A task context-driven unmanned aerial vehicle swarm quantum key splitting method and system

Abstract

The present application relates to the technical field of quantum communication, and especially to a task context driven UAV group quantum key splitting method and system, the method comprising the following steps: obtaining task context data for preprocessing to generate a standardized group situation tensor; using a group time series prediction model to obtain a group prediction atlas; presetting a global resource upper limit constraint and introducing a resource auction mechanism to determine individual difference resource quotas based on the group prediction atlas; constructing a key splitting strategy table for a quantum key pool to obtain key splitting segments; issuing the key splitting strategy table and the key splitting segments for execution and collecting an execution feedback dataset; obtaining a deviation comparison result of the execution feedback dataset and the group prediction atlas to update the group time series prediction model, thereby realizing a closed loop for splitting. The present application dynamically adapts to the situation of a UAV group, optimizes key splitting in a closed loop, and improves resource utilization.

Description

Technical Field

[0001] This invention relates to the field of quantum communication technology, and in particular to a task context-driven method and system for quantum key distribution in unmanned aerial vehicle (UAV) swarms. Background Technology

[0002] When drone swarms perform collaborative reconnaissance, strike, and relay missions, each node needs to establish an unconditionally secure confidential communication channel through quantum key distribution. Existing quantum key management schemes mainly fall into two categories: one where drones generate their own quantum keys using quantum true random number generation; and another where a ground control center or quantum key distribution infrastructure generates keys and pre-installs them into each drone terminal. Regarding key allocation, existing schemes typically employ static hierarchical allocation based on swarm identifiers, or unified key resource configuration by a central node. For key update mechanisms, a threshold-triggered approach is used, where a drone requests key replenishment or updates from the control center when it detects that its local key pool balance is below a preset threshold.

[0003] However, existing technologies have the following shortcomings: First, key allocation lacks task adaptability. Static pre-configuration or group identifier allocation methods cannot dynamically adjust key resources according to the actual task types and priorities undertaken by each node, resulting in low-risk nodes consuming excessive keys while high-risk nodes have insufficient key strength. Second, key updates are passive responses. Threshold triggering or risk level response can only reflect the current situation and cannot predict trend changes such as drones entering high-threat areas. Delayed key availability may cause a vacuum period in secure communication. Third, there is a lack of global consideration of drone swarm resource constraints. Onboard energy and computing overhead are not uniformly scheduled as constraints for key management. Each node manages independently, which can easily cause some nodes to run out of power prematurely due to excessive key management overhead. Fourth, functional modules are relatively independent and lack a systematic closed-loop architecture, making it impossible to achieve adaptive iterative optimization of key management strategies. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a task context-driven method and system for quantum key distribution in unmanned aerial vehicle (UAV) swarms.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a task context-driven quantum key distribution method for UAV swarms. The method includes the following steps: acquiring task context data of each node in the UAV swarm; preprocessing the task context data to generate a standardized swarm state tensor; using a swarm temporal prediction model to predict the standardized swarm state tensor to obtain a swarm prediction graph of the UAV swarm; pre-setting a global resource upper limit constraint and introducing a resource auction mechanism to determine the individual resource quotas of each node based on the swarm prediction graph; constructing a key distribution strategy table for the quantum key pool based on the swarm prediction graph and the individual resource quotas to obtain key distribution fragments; distributing the key distribution strategy table and the key distribution fragments to the nodes for execution and collecting the execution feedback dataset of the UAV swarm; obtaining a deviation comparison result between the execution feedback dataset and the swarm prediction graph to update the swarm temporal prediction model and achieve a closed-loop distribution. This invention matches the dynamic task state of the UAV swarm, ensures the rationality of key allocation, and improves the security and dynamic adaptability of UAV swarm key communication.

[0006] Optionally, the step of acquiring the task context data of each node in the UAV swarm and preprocessing the task context data to generate a standardized swarm situation tensor includes: acquiring the task context data, including node spatial coordinates, velocity vectors, qubit error rates, and remaining battery power; performing outlier removal and missing value imputation on the task context data to obtain a preprocessed dataset to construct an initial situation data cube; performing feature normalization on the initial situation data cube and tensor decomposition to extract a low-dimensional core feature matrix as the standardized swarm situation tensor. This invention provides accurate and standardized data support for situation prediction and key splitting, ensuring the accuracy of subsequent calculations.

[0007] Optionally, the step of using a swarm temporal prediction model to predict the standardized swarm situation tensor to obtain a swarm prediction map of the UAV swarm includes: establishing the swarm temporal prediction model through a gated recurrent unit network; extracting temporal features from the standardized swarm situation tensor to obtain a predicted hidden state sequence; obtaining a predicted feature vector sequence of the nodes based on the predicted hidden state sequence; obtaining the situation prediction results of the nodes based on the predicted feature vector sequence, including motion trajectory prediction results, high-threat area intrusion probability, link quality degradation probability, and resource depletion probability; and establishing the swarm prediction map based on the situation prediction results using the node index, future time index, and prediction type index. This invention overcomes the limitations of traditional static situation awareness, providing accurate pre-judgment basis for resource quota allocation and key differentiation.

[0008] Optionally, the preset global resource ceiling constraint and the introduction of a resource auction mechanism to determine the individual-differentiated resource quotas of the nodes based on the swarm prediction graph include: preset the global resource ceiling constraint, including a global energy consumption ceiling constraint and a global computing overhead ceiling constraint; allocating a minimum budget to the nodes, including a minimum energy consumption budget and a minimum computing power budget; introducing the resource auction mechanism to deduct the minimum budget from the global resource ceiling constraint to form an auction pool; obtaining task urgency, resource abundance, and predicted risk value to generate the bidding price for the nodes; and allocating a budget from the auction pool based on the bidding price to determine the individual-differentiated resource quotas, including differentiated energy consumption quotas and differentiated computing power quotas. This invention helps improve the utilization rate of computing power and energy resources of UAV swarms and adapts to the operational needs of differentiated tasks.

[0009] Optionally, the step of obtaining the task urgency, resource abundance, and predicted risk value to generate the bidding price for the node includes: obtaining the task urgency by normalizing and weighting the remaining execution time and expected data volume of the node; obtaining the resource abundance by normalizing and weighting the available energy and available computing resources of the node; calculating the predicted risk value of the node based on the population prediction graph; and generating the bidding price by weighted summation of the task urgency, resource abundance, and predicted risk value. This invention aligns with actual task situations and operational risks, making resource bidding allocation results more scientific and matching dynamic task requirements and security needs.

[0010] Optionally, the step of constructing a key segmentation strategy table for the quantum key pool based on the population prediction graph and the individual-differentiated resource quota to obtain key segmentation fragments includes: calculating the security requirement weight of the node based on the population prediction graph; calculating the upper limit of the total key volume for the node based on the individual-differentiated resource quota; allocating the upper limit of the total key volume according to the security requirement weight in the time dimension to obtain the key allocation amount, as the time dimension segmentation result; allocating high-entropy key segments and low-entropy key segments to the node based on a preset risk threshold in the node dimension, as the node dimension segmentation result; and generating the key segmentation strategy table by combining the time dimension segmentation result and the node dimension segmentation result to segment the quantum key pool and obtain the key segmentation fragments. This invention adapts to the security risk level of each node, significantly improving the resource utilization rate of the quantum key pool.

[0011] Optionally, the step of distributing the key segmentation strategy table and the key segmentation fragments to the nodes for execution and collecting the execution feedback dataset of the UAV swarm includes: distributing the key segmentation strategy table and the key segmentation fragments to the nodes via a quantum-secure channel; in the local key cache, the nodes load the corresponding key segmentation fragments according to the key segmentation strategy table and establish a key segmentation index table; performing key rotation execution according to the key segmentation index table, and collecting the execution feedback dataset of the nodes. This invention provides comprehensive and realistic data support for model deviation verification and iterative optimization through execution feedback data.

[0012] Optionally, obtaining the deviation comparison result between the execution feedback dataset and the swarm prediction graph to update the swarm temporal prediction model and achieve segmentation closure includes: calculating the prediction deviation vector of the node based on the execution feedback dataset and the swarm prediction graph, as the deviation comparison result; and using the deviation comparison result to update the swarm temporal prediction model to obtain network update parameters, thereby achieving segmentation closure. This invention continuously reduces the deviation between situation prediction and actual operation, enabling the key segmentation strategy to dynamically adapt to the complex and ever-changing task scenarios of UAV swarms.

[0013] Optionally, updating the group time-series prediction model using the deviation comparison results to obtain network update parameters to achieve key segmentation closure includes: calculating the cross-correlation entropy of the deviation comparison results; determining a threshold for the cross-correlation entropy to trigger online fine-tuning of the group time-series prediction model; the online fine-tuning employs mini-batch gradient descent, using the execution feedback dataset as new training samples to update the group time-series prediction model to obtain the network update parameters; and using the network update parameters as initial parameters for the next round of prediction by the group time-series prediction model to achieve key segmentation closure. This invention improves the model's dynamic adaptability and prediction accuracy, ensuring the stability and efficiency of the key segmentation closure.

[0014] Secondly, this invention provides a task context-driven UAV swarm quantum key distribution system. The system executes the task context-driven UAV swarm quantum key distribution method provided by this invention. The system includes an input device, an output device, a processor, and a memory, which are interconnected. The memory stores a computer program, which includes program instructions, and the processor is configured to invoke the program instructions. This invention, through high-performance hardware collaboration, ensures stable execution of the method and facilitates business deployment and expansion. Attached Figure Description

[0015] Figure 1This is a flowchart of a task context-driven unmanned aerial vehicle swarm quantum key distribution method according to an embodiment of the present invention; Figure 2 This is a framework diagram of a task context-driven unmanned aerial vehicle swarm quantum key distribution system according to an embodiment of the present invention. Detailed Implementation

[0016] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0017] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0018] Please see Figure 1 One embodiment of the present invention provides a task context-driven quantum key distribution method for unmanned aerial vehicle swarms, the method comprising the following steps: S1. Obtain the task context data of each node in the drone swarm, and preprocess the task context data to generate a standardized swarm situation tensor.

[0019] Specifically, S1 includes the following steps: S11. Obtain the task context data, including node spatial location coordinates, velocity vector, qubit error rate, and remaining battery power.

[0020] In this embodiment, each drone node in the drone swarm collects its task context data in real time through its onboard sensor array and task management module.

[0021] Specifically, the spatial coordinates of the node are acquired by the airborne GPS receiver at a fixed sampling period, outputting the longitude, latitude, and altitude information of the UAV in three-dimensional space; the velocity vector is acquired by the inertial measurement unit, including the linear velocity components and attitude angular velocity components of the UAV in three axes; the quantum bit error rate (QBER) is statistically recorded in real time by the airborne quantum key distribution communication module during each quantum key negotiation, reflecting the degree of interference from environmental noise or potential eavesdropping in the current quantum channel; and the remaining battery power is monitored in real time by the airborne power management module.

[0022] In an optional embodiment, the task type code and task priority level are obtained as background information and pre-allocated by the ground control center through the task planning system and distributed to each node along with the task instructions. The task type code is used to identify the specific task category currently undertaken by the node, and the task priority level is used to characterize the criticality of the node in the entire task group.

[0023] Furthermore, to ensure the accuracy and reliability of the collected data, a confidence level verification is performed on the collected data. After verification, each node adds a collection timestamp and node identifier to the high-confidence task context data, packages it into a raw data message to be transmitted, and prepares to report it to the edge computing node or ground control center in a unified data format.

[0024] S12. Perform outlier removal and missing value imputation on the task context data to obtain a preprocessed dataset, so as to construct an initial situation data cube.

[0025] First, the task context data is grouped and sorted according to node identifiers and timestamps to form a raw data sequence that is aggregated by node and arranged in ascending order by time.

[0026] Subsequently, the outlier removal phase begins, employing differentiated outlier detection strategies for different data dimensions: For spatial location coordinates and velocity vectors, the system uses a sliding window-based local outlier detection method. This involves taking the current sampling point as the center, selecting several sampling points before and after it to form a local neighborhood, and calculating the local reachability density ratio of this point relative to other points in the neighborhood. If this ratio is lower than a preset local outlier threshold, the sampling point is determined to be an outlier and removed. For the quantum bit error rate (QBER), the system uses an adaptive threshold method based on historical statistics. This method calculates the mean and standard deviation using historical QBER data within a preset time window prior to the current moment. If the current QBER value deviates from the mean... If the degree of QBER fluctuation exceeds a preset multiple of the standard deviation (the preset multiple takes into account the normal fluctuation range of the quantum channel under different environmental conditions: QBER fluctuation is small in clear weather, so the preset multiple can be appropriately reduced to improve sensitivity; QBER fluctuation is large in severe weather or strong electromagnetic interference environment, so the preset multiple should be appropriately relaxed to avoid misjudgment), it is judged as an outlier and removed. For the remaining battery power, the system adopts the rate of change verification method to calculate the theoretical energy consumption rate limit based on the current node's task load type (such as hovering, cruising, acceleration maneuvering, etc.). If the power decrease in adjacent sampling intervals exceeds the product of the theoretical limit and the preset redundancy coefficient, the sampled value is judged as an outlier and removed.

[0027] Next, after outlier removal, the system fills in the removed data points and their adjacent missing values ​​using cubic spline interpolation. This method uses the numerical values ​​of valid data points before and after the missing value and the continuity condition of the first derivative to construct a piecewise cubic polynomial, and performs smooth interpolation between adjacent valid data points to maintain the continuity of data changes.

[0028] Finally, after obtaining the preprocessed dataset, the system organizes the initial situation data cube according to a three-dimensional structure: the first dimension is the node index, the second dimension is the context feature type, and the third dimension is the time window sequence.

[0029] Specifically, the system takes the current moment as the endpoint, extracts a historical time window of fixed length, arranges the preprocessed datasets of each node at each sampling moment within the time window into a two-dimensional matrix, with the rows of the matrix corresponding to the node index and the columns corresponding to the feature types; then, the two-dimensional matrices of multiple moments are stacked along the time dimension to form an initial situation data cube.

[0030] S13. Perform feature normalization processing on the initial situation data cube, and perform tensor decomposition to extract a low-dimensional core feature matrix as the standardized group situation tensor.

[0031] In this embodiment, after the initial situation data cube is constructed, the original values ​​of each feature dimension have significant differences in terms of units and ranges. Therefore, the system performs min-max normalization on each feature dimension in the initial situation data cube, mapping all feature values ​​to a unified numerical range, making each feature dimension comparable, while preserving the relative differences between nodes and the trend of change over time.

[0032] Furthermore, after feature normalization, a normalized situation data cube is obtained. Although this cube completely retains all feature information of all nodes, its dimensionality is high. Directly inputting it into the subsequent prediction model will face a large computational burden. Moreover, redundant information and noise inevitably exist in high-dimensional data. In order to extract low-dimensional core feature matrices and compress data volume, the system introduces tensor decomposition technology.

[0033] Specifically, a Higher-Order Singular Value Decomposition (HOSVD) method is employed to decompose the situation data cube into the product of a core tensor and three factor matrices. These three factor matrices correspond to the three modes of the data cube: node mode, feature mode, and temporal mode, each describing the principal component pattern in its corresponding dimension. The core tensor, through the projection of the three factor matrices, compresses the original high-dimensional data into a low-dimensional representation while preserving as much important information as possible from the original data. During the decomposition process, the system controls the degree of dimensionality reduction by setting the compression rank of each mode. The value of the compression rank is empirically preset based on the actual UAV swarm size, feature dimensions, and time window length, achieving a balance between reducing data dimensionality and preserving effective information.

[0034] In this embodiment, after decomposition, the system expands the core tensor along the time mode to obtain a low-dimensional core feature matrix. Each row of the matrix corresponds to the compressed situation feature of a node, and each column corresponds to the compressed situation dimension. This low-dimensional core feature matrix not only retains the key information reflecting the task situation in the original data, but also significantly reduces the input dimension of the subsequent prediction model. It is used as a standardized group situation tensor to provide high-quality input data for subsequent situation prediction.

[0035] S2. Use the swarm temporal prediction model to predict the standardized swarm situation tensor to obtain the swarm prediction map of the UAV swarm.

[0036] Specifically, S2 includes the following steps: S21. Establish the group temporal prediction model through a gated recurrent unit network, and extract the temporal features of the standardized group situation tensor to obtain the predicted hidden state sequence.

[0037] In this embodiment, the swarm temporal prediction model adopts a gated recurrent unit network (GRNN) as its core architecture. During the model building phase, the system pre-constructs a multi-layer GRNN. The input layer dimension of this network matches the dimension of the standardized swarm situation tensor in the feature modes. The hidden layer contains several GRNNs, which are sequentially connected according to time steps. The output layer is used to generate the predicted hidden state sequence. During the model training phase, the system uses historical situation data for supervised training of the network, using the actually observed future situation sequences as supervision labels. The time backpropagation algorithm is used to optimize the network parameters, enabling the network to learn the patterns of UAV swarm motion evolution and state changes from historical situation sequences.

[0038] Furthermore, after training, during the prediction phase, the system expands the standardized group situation tensor along the time dimension to form an input sequence arranged by time steps. This sequence is then sequentially input into the gated recurrent unit network. The network input at each time step is the situation feature vector of all nodes at the current moment. The network extracts temporal features from the input through the synergistic effect of update gates and reset gates: the update gate controls how much information from the hidden state at the previous moment needs to be retained in the current moment, and the reset gate controls how much information from the hidden state at the previous moment needs to be ignored to receive the current input. Together, they determine how the network's internal state evolves over time. After iterative calculations at all time steps, the network outputs the predicted hidden state vectors for each future moment to form a predicted hidden state sequence.

[0039] S22. Obtain the predicted feature vector sequence of the node based on the predicted hidden state sequence.

[0040] In this embodiment, the system connects a fully connected output layer to the output of the gated recurrent unit network. This output layer is a linear mapping layer without a nonlinear activation function. Its input is the predicted hidden state vector at each time step, and its output is the predicted feature vector at the corresponding time step. The weight matrix and bias vector of the fully connected output layer are optimized together with the gated recurrent unit network during the model training phase.

[0041] Specifically, the system inputs the predicted hidden state sequence into the fully connected output layer. The predicted hidden state vector at each time step is multiplied by the weight matrix and then the bias vector is added to obtain the predicted feature vector at that time step. Each component in the predicted feature vector corresponds one-to-one with the feature dimension in the standardized population state tensor, and corresponds to the predicted values ​​of physical quantities such as spatial position coordinates, velocity vector, qubit error rate, and remaining battery power.

[0042] Furthermore, by arranging the predicted feature vectors of all future moments in chronological order, a predicted feature vector sequence corresponding to the historical situation sequence is formed. The predicted feature vector of each moment in this sequence contains all the predicted situation information of each node at that moment, providing a complete data foundation for the subsequent extraction of various situation prediction results.

[0043] S23. Obtain the situation prediction results of the node based on the predicted feature vector sequence, including the motion trajectory prediction results, the probability of intrusion into high-threat areas, the probability of link quality degradation, and the probability of resource depletion.

[0044] First, regarding the motion trajectory prediction results: Spatial position coordinate components are directly extracted from the predicted feature vector to obtain the predicted spatial position coordinates of each node at each time. The predicted spatial position coordinates are organized by node index and future time index and stored as the position sequence of each node at future time, which serves as the motion trajectory prediction result.

[0045] Second, regarding the intrusion probability of high-threat areas: Based on the spatial boundary information of high-threat areas preset during the mission planning phase (stored in the form of spatial boundary equations), the system obtains the intrusion probability of high-threat areas based on the spatial topological relationship between the predicted spatial location coordinates of each node at each future time and the preset high-threat area boundary, satisfying the following relationship:

[0046] in, For the first The node at the th The probability of intrusion into a high-threat area at a future moment. For the Sigmoid function, These are the weighting coefficients. For the first The node at the th The symbolic distance from the predicted spatial location at a future moment to the boundary of the high-threat area. For the first The node at the th The projection component of the predicted velocity vector at a future time point onto the normal direction of the boundary of the high-threat area. For the first The node at the th The time interval between a future moment and the present moment.

[0047] It should be noted that the symbolic distance is positive when the predicted location is inside the threat region and negative when it is outside. It is calculated by substituting the predicted spatial coordinates into the spatial boundary equation of the threat region. The projection component is positive when the velocity direction points inside the threat region and negative when it points outside. It is obtained by calculating the dot product of the predicted velocity vector and the unit vector of the normal direction of the threat region boundary. Taking the reciprocal of the time interval ensures that nodes with more pressing time constraints receive a higher probability contribution.

[0048] Third, regarding the link quality degradation probability: The system extracts the predicted value of the qubit bit error rate from the predicted feature vector, and calculates the probability that the link will degrade in quality at a future time, thus failing to guarantee secure communication, by combining the link quality-related components. This probability of link quality degradation satisfies the following relationship:

[0049] in, For the first The node at the th The probability value of link quality degradation at a future time. For the Sigmoid function, These are the weighting coefficients. For the first The node at the th The predicted value of the bit error rate of a qubit at a future time. The preset secure communication bit error rate threshold, For the first The node at the th The predicted value of the bit error rate of a qubit at a future time. This represents the time step between two adjacent prediction times.

[0050] Fourth, regarding the probability of resource depletion: The system extracts the predicted value of the remaining battery power from the predicted feature vector, and calculates the probability that the battery power will be exhausted within the prediction window by combining the node's task load type information. This probability of resource depletion is obtained and satisfies the following relationship:

[0051] in, For the first The node at the th The probability value of resource depletion at a future moment. For the Sigmoid function, These are the weighting coefficients. The preset minimum safe operating power threshold, For the first The node at the th The predicted value of the remaining battery power at a future time. This represents the node's initial battery level. For the first The node at the th The predicted value of the remaining battery power at a future time. This represents the time step between two adjacent prediction times.

[0052] It should be noted that regarding the weighting coefficient ( , , All of these are preset positive numbers, and their specific values ​​are determined by fitting historical data or by experience. Furthermore, the weight coefficients within each formula satisfy the condition that the sum is 1.

[0053] S24. Based on the node index, future time index, and prediction type index, establish the population prediction map using the situation prediction results.

[0054] In this embodiment, the population prediction graph is a three-dimensional data structure with three dimensions: node dimension, future time dimension, and prediction type dimension.

[0055] During the organization process, the system first groups all prediction results according to the node index, and assigns all future time prediction results belonging to the same node to the data group corresponding to that node; then, within each node group, the prediction results for each time are sorted in ascending order of the future time index to form the prediction time series for that node; finally, for the prediction results of each node at each time, the corresponding prediction type position is filled in according to the prediction type index: the first layer is for the probability of high threat area intrusion, the second layer is for the probability of link quality degradation, and the third layer is for the probability of resource depletion.

[0056] Furthermore, the motion trajectory prediction results are stored separately as auxiliary information in the trajectory data field corresponding to each node at each time. They are associated with the above three types of probability data through node index and time index, and are not directly used as an independent layer in the prediction type dimension.

[0057] Through the data organization steps described above, the system obtains and stores a complete population prediction map.

[0058] S3. Preset the upper limit constraint of global resources and introduce a resource bidding mechanism to determine the individual resource quota of the node based on the group prediction map.

[0059] Specifically, S3 includes the following steps: S31. Preset the global resource upper limit constraint, including the global energy consumption upper limit constraint and the global computing overhead upper limit constraint.

[0060] In this embodiment, the system presets a global resource upper limit constraint based on the actual needs of the drone swarm to perform the current task and the hardware capabilities of each node.

[0061] The method for setting the upper limit of global energy consumption is as follows: The system calculates the total remaining battery power of all nodes in the current task phase, multiplies the total by the preset energy availability ratio coefficient, and obtains the upper limit of global energy consumption. This ratio coefficient is used to ensure that each node still retains a certain amount of safe return power reserve after the task is completed. Its specific value is dynamically adjusted within the preset range according to the task type and flight distance.

[0062] The method for setting the upper limit constraint of global computing overhead is as follows: the system calculates the total amount of remaining computing resources of all onboard processors of all nodes in the current task phase, multiplies the total by a preset computing resource availability ratio coefficient, and obtains the upper limit value of global computing overhead. This ratio coefficient is used to ensure that each node still has enough computing margin to meet the real-time computing needs of flight control and task execution while performing key management operations.

[0063] S32. Allocate a minimum budget to the node, including a minimum energy consumption budget and a minimum computing power budget.

[0064] In this embodiment, the allocation of the minimum budget is based on the mission criticality level of each node. This level is pre-set by the ground control center during the mission planning phase according to the importance of the sub-tasks undertaken by each node and is issued to each node along with the mission instructions.

[0065] Specifically, the system calculates the minimum allocation coefficient for each node based on the task criticality level of each node. The higher the task criticality level, the larger the minimum allocation coefficient.

[0066] For the minimum energy consumption budget, the system multiplies the upper limit constraint of the total energy consumption by the preset minimum reserve ratio to obtain the total minimum energy consumption budget. Then, according to the minimum allocation coefficient of each node, the total minimum energy consumption budget is allocated to each node to obtain the minimum energy consumption budget of each node.

[0067] For the guaranteed computing power budget, the system multiplies the upper limit constraint of the global computing overhead by the same guaranteed reserve ratio to obtain the total guaranteed computing power budget. Then, according to the guaranteed allocation coefficient of each node, the total guaranteed computing power budget is allocated to each node to obtain the guaranteed computing power budget for each node.

[0068] It should be noted that the minimum reserve ratio is determined based on the security requirements of the current task: for high-security tasks, the minimum reserve ratio is set to be larger to ensure that each node, especially critical nodes, has sufficient basic key management resources; for regular tasks, the minimum reserve ratio is set to be moderate to retain more resources for dynamic allocation in the auction pool.

[0069] S33. Introduce the resource bidding mechanism and deduct the guaranteed budget from the overall resource limit constraint to form a bidding pool.

[0070] In this embodiment, the resource auction mechanism is introduced as follows: The system first deducts the guaranteed energy consumption budget for all nodes from the overall energy consumption ceiling constraint, obtaining the remaining energy consumption budget as the energy consumption auction pool; then, it deducts the guaranteed computing power budget for all nodes from the overall computing overhead ceiling constraint, obtaining the remaining computing power budget as the computing power auction pool. The energy consumption auction pool and the computing power auction pool together constitute a complete auction pool, used to allocate resources to nodes that still have additional key resource requirements beyond the guaranteed budget.

[0071] S34. Obtain the task urgency, resource abundance, and predicted risk value to generate the bidding price for the node.

[0072] Specifically, S34 includes the following steps: S341. Based on the remaining execution time and expected data volume of the node's task, the urgency of the task is obtained by normalization and weighting.

[0073] In this embodiment, the system calculates the urgency of the task based on the remaining execution time and expected data volume of each node. The remaining execution time refers to the remaining time required for the current node to complete its assigned sub-tasks. This information is updated in real time by the ground control center according to the overall task plan and distributed to each node. The expected data volume refers to the total amount of encrypted data that the current node is expected to transmit during the subsequent task execution process.

[0074] The system normalizes the two original indicators mentioned above: for the remaining execution time of the task, it maps it to a preset time urgency range. The shorter the remaining time, the higher the normalized value, which indicates a greater time urgency. For the expected data volume, it maps it to a preset data load range. The larger the expected data volume, the higher the normalized value, which indicates a heavier communication load.

[0075] Furthermore, the two normalized indicators are weighted and summed to obtain the task urgency of the node. The value range of the task urgency is normalized to the interval [0,1]. The larger the value, the more urgent the current task of the node is and the more urgent the demand for key resources.

[0076] S342. Based on the available energy margin and available computing resource margin of the node, the resource abundance is obtained by normalization and weighting.

[0077] In this embodiment, the system calculates the resource sufficiency based on the available energy margin and available computing resource margin of each node; the available energy margin refers to the remaining battery power of the current node minus the minimum power reserve required to ensure a safe return; the available computing resource margin refers to the amount of idle computing resources of the current node's processor at the current moment.

[0078] Furthermore, the system uses the same normalization process and weighted summation as step S341 to obtain the resource abundance of the node. The value range of the resource abundance is normalized to the interval [0,1]. The larger the value, the more abundant the node's own resources are, and the lower the urgency of the need for external additional key resources.

[0079] S343. Calculate the predicted risk value of the node based on the population prediction map.

[0080] In this embodiment, the system extracts the probability of a node entering a high-threat area and the probability of link quality degradation at each future time from the group prediction map. The system then sums the two probabilities at each time of the same node and takes the average to obtain the average probability of the node entering and the average probability of link degradation within the prediction time window.

[0081] Furthermore, the predicted risk value of the node is obtained by weighted summing of the average intrusion probability and the average link degradation probability. The range of the predicted risk value is normalized to the interval [0,1]. The larger the value, the higher the comprehensive security risk faced by the node in the future task execution process, and the more key resources need to be allocated to ensure communication security.

[0082] S344. The bidding price is generated by weighted summation of the task urgency, the resource abundance, and the predicted risk value.

[0083] In this embodiment, resource abundance is taken as its normalized reciprocal: the higher the resource abundance, the smaller its reciprocal, indicating that the node has abundant resources and a lower willingness to require additional budget, thus lowering the bidding price accordingly; conversely, the lower the resource abundance, the larger its reciprocal, indicating that the node has scarce resources and an urgent need for additional budget, thus increasing the bidding price accordingly. Task urgency and predicted risk both contribute positively to the bidding price: the more urgent the task or the higher the predicted risk, the higher the bidding price.

[0084] The three indicators are weighted and summed according to preset weight coefficients to obtain the bidding price of the node. The bidding price of each node reflects its comprehensive need for additional key budget under the current task situation. It serves as the sole ranking basis for subsequent budget allocation from the auction pool. The higher the bidding price, the higher the priority of the node in the auction, so that the limited remaining budget is allocated to the node with the most urgent comprehensive needs.

[0085] The above bidding prices satisfy the following relationship:

[0086] in, For the bidding price, These are the weighting coefficients. For the first The task urgency of each node For the first The resource abundance of each node, It is a very small positive number. For the first The predicted risk value of each node.

[0087] It should be noted that the above weighting coefficients are preset positive numbers that sum to 1. Their specific values ​​are configured by the ground control center during the mission initialization phase based on the current mission type. For example, in time-sensitive missions... The highest value is chosen to highlight the urgency of the task in driving the bidding process, especially in security-sensitive tasks. The highest value is chosen to highlight the dominant role of risk prediction in resource-sensitive tasks. The highest value is selected to highlight the regulatory role of resource abundance on auction prices; each weight coefficient can also be periodically recalibrated based on historical execution data through offline learning during system operation to adapt to changing task scenarios.

[0088] S35. Allocate a budget from the auction pool based on the bidding price to determine the individual-differentiated resource quota, including differentiated energy consumption quota and differentiated computing power quota.

[0089] In this embodiment, after collecting the bidding prices reported by all nodes, the system sorts the nodes in descending order of bidding price to form a bidding priority queue.

[0090] Before the allocation begins, the system first checks the remaining budget of the energy consumption auction pool and the computing power auction pool. If the budget of either auction pool has been exhausted, the allocation process is terminated.

[0091] For each node in the bidding priority queue, the system allocates budget to that node sequentially according to the queue order. Specifically, the allocation method is as follows: based on the task urgency, predicted risk value, and resource abundance reported by the node during bidding, combined with the node's guaranteed budget, the system obtains the additional energy consumption budget and additional computing power budget required to ensure secure communication within the current predicted time window as the total demand. If the remaining budget in the current bidding pool can meet the node's total demand, the system allocates the full amount of the additional budget required by the node and deducts this amount from the remaining budget in the corresponding bidding pool. If the remaining budget in the current bidding pool is insufficient to meet the node's total demand, the system allocates the remaining budget in the bidding pool to the node according to the proportion of the node's demand in both energy consumption and computing power dimensions, and resets the remaining budget in the corresponding bidding pool to zero.

[0092] After allocation, the system records the actual amount of additional budget obtained by each node. The minimum budget of each node and the additional budget obtained through bidding are added together to form the individual differentiated resource quota of that node, namely the differentiated energy consumption quota and the differentiated computing power quota.

[0093] S4. Construct a key segmentation strategy table for the quantum key pool based on the group prediction graph and the individual difference resource quota to obtain key segmentation fragments.

[0094] Specifically, S4 includes the following steps: S41. Calculate the security requirement weight of the node based on the group prediction graph.

[0095] In this embodiment, the predicted risk value is non-linearly amplified through an exponential function, so that the high comprehensive risk moment receives a significantly higher security requirement weight, thereby tilting towards the high risk moment in the subsequent key distribution.

[0096] Furthermore, the system normalizes the security requirement weights so that the sum of the security requirement weights of the same node at each future time is 1, which facilitates proportional allocation in the future.

[0097] In this way, the system calculates a standardized security requirement weight for each node at each future time. The larger the weight value, the higher the security risk faced by the node at that time, and the more key resources need to be allocated.

[0098] S42. Calculate the upper limit of the total number of keys for the node based on the individual difference resource quota.

[0099] In this embodiment, after the system obtains the individual resource quotas of each node, it converts the differentiated energy consumption quota and differentiated computing power quota in the quota into the upper limit of key length.

[0100] Specifically, for differentiated energy consumption quotas, the system converts them into the upper limit of the first key length according to a preset energy consumption-key length conversion factor, satisfying the following relationship:

[0101] in, This is the upper limit of the first key length. This is the energy consumption-key length conversion factor. For differentiated energy consumption quotas.

[0102] It should be noted that the energy consumption-key length conversion factor represents the amount of key management operations that can be supported by a unit of energy consumption. Its value is dynamically adjusted according to the current communication distance and channel quality of the node. The greater the communication distance or the worse the channel quality, the lower the key length that can be supported by a unit of energy consumption.

[0103] Specifically, for differentiated computing power quotas, the system converts them into the second key length limit based on a preset computing power-key length conversion factor, satisfying the following relationship:

[0104] in, This is the upper limit of the second key length. This is the conversion factor between computing power and key length. Differentiated computing power quotas.

[0105] It should be noted that the computing power-key length conversion factor represents the amount of key material that can be processed per unit of computing power, and its value is determined based on the computational complexity of the key splitting algorithm.

[0106] Furthermore, the system uses the smaller of the first and second key length limits as the upper limit of the total number of keys for that node, to ensure that key management operations do not exceed the available resource limits in terms of both energy consumption and computing power.

[0107] S43. In the time dimension, the key allocation amount is obtained by allocating the upper limit of the total key amount according to the security requirement weight, which is used as the time dimension segmentation result.

[0108] In this embodiment, for each node, the system treats the security requirement weight of the node within the current prediction time window as a set of probability distributions, which reflect the relative demand for key resources of the node at each time. The system multiplies the upper limit of the total key amount by the security requirement weight at each time to obtain the initial key allocation amount at each time.

[0109] To prevent insufficient key allocation at certain times from failing to meet basic encrypted communication requirements, the system sets a minimum key allocation threshold for a single moment. If the initial key allocation at a certain moment is lower than this threshold, the system increases the key allocation for that moment to the threshold and deducts the corresponding difference proportionally from the key allocation at other moments for that node to ensure the total amount remains unchanged. Conversely, to prevent excessive key allocation at a single moment from overconsuming key resources, the system sets a maximum key allocation threshold for a single moment. If the initial key allocation at a certain moment exceeds this threshold, the excess portion is redistributed to other moments.

[0110] After the above adjustments, the key allocation amount of each node at each time moment constitutes the time dimension partitioning result. This result is organized by the node index and the future time index, and stored as the key length index corresponding to each node at each time moment.

[0111] S44. At the node level, high-entropy key fragments and low-entropy key fragments are allocated to the node based on a preset risk threshold, as the node-dimensional segmentation result.

[0112] In this embodiment, the system determines the overall risk level of a node within the entire prediction time window based on the predicted risk value of the node in the group prediction graph: if the predicted risk value of the node exceeds a preset risk threshold, the node is identified as a high-risk node and a high-entropy key fragment is allocated to it; if the predicted risk value of the node does not exceed the preset risk threshold, the node is identified as a low-risk node and a low-entropy key fragment is allocated to it.

[0113] It should be noted that the preset risk threshold is dynamically adjusted based on the security requirements of the current task. For high-security tasks, this threshold is set lower, allowing more nodes to be identified as high-risk and thus receive high-entropy key protection; for regular tasks, the threshold is set at a moderate value. High-entropy key fragments refer to key materials with longer key lengths and higher update frequencies, providing higher security strength and suitable for nodes facing higher threat risks. Low-entropy key fragments refer to key materials with shorter key lengths and lower update frequencies, meeting basic secure communication requirements while consuming fewer key resources and energy, and suitable for nodes facing lower threat risks.

[0114] During the allocation process, the system uses the task role of the node to make auxiliary judgments: for the master node undertaking the core task, even if its predicted risk value is lower than the preset risk threshold, the system still allocates a high-entropy key fragment to ensure the security of the core task; for the relay node undertaking the auxiliary task, the allocation is strictly based on the comparison result between the predicted risk value and the preset risk threshold.

[0115] After the above allocation is completed, the key fragment type of each node constitutes the node dimension splitting result, which clarifies whether each node obtains a high-entropy key or a low-entropy key within the entire prediction time window.

[0116] S45. Combine the time dimension segmentation result and the node dimension segmentation result to generate the key segmentation strategy table, and segment the quantum key pool to obtain the key segmentation fragment.

[0117] In this embodiment, the system traverses each node and each future time, using the key allocation amount of each node at each time as the allocation basis for the length dimension, and the key fragment type (high entropy or low entropy) corresponding to the node as the allocation basis for the type dimension. The two are combined to form a complete key splitting strategy record. This record contains four fields: node identifier, time identifier, key fragment length, and key fragment type, indicating that the node should obtain a key fragment of a specified length and type at that time. After summarizing the strategy records of all nodes at all times, a complete key splitting strategy table is formed.

[0118] Furthermore, after the key segmentation strategy table is generated, the quantum key pool is segmented. According to the key length specified in each record of the strategy table, the original key material of the corresponding length is extracted from the quantum key pool. According to the key type (high entropy or low entropy) specified in the strategy table, the extracted key material is post-processed accordingly: high entropy key fragments are updated with parameters at a higher frequency, and low entropy key fragments are updated with parameters at a lower frequency.

[0119] After extraction and configuration, each key material is encapsulated and packaged according to its corresponding node identifier and time identifier to form the key segmentation fragment corresponding to each node at each time.

[0120] S5. The key splitting strategy table and the key splitting fragments are sent to the node for execution, and the execution feedback dataset of the drone swarm is collected.

[0121] Specifically, S5 includes the following steps: S51. The key segmentation strategy table and the key segmentation fragments are sent to the node through a quantum-secure channel.

[0122] In this embodiment, the quantum secure channel is established as follows: the system pre-negotiates and generates a pair of symmetric quantum session keys between the ground control center and each UAV node through a quantum key distribution protocol, which serve as the transmission encryption keys for this distribution process.

[0123] Before the distribution begins, the system encapsulates the key segmentation strategy table and key segmentation fragments according to a preset data packet format. The strategy table is encapsulated as a control message, and the key fragments are encapsulated as data messages. The two are distinguished by different message type identifiers to facilitate identification and processing by the receiving end.

[0124] After encapsulation, the system uses a quantum session key to encrypt the encapsulated message and sends the ciphertext to each node via a classical data link. After receiving the ciphertext, each node uses the same quantum session key stored locally to decrypt it and restore the key splitting strategy table and key splitting fragments.

[0125] During the transmission process, the system verifies the integrity of the message and attaches a message authentication code generated based on the quantum session key to the message. The receiving end verifies the authentication code to confirm that the message has not been tampered with during transmission.

[0126] For nodes that fail to receive data, the system determines the receiving result based on the receiving status feedback reported by the node and triggers a retransmission mechanism: when a node is detected to have failed to receive data or has not responded after a timeout, the system retransmits the corresponding policy table and key fragment so that all nodes can receive the data successfully.

[0127] S52. In the local key cache area, the node loads the corresponding key segmentation fragment according to the key segmentation strategy table and establishes a key segmentation index table.

[0128] In this embodiment, the local key cache is a dedicated storage area pre-allocated in the onboard storage device of each node, used to store key fragments that are currently in use and will be used in the future.

[0129] During the loading process, the node stores the corresponding key segmentation fragments into the designated location in the cache, one by one, according to the field content of each record in the key segmentation strategy table.

[0130] Specifically, the node determines whether the record belongs to its own node based on the node identifier field in the policy table. If it does, it extracts the time identifier, key fragment length, and key fragment type information from the record, determines the storage location index of the key fragment in the cache based on the time identifier, and writes the key fragment to that location.

[0131] In an optional embodiment, after the writing is completed, a data consistency check is performed by rereading the written key fragment and comparing it byte by byte with the original fragment sent. If the comparison is consistent, it is marked as successful loading; otherwise, it is marked as failed loading and a retransmission request is triggered.

[0132] After loading is complete, a key fragment index table is created for the node. The key fragment index table is a data structure stored in the onboard memory. Each row corresponds to a loaded key fragment and contains three fields: the effective time field records the time when the key fragment becomes effective, the key identifier field points to the specific storage address of the corresponding key fragment in the cache, and the lifespan field records the effective duration of the key fragment.

[0133] It should be noted that the key fragment index table is sorted in chronological order of its effective time, enabling nodes to quickly locate the key fragment to be used at the current time and the key fragment to be switched to at the next time, providing efficient query support for subsequent key rotation execution.

[0134] S53. Perform key rotation execution according to the key fragment index table, and collect the execution feedback dataset of the node.

[0135] In this embodiment, key rotation refers to the node switching the key segment it uses according to a preset time window during communication. When the current window is about to end, the node loads the key segment from the cache into the working register in advance according to the effective time and storage address of the next key segment recorded in the index table, and automatically switches to the new key at the moment the current window ends, so as to achieve zero-delay key rotation and ensure the continuity and security of key use during communication.

[0136] During the key rotation process, the system synchronously collects execution feedback datasets from each node. The actual intrusion flag is determined by the node based on threat signals detected by airborne sensors or threat area notifications issued by the ground control center. If the node's actual spatial location at the current moment is within a preset high-threat area, the actual intrusion flag is recorded as 1; otherwise, it is recorded as 0. The actual qubit error rate is statistically recorded in real time by the node's airborne quantum key distribution communication module during each quantum key negotiation. The actual remaining battery power is obtained in real time by the airborne power management module. All feedback data are correlated according to timestamps and node identifiers to form an execution feedback dataset.

[0137] S6. Obtain the deviation comparison result between the execution feedback dataset and the population prediction map to update the population time series prediction model and realize the segmentation closure.

[0138] Specifically, S6 includes the following steps: S61. Based on the execution feedback dataset and the population prediction map, calculate the prediction deviation vector of the node as the deviation comparison result.

[0139] In this embodiment, the system compares the actual intrusion identifier in the execution feedback dataset with the intrusion probability of high-threat areas in the swarm prediction map, and calculates the difference as the threat prediction bias; it obtains the actual link degradation probability based on the actual qubit error rate, and compares it with the link quality degradation probability in the swarm prediction map, calculating the difference as the link quality prediction bias; it obtains the actual resource depletion probability based on the actual remaining battery power, and compares it with the resource depletion probability in the swarm prediction map, calculating the difference as the resource prediction bias. These three bias values ​​are organized according to node index and time index to form a three-dimensional prediction bias vector for each node at each time moment, which serves as the bias comparison result.

[0140] S62. The deviation comparison results are used to update the group time series prediction model to obtain network update parameters, so as to realize the segmentation closure.

[0141] Specifically, S62 includes the following steps: S621. Calculate the cross-correlation entropy of the deviation comparison results, and perform a threshold determination on the cross-correlation entropy to trigger online fine-tuning of the population time series prediction model.

[0142] In this embodiment, for the prediction deviation vector of any node at any time, the system uses a preset kernel width parameter as the bandwidth and a Gaussian kernel function to convert the deviation values ​​of each dimension in the deviation vector into similarity scores in the (0,1) interval. Then, the mean of the similarity scores of each dimension is taken as the cross-correlation entropy of the prediction deviation vector, satisfying the following relationship:

[0143] in, For cross-correlation entropy, For the prediction bias vector, For the dimension index of the predicted bias vector, For the first The components of the prediction bias vector. This is the kernel width parameter.

[0144] It should be noted that the cross-correlation entropy ranges from (0,1). The closer the value is to 1, the smaller the prediction bias, that is, the closer the predicted value is to the actual value, and the higher the accuracy of the prediction model. The farther the value is from 1, the larger the prediction bias, that is, the greater the deviation between the predicted value and the actual value.

[0145] In this embodiment, the value of the kernel width parameter directly affects the sensitivity of the cross-correlation entropy to the deviation: when the kernel width is small, the cross-correlation entropy is more sensitive to changes in the deviation, and even a small deviation can lead to a significant decrease in the cross-correlation entropy value; when the kernel width is large, the cross-correlation entropy changes more smoothly to the deviation.

[0146] Specifically, the system dynamically adjusts the kernel width parameter based on the current task's security requirement level: in high-security tasks, a smaller kernel width is used, making the cross-correlation entropy more sensitive to deviations, thereby triggering more frequent model updates to maintain prediction accuracy; in regular tasks, a moderate kernel width is used, ensuring prediction accuracy while reducing unnecessary update overhead.

[0147] After the calculation is completed, the system compares the cross-correlation entropy value with a preset threshold: if the cross-correlation entropy value of a certain deviation vector is lower than the preset threshold, it indicates that there is a significant deviation between the actual execution result and the prediction result of the node at that time, and the system determines the deviation vector as a valid trigger sample; if the cross-correlation entropy value is not lower than the preset threshold, it indicates that the prediction deviation is within an acceptable range, and the system does not trigger the model update.

[0148] It should be noted that the preset threshold value is dynamically adjusted according to the security requirement level of the current task. In high security requirement tasks, the threshold value is higher, so that more deviation vectors are judged as significant deviations, thereby triggering more frequent model updates to ensure prediction accuracy. In regular tasks, the threshold value is moderate.

[0149] S622. The online fine-tuning adopts the mini-batch gradient descent method, using the execution feedback dataset as new training samples to update the population time series prediction model, so as to obtain the network update parameters.

[0150] In each fine-tuning iteration, the system randomly selects a preset number of samples (such as 32 or 64) from the newly added training sample set to form a mini-batch. The mini-batch samples are input into the current model, the prediction error between the model output and the actual observation value is calculated, and then the gradient of the prediction error with respect to each learnable parameter in the model (including the weight matrix and bias vector of each time step in the gated recurrent unit network, and the weight matrix and bias vector of the fully connected output layer) is calculated through the backpropagation algorithm. Each gradient is multiplied by the learning rate and then subtracted from the current parameter value to complete the parameter update of the mini-batch.

[0151] The system repeats the above random sampling and parameter update process until all samples in the newly added training sample set have been used for training, completing one round of online fine-tuning and obtaining the network update parameters. If the number of samples in the newly added training sample set is small and insufficient to form multiple mini-batches, the system updates all new samples as a single mini-batch to make full use of the limited feedback data to correct the model.

[0152] S623. Use the network update parameters as the initial parameters for the next round of prediction of the population time series prediction model to achieve segmentation and closure.

[0153] In this embodiment, the system packages the updated network parameters after the gated recurrent unit network update, including the weight matrix and bias vector at each time step, and the weight matrix and bias vector of the fully connected output layer, into a complete model parameter file according to their structural organization in the model, writes it into the storage location corresponding to the model in the parameter library, and records the current parameter version number and timestamp.

[0154] At the start of the next prediction cycle, the population time series prediction model loads the network update parameters as initial parameters, replacing the original parameters to predict the standardized population situation tensor at the current moment, thus achieving closed-loop iteration.

[0155] Please see Figure 2 In an optional embodiment, the present invention provides a task context-driven UAV swarm quantum key distribution system. The system includes an input device, an output device, a processor, and a memory, all interconnected. The memory stores a computer program comprising program instructions, and the processor is configured to invoke the program instructions to execute specific steps as described in the relevant embodiments of the task context-driven UAV swarm quantum key distribution method provided by the present invention. The task context-driven UAV swarm quantum key distribution system provided by the present invention is structurally complete, objectively stable, and enhances the overall applicability and practical application capabilities of the present invention.

[0156] In summary, the present invention provides a task context-driven quantum key distribution method and system for UAV swarms. This method acquires the task context data of each node in the UAV swarm and preprocesses it to generate a standardized swarm situation tensor. A swarm temporal prediction model is then used to predict the future situation and obtain a swarm prediction graph. Based on this, a global resource upper limit constraint is preset, and a resource auction mechanism is introduced to determine the individual resource quotas for each node based on the prediction graph. A key distribution strategy table for the quantum key pool is then constructed to obtain key distribution fragments, which are then distributed for execution. Finally, execution feedback data is collected and compared with the swarm prediction graph to update the prediction model and achieve closed-loop iteration. The method is easy to understand, computationally simple, requires minimal workload, and is suitable for engineering applications, providing a theoretical foundation and technical support for the further development of quantum communication technology.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A task context driven UAV swarm quantum key splitting method, characterized in that, Includes the following steps: The task context data of each node in the drone swarm is obtained, and the task context data is preprocessed to generate a standardized swarm situation tensor. The standardized swarm situation tensor is predicted using a swarm temporal prediction model to obtain the swarm prediction map of the UAV swarm. A preset global resource ceiling constraint is established, and a resource auction mechanism is introduced to determine the individual resource quota of the node based on the population prediction graph. Based on the group prediction graph and the individual difference resource quota, a key segmentation strategy table for the quantum key pool is constructed to obtain key segmentation fragments; The key splitting strategy table and the key splitting segments are sent to the nodes for execution, and the execution feedback dataset of the drone swarm is collected. Obtain the deviation comparison results between the execution feedback dataset and the population prediction map to update the population time series prediction model and achieve segmentation closure.

2. The task context driven UAV swarm quantum key splitting method of claim 1, wherein, The step of acquiring the task context data of each node in the drone swarm and preprocessing the task context data to generate a standardized swarm situation tensor includes: Obtain the task context data, including node spatial location coordinates, velocity vector, qubit error rate, and remaining battery power; The task context data is subjected to outlier removal and missing value imputation to obtain a preprocessed dataset, which is then used to construct an initial situation data cube; The initial situation data cube is subjected to feature normalization processing, and tensor decomposition is performed to extract a low-dimensional core feature matrix as the standardized group situation tensor.

3. The task context-driven UAV swarm quantum key distribution method according to claim 1, characterized in that, The step of using a swarm temporal prediction model to predict the standardized swarm situation tensor to obtain the swarm prediction map of the UAV swarm includes: The population temporal prediction model is established by gated recurrent unit network, and the predicted hidden state sequence is obtained by extracting temporal features from the standardized population situation tensor. The predicted feature vector sequence of the node is obtained based on the predicted hidden state sequence; The situation prediction results of the node are obtained based on the predicted feature vector sequence, including the motion trajectory prediction results, the probability of intrusion into high-threat areas, the probability of link quality degradation, and the probability of resource depletion. Based on the node index, future time index, and prediction type index, the population prediction map is established using the situation prediction results.

4. The task context-driven UAV swarm quantum key distribution method according to claim 1, characterized in that, The preset global resource upper limit constraint, and the introduction of a resource auction mechanism, determine the individual-differentiated resource quotas of the nodes based on the population prediction graph, including: The preset global resource upper limit constraints include global energy consumption upper limit constraints and global computing overhead upper limit constraints; Allocate a minimum budget to the node, including a minimum energy consumption budget and a minimum computing power budget; The resource bidding mechanism is introduced to deduct the minimum budget from the overall resource limit constraint to form a bidding pool; The bidding price for the node is generated by obtaining the task urgency, resource abundance, and predicted risk value. Budgets are allocated from the auction pool based on the bidding price to determine the individualized resource quotas, including differentiated energy consumption quotas and differentiated computing power quotas.

5. The task context-driven UAV swarm quantum key distribution method according to claim 4, characterized in that, The process of generating the bidding price for the node by obtaining the task urgency, resource abundance, and predicted risk value includes: The task urgency is obtained by normalizing and weighting the remaining execution time and expected data volume of the node. The resource abundance is obtained by normalizing and weighting the available energy and available computing resources of the node. The predicted risk value of the node is calculated based on the population prediction map; The bidding price is generated by weighting and summing the task urgency, resource abundance, and predicted risk value.

6. The task context-driven UAV swarm quantum key distribution method according to claim 1, characterized in that, The key segmentation strategy table for constructing a quantum key pool based on the population prediction graph and the individual difference resource quota, to obtain key segmentation fragments, includes: Calculate the security requirement weight of the node based on the group prediction graph; The maximum total number of keys for the node is calculated based on the individual resource quota. In the time dimension, the key allocation amount is obtained by allocating the upper limit of the total key amount according to the security requirement weight, which serves as the time dimension segmentation result; At the node level, high-entropy key fragments and low-entropy key fragments are allocated to the nodes based on a preset risk threshold, which serve as the node dimension segmentation results. The key segmentation strategy table is generated by combining the time dimension segmentation result and the node dimension segmentation result, and the quantum key pool is segmented to obtain the key segmentation fragment.

7. The task context-driven UAV swarm quantum key distribution method according to claim 1, characterized in that, The step of distributing the key segmentation strategy table and the key segmentation fragments to the node for execution, and collecting the execution feedback dataset of the drone swarm, includes: The key segmentation strategy table and the key segmentation fragments are sent to the node via a quantum-secure channel; In the local key cache, the node loads the corresponding key segmentation fragment according to the key segmentation strategy table and establishes a key segmentation index table; The key is rotated according to the key fragment index table, and the execution feedback dataset of the node is collected.

8. The task context-driven UAV swarm quantum key distribution method according to claim 1, characterized in that, The step of obtaining the deviation comparison result between the execution feedback dataset and the population prediction map to update the population time series prediction model and achieve segmentation closure includes: Based on the execution feedback dataset and the population prediction map, the prediction deviation vector of the node is calculated as the deviation comparison result; The deviation comparison results are used to update the population time series prediction model to obtain network update parameters, so as to achieve segmentation closure.

9. The task context-driven UAV swarm quantum key distribution method according to claim 8, characterized in that, The step of updating the population time series prediction model using the deviation comparison results to obtain network update parameters, in order to achieve segmentation closure, includes: Calculate the cross-correlation entropy of the deviation comparison results, and apply a threshold to the cross-correlation entropy to trigger online fine-tuning of the population time series prediction model; The online fine-tuning uses mini-batch gradient descent to update the population time-series prediction model with the execution feedback dataset as new training samples, in order to obtain the network update parameters. The network update parameters are used as the initial parameters for the next round of prediction in the population time series prediction model to achieve segmentation and closure.

10. A task context-driven unmanned aerial vehicle swarm quantum key distribution system, characterized in that, The system includes an input device, an output device, a processor, and a memory, which are interconnected. The memory stores a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute the task context-driven UAV swarm quantum key distribution method as described in any one of claims 1-9.

Images