corresponds to step S3, and its output future topology sample set is transmitted to the next confidence resilience calculation module.

[0108] In one implementation, the lower confidence resilience calculation module is used to calculate the structural evolution uncertainty from a future topological sample set. Service confidence index Risk lower bound and robustness value This module corresponds to steps S4 and S5, and its output is the structural evolution uncertainty. Service confidence index Risk lower bound and robustness value The data is transmitted to the game closed-loop optimization module and the evaluation output module.

[0109] In one implementation, the game-theoretic closed-loop optimization module is used to optimize the robustness value. and risk lower bound Inputting a constrained Markov game between the attacking agent and the recovering agent, updating the joint action. and the updated joint action The data is fed back to the graph state and constraint coding module; the graph state and constraint coding module then uses the updated joint action... Regenerate condition vector and joint action feasible domain mask Then regenerate the condition vector and joint action feasible domain mask The data is transmitted to the constraint diffusion sampling module. This module corresponds to the strategy update part in step S7, and receives the local detour reconstruction result and temporary relay deployment result output in step S6, so that the local detour reconstruction result and temporary relay deployment result are used as part of the recovery sub-action in the next closed-loop round.

[0110] In one implementation, the evaluation output module is used to evaluate the robustness value obtained after meeting the convergence or stopping conditions. The module calculates the counterfactual robustness value obtained from the future topology sample set, performs counterfactual marginal contribution calculation, and outputs node resilience, link resilience, attack strategy ranking, and recovery strategy ranking. This module corresponds to the counterfactual evaluation part in step S7.

[0111] In the above system, the modules sequentially form the following closed-loop coupling: , in, For closed-loop round index, For the first The combined actions of a closed-loop cycle, For the first The condition vector for each closed-loop cycle. For the first Joint action feasible domain mask for each closed-loop round , , and The first The structural evolution uncertainty, service lower confidence index, risk lower bound, and robustness value obtained from each closed-loop cycle are: According to the first The next closed-loop round's joint action is obtained by updating the evaluation results of the previous closed-loop round.

[0112] This system, through strict data input-output relationships between its modules, enables the complete implementation of joint action generation, constraint diffusion sampling, lower confidence resilience calculation, constraint game update, and counterfactual evaluation output at the system level. The system's modules are not simply parallel combinations, but rather work collaboratively around the same closed-loop coupling chain, improving the reliability, interpretability, and engineering feasibility of urban communication system resilience assessment and attack strategy evaluation.

[0113] By adopting the method and system of this embodiment, at least the following overall beneficial effects can be obtained.

[0114] First, the attack sub-action and the recovery sub-action are unified into a combined action. This can avoid data breakage caused by separating the attack process and the recovery process in the model.

[0115] Second, condition vector and joint action feasible domain mask The common input conditional graph diffusion model enables future topological sample sets to simultaneously satisfy action semantic consistency and engineering feasibility.

[0116] Third, the introduction of feasible domain projection constraints in the reverse denoising stage can prevent nodes, links, path switching locations, and temporary relay deployment locations from appearing in future topology samples that violate attack budget, recovery resource, deployment reachability, or business continuity constraints.

[0117] Fourth, structural evolution uncertainty Service confidence index It can incorporate future fluctuations in topology distribution and fluctuations in service capabilities of key businesses into resilience assessments, reducing the optimistic overestimation of the service capabilities of urban communication systems in high-risk scenarios.

[0118] Fifth, lower bound of risk and robustness value Markov games constrained by common inputs enable the updating of attack and recovery strategies to take into account service continuity, tail risk, attack budget, and recovery resource constraints.

[0119] Sixth, local detour reconstruction is limited to the affected subgraph. Internal execution, and temporary relay deployment when necessary, can reduce searches in irrelevant areas and improve the targeting and engineering feasibility of recovery actions.

[0120] Seventh, the counterfactual marginal contribution calculation can simultaneously output node resilience, link resilience, attack strategy ranking, and recovery strategy ranking, providing interpretable evaluation results at the node, link, and policy levels.

[0121] Eighth, shadow topology consistency verification, service-based confidence adaptive calibration, temporary relay cooling flag, and dual-threshold hysteresis convergence determination can further improve sample generation stability, risk calibration accuracy, recovery resource utilization efficiency, and closed-loop convergence stability.

[0122] The present invention has been described above with reference to specific embodiments. Those skilled in the art should understand that, without departing from the concept of the present invention, equivalent substitutions or transformations can be made to the attribute organization method of the current graph state, the encoding method of the condition vector, the generation method of the joint action feasible region mask, the network structure of the conditional graph diffusion model, the implementation method of the reverse denoising process, the weight setting method of the confidence index under the service, the calculation method of the risk lower bound, the training method of the constrained Markov game, and the calculation details of the counterfactual marginal contribution. All technical solutions based on the technical concept of the present invention and falling within the scope of protection of the claims should be protected by the present invention.

Claims

1. A method for assessing the resilience of urban communication systems and evaluating attack strategies, characterized in that, include: Build Time Current graph state and generate time joint action The current graph state This includes a city communication map, a set of node attributes, a set of link attributes, and a set of key service pairs; the joint action... It includes attack sub-actions and recovery sub-actions, wherein the recovery sub-actions include local bypass reconstruction and temporary relay deployment; By the aforementioned joint action Generation time condition vector and time Joint action feasible domain mask ; The current graph state Conditional vector and joint action feasible domain mask The input conditional graph diffusion model with feasible region projection is used to make the inverse denoising output subject to the joint action feasible region mask. Projection constraints are used to obtain the joint action. A consistent set of future topological samples; The time is calculated from the future topology sample set. Structural evolution uncertainty Service confidence index Risk lower bound and robustness value ; The robustness value and risk lower bound Input a constrained Markov game consisting of an attacking agent and a recovering agent to update the joint action. and the updated joint action As the joint action input for the next closed-loop round, it is used to regenerate the condition vector for the next closed-loop round. and joint action feasible domain mask ; When the closed-loop update meets the preset convergence condition and the number of closed-loop update rounds reaches the maximum number of closed-loop rounds. Alternatively, if a preset stopping condition is met, the data transmission update is stopped, and the robustness value obtained when the data transmission update is stopped is used as the basis for the update. The counterfactual marginal contribution outputs node resilience, link resilience, attack strategy ranking, and recovery strategy ranking.

2. The method for assessing the resilience of urban communication systems and evaluating attack strategies according to claim 1, characterized in that, During the build process Current graph state and generate time joint action In this process, the city communication map, node attribute set, link attribute set, and key business pair set are combined into the current map state. It combines the attack sub-action and the recovery sub-action into a combined action. In step S2, the attack location, recovery location, local bypass reconstruction index, and temporary relay deployment index are jointly encoded into a condition vector. ,satisfy: ， in, For a moment The current state of the graph. For a moment City communication map, For a moment The set of node attributes, For a moment The set of link attributes, For a moment Key business pairs; For a moment joint actions, For a moment attack sub-actions, For a moment The recovery sub-action; For a moment The condition vector, For a moment The node attack mask vector, For a moment Link attack mask vector, For a moment The vector of the percentage decrease in node capacity. For a moment The node recovery mask vector, For a moment The link recovery mask vector, For a moment The local bypass reconstruction index vector, For a moment Temporary relay deployment index vector; For a moment The joint action feasible domain mask is used to identify the allowed action locations that simultaneously satisfy attack budget constraints, recovery resource constraints, deployment location reachability constraints, and business continuity constraints; Represents node dimension, Indicates the link dimension. Indicates attack attribute. This indicates the attribute to be restored.

3. The method for assessing the resilience of urban communication systems and evaluating attack strategies according to claim 2, characterized in that, After obtaining the condition vector and joint action feasible domain mask Then, the current graph state is... Conditional vector and joint action feasible domain mask The conditional graph diffusion model, which serves as a conditional input with feasible region projection, is then used in the reverse denoising process based on the joint action feasible region mask. Projection constraints are applied to the node, link, path switching location, and temporary relay deployment location to satisfy: ， in, For a moment The A future topological sample, For a moment The target future topology, To predict the step size, The sample number. The total number of future topological samples and to ; For parameters The conditional diffusion distribution These are the parameters of the conditional graph diffusion model; For the first Step-by-step reverse denoising of the input image. For the first Step-by-step reverse denoising output image This is the sequence number of the reverse denoising step; For parameters The inverse denoising operator; Based on the feasible domain mask of joint action Constraint operators that perform feasible region projections.

4. The method for assessing the resilience of urban communication systems and evaluating attack strategies according to claim 3, characterized in that, After obtaining the future topology sample set, the maximum connected subgraph size, average weighted shortest path length, total betweenness of key nodes, and reachability of key business pairs are extracted for each future topology sample. Based on the sample variance of each topology statistic within the future topology sample set, the structural evolution uncertainty is calculated. ,satisfy: ， in, For a moment The The topological statistics vector corresponding to each future topological sample. The size of the maximum connected subgraph. The average weighted shortest path length, The total betweenness of critical nodes. For key business reachability; For a moment Structural evolution uncertainty, Index for topological statistics components. topological statistics vector The One portion, For the first The weighting coefficients of each topological statistic component. For sample serial number The variance operator for finding the variance.

5. The method for assessing the resilience of urban communication systems and evaluating attack strategies according to claim 4, characterized in that, Calculate the confidence index under the service. In the process, the comprehensive path cost of the link is first calculated based on the link cost components in the future topology sample, then the minimum path cost is calculated for the candidate path set based on the key business, and the sample service index is obtained based on whether the minimum path cost meets the acceptable path cost threshold. Subsequently, the mean of the sample service index, the uncertainty of structural evolution, and the standard deviation of the sample service index are coupled to obtain the service confidence index. ,satisfy: ， in, For the first Links in a future topology sample The overall path cost For cost component index, For the first Class cost component weighting coefficient, For link The Class cost components, and to These correspond to the inverse bandwidth cost component, the latency cost component, the reliability loss cost component, the maintenance cost component, and the risk cost component, respectively. For business The minimum path cost, As the source node, For the destination node, For business The set of candidate paths, Candidate paths, Candidate paths The link in; For the first Sample service metrics corresponding to each future topology sample For business At any moment Business weight, For indicator functions, For business At any moment The acceptable path cost threshold; The mean of the service index for the sample. The standard deviation of the service index for the sample. To serve the lower confidence index, This is the structural uncertainty penalty coefficient. To serve the fluctuation penalty coefficient, This represents the total number of future topological samples.

6. The method for assessing the resilience of urban communication systems and evaluating attack strategies according to claim 5, characterized in that, Confidence index under the aforementioned service Subsequently, a sample service loss set is calculated based on the sample service metrics in the future topology sample set, and a conditional value-at-risk (VAT) calculation is performed on the sample service loss set to obtain a lower bound for risk. Then, the service confidence index Baseline service indicators and risk lower bound Co-coupling results in robust toughness value ,satisfy: ，， ， in, For a moment The The service loss corresponding to each future topology sample For baseline service indicators and , For the first Sample service metrics corresponding to each future topology sample; For a moment The lower bound of risk, Confidence level Conditional Value at Risk (VaR) operator, For risk confidence level, For the reason The sample service loss set consists of the service loss corresponding to each future topology sample; For a moment robustness value To truncate the input value to Interval cutoff function, It is a natural exponential function. This is the tail risk penalty coefficient; To evaluate the overall robustness value within the window, To evaluate the window length.

7. The method for assessing the resilience of urban communication systems and evaluating attack strategies according to claim 6, characterized in that, In updating the joint action In the steps, the current graph state is... Structural evolution uncertainty Service confidence index Risk lower bound The cumulative attack budget consumption and cumulative recovery resource consumption are combined into a game state, and the lower bound of the risk is defined. and robustness value Input the reward functions for the attacking agent and the recovering agent respectively, and update the attack and recovery sub-actions under budget and resource constraints, satisfying the following: ， in, For a moment The state of the game, For a moment The cumulative attack budget consumption, For a moment The cumulative recovery resource consumption, Add a time-based index; For a moment The cost of attack sub-actions, The attack budget cap; For a moment Resource consumption of recovery sub-actions To restore the resource limit; For a moment The robustness increment, The robustness value at the previous moment; For a moment Rewards for attacking intelligent agents, For a moment The reward for the recovery agent, to The reward weighting coefficient; the updated attack sub-action. and recovery sub-action Reorganize the updated joint actions and the updated joint action As input to the next closed-loop iteration, it is fed back to the condition vector. and joint action feasible domain mask The generation steps are as follows: when continuous closed-loop updates meet the preset convergence conditions and the number of closed-loop update rounds reaches the maximum number of closed-loop rounds. Alternatively, when preset stop conditions are met, the combined action feedback will be stopped.

8. The method for assessing the resilience of urban communication systems and evaluating attack strategies according to claim 7, characterized in that, In the steps of performing the local bypass reconstruction and temporary relay deployment, a subset of affected service pairs is generated only for critical service pairs affected by the attack sub-action or recovery sub-action. and the affected service pairs subset Inducing the affected subgraph ; In the affected subgraph Perform partial relocation and reconstruction within the system, and if the partial relocation and reconstruction cannot meet business continuity constraints, execute temporary relay deployment, satisfying the following: ， in, For a moment The affected business subset For a moment The affected subgraph For link The average comprehensive path cost of the sample; For business The optimal local detour path, For business In the affected subgraph The set of candidate paths within, This is the resilience increment trade-off coefficient. This is the risk change trade-off coefficient. To select a path Post-business The corresponding robustness increment, To select a path Post-business The corresponding change in the lower bound of risk; For a moment temporary links capacity, For a moment temporary links The time delay, For a moment The The processing capacity of a temporary relay node For temporary relay node index; and These are the lower and upper limits of temporary link capacity, respectively. and These are the lower and upper limits of temporary link latency, respectively. This represents the lower limit of the processing capacity of temporary relay nodes. Indicates a temporary attribute; the optimal local detour path The results of temporary relay deployments that satisfy the constraints are encoded as part of the recovery sub-action and used as the joint action after the next closed-loop round update. The recovery action input is used as the condition vector for the next closed-loop round. and joint action feasible domain mask The generation of .

9. The method for assessing the resilience of urban communication systems and evaluating attack strategies according to claim 8, characterized in that, In the steps of outputting node resilience, link resilience, attack strategy ranking, and recovery strategy ranking, the robustness value is based on the robustness value obtained after satisfying the convergence or stopping condition. Counterfactual marginal contributions are calculated for nodes, links, candidate attack strategies, and candidate recovery strategies, respectively. Specifically, counterfactual removal is performed on nodes and links to obtain node resilience and link resilience; counterfactual attacks are performed on candidate attack strategies to obtain attack strategy ranking values; and counterfactual recovery is performed on candidate recovery strategies to obtain recovery strategy ranking values, satisfying the following: ， in, For nodes Node resilience, For nodes The robustness value obtained after performing counterfactual removal. For link Link resilience, For the link The robustness value obtained after performing counterfactual removal; For the first One candidate attack strategy Counterfactual marginal damage, For a moment The benchmark robustness value, To execute the candidate attack strategy The robustness value obtained afterwards; For the first One candidate recovery strategy The counterfactual marginal recovery gain, To execute candidate recovery strategies The robustness value obtained afterwards; Candidate attack strategies The sorting value, Candidate attack strategies The corresponding lower bound of risk, Candidate attack strategies The strategy cost; Candidate recovery strategies The sorting value, Candidate recovery strategies Strategic resource consumption, Candidate recovery strategies The corresponding lower bound of risk; Index of candidate attack strategies For candidate recovery strategy index, to These are the ranking weight coefficients. To prevent positive numbers with a denominator of zero.

10. A system for assessing the resilience of urban communication systems and evaluating attack strategies, characterized in that, include: The graph state and constraint encoding module is used to construct the current graph state. and joint actions And by the joint action Generate condition vector and joint action feasible domain mask ; The constraint diffusion sampling module is used to sample the current graph state. Conditional vector and joint action feasible domain mask The input conditional graph diffusion model with feasible region projection is used to make the inverse denoising output subject to the joint action feasible region mask. Projection constraints are applied to obtain a future topology sample set. The lower confidence resilience calculation module is used to calculate the structural evolution uncertainty from the future topological sample set. Service confidence index Risk lower bound and robustness value ; The game-theoretic closed-loop optimization module is used to optimize the robustness value. and risk lower bound Inputting a constrained Markov game between the attacking agent and the recovering agent, updating the joint action. and the updated joint action The data is fed back to the graph state and constraint encoding module, enabling the graph state and constraint encoding module to base its actions on the updated joint actions. Regenerate condition vector and joint action feasible domain mask Then regenerate the condition vector and joint action feasible domain mask Transmitted to the constrained diffusion sampling module; The evaluation output module is used to evaluate the robustness value obtained after meeting the convergence or stopping conditions. The counterfactual robustness value calculated from the future topology sample set is used to perform counterfactual marginal contribution calculation, and output node resilience, link resilience, attack strategy ranking, and recovery strategy ranking.

Images