Unmanned aerial vehicle cluster test method and system

Abstract

The invention provides an unmanned aerial vehicle cluster test method and system, and the method comprises the steps: jointly executing a first attack strategy for all unmanned aerial vehicles in a target unmanned aerial vehicle cluster in different time step ranges, and predicting a first global operation state sequence after the target unmanned aerial vehicle cluster executes each first attack strategy based on a predictor, determining a target time step range in the plurality of time step ranges; independently executing a second attack strategy on different unmanned aerial vehicles in the target unmanned aerial vehicle cluster within the target time step range, and predicting a second global operation state sequence after the target unmanned aerial vehicle cluster executes each second attack strategy based on the predictor, so as to determine at least one target unmanned aerial vehicle, and according to the target time step range and each target unmanned aerial vehicle, performing an anti-attack test on the target unmanned aerial vehicle cluster in a real test environment. According to the invention, the attack range is narrowed, the anti-attack concealment is improved, the lightweight black box anti-attack test is realized, and the calculation cost is reduced.

Description

Technical Field

[0001] The present invention relates to the field of safety protection technology, and in particular to a drone cluster testing method and system. Background Art

[0002] Drone swarms are playing an increasingly important role in both real life and the military. However, the dynamic and complex environments in which they operate, coupled with the collaboration and confrontation of multiple drones, pose significant challenges to their coordinated security and stability. Therefore, developing effective drone swarm testing methods and evaluating their performance are crucial for ensuring the safe and stable operation of drone swarms.

[0003] At present, some scholars have explored the technology of drone swarm testing in multi-agent systems, but the adversarial attacks in such testing methods often cover the entire testing time steps in the testing process, resulting in poor concealment of the attack and mismatch with actual application scenarios. In addition, the adversarial attacks in such testing methods are usually white-box attacks, resulting in high computational costs.

[0004] Therefore, there is an urgent need to provide a drone cluster testing method and system to solve the above technical problems. Summary of the Invention

[0005] The present invention provides a UAV cluster testing method and system to solve the defects in the prior art.

[0006] The present invention provides a UAV cluster testing method, comprising: Within different time step ranges, jointly executing a first attack strategy on all drones in the target drone cluster, and predicting a first global operating state sequence of the target drone cluster after each execution of the first attack strategy based on a predictor; Determining a target time step range among the plurality of time step ranges according to the first global operating state sequence; Within the target time step range, sequentially executing a second attack strategy on different drones in the target drone cluster, and predicting, based on the predictor, a second global operating state sequence of the target drone cluster after each execution of the second attack strategy; According to the second global operating state sequence, at least one target drone is determined in the target drone cluster, and based on the target time step range and each target drone, an adversarial attack test is performed on the target drone cluster in a real test environment to obtain an adversarial attack test result of the target drone cluster.

[0007] According to a drone cluster testing method provided by the present invention, the predictor is constructed based on the following steps: Obtaining a sample operating state of a sample UAV cluster, a sample operating action corresponding to the sample operating state, a true result of the next operating state corresponding to the sample operating action, and a true result of the global operating state sequence corresponding to the sample operating action; Inputting the sample running state and the sample running action into the initial neural network model for single-step prediction to obtain a next running state prediction result corresponding to the sample running action; Inputting the sample running state and the sample running action into the initial neural network model for global prediction, and obtaining a global running state sequence prediction result corresponding to the sample running action; Calculating a first loss value according to the actual result of the next operating state and the predicted result of the next operating state; Calculating a second loss value according to the actual result of the global operating state sequence and the predicted result of the global operating state sequence; The initial neural network model is iteratively trained according to the first loss value and the second loss value to obtain the predictor.

[0008] According to a drone cluster testing method provided by the present invention, the initial neural network model is constructed according to the task type of the task performed by the target drone cluster, and the task type includes a positioning search task or a confrontation combat task.

[0009] According to a drone cluster testing method provided by the present invention, determining a target time step range in a plurality of time step ranges based on the first global operating state sequence includes: Based on the predictor, in the initial operating state, predicting a third global operating state sequence after all drones in the target drone cluster jointly execute a normal strategy; Calculating, based on each first termination operating state and a reference state, a first attack impact degree after the target UAV cluster executes each of the first attack strategies; each first termination operating state is a termination operating state in each of the first global operating state sequences, and the reference state is a termination operating state in the third global operating state sequence; The target time step range is determined within the plurality of time step ranges according to each of the first attack impacts.

[0010] According to a drone cluster testing method provided by the present invention, the first attack impact degree after the target drone cluster executes each of the first attack strategies is calculated based on each first termination operation state and a reference state, including: Determining a target evaluation function according to the task type of the task performed by the target UAV cluster; Based on the target evaluation function, respectively evaluating the task completion degree corresponding to each of the first terminated running states and the task completion degree corresponding to the reference state; A difference calculation is performed based on the task completion degree corresponding to each of the first termination operation states and the task completion degree corresponding to the baseline state to obtain the first attack impact degree after the target drone cluster executes each of the first attack strategies.

[0011] According to a drone cluster testing method provided by the present invention, determining at least one target drone in the target drone cluster according to the second global operating state sequence includes: Based on the predictor, in the initial operating state, predicting a third global operating state sequence after all drones in the target drone cluster jointly execute a normal strategy; Calculating, according to each second termination operating state and a reference state, a second attack impact degree after the target UAV cluster executes each second attack strategy; each second termination operating state is a termination operating state in each second global operating state sequence, and the reference state is a termination operating state in the third global operating state sequence; Determining an attack priority level corresponding to each of the drones according to each of the second attack impacts and the total number of time steps of each of the second global operating state sequences; At least one target drone is determined in the target drone cluster according to the attack priority level corresponding to each drone.

[0012] According to a drone cluster testing method provided by the present invention, determining the attack priority level corresponding to each drone based on each second attack impact and the total number of time steps of each second global operating state sequence includes: For each drone, when the second attack impact degree corresponding to the drone is greater than the second attack impact degree corresponding to any other drone, the attack priority level corresponding to the drone is determined to be higher than the attack priority level corresponding to any other drone; Alternatively, when the second attack influence degree corresponding to the drone is the same as the second attack influence degree corresponding to any other drone, and the total number of time steps of the second global operating state sequence corresponding to the drone is greater than the total number of time steps of the second global operating state sequence corresponding to any other drone, it is determined that the attack priority level corresponding to the drone is higher than the attack priority level corresponding to any other drone.

[0013] The present invention also provides a UAV cluster testing system, comprising: A first prediction unit is configured to jointly execute a first attack strategy on all drones in a target drone cluster at different time steps, and predict, based on a predictor, a first global operating state sequence of the target drone cluster after each execution of the first attack strategy; A search unit, configured to determine a target time step range from among the plurality of time step ranges according to the first global operating state sequence; A second prediction unit is configured to sequentially execute a second attack strategy on different drones in the target drone cluster within the target time step range, and predict, based on the predictor, a second global operating state sequence of the target drone cluster after each execution of the second attack strategy; The testing unit is used to determine at least one target drone in the target drone cluster according to the second global operating state sequence, and perform an adversarial attack test on the target drone cluster in a real test environment according to the target time step range and each target drone, to obtain an adversarial attack test result of the target drone cluster.

[0014] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the drone cluster testing method described above is implemented.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described drone cluster testing methods.

[0016] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-described drone cluster testing methods.

[0017] The drone cluster testing method and system provided by the present invention construct a predictor that can simulate an environment and interact with drones in the drone cluster, and predict the effects of the target drone cluster executing the attack strategy within different time step ranges based on the predictor to determine the target time step range, and combine the predictor to predict the effects of different drones executing the attack strategy within the target time step range, determine at least one target drone, and execute the attack strategy on the target drone cluster in a real test environment based on the target time step range and the at least one target drone, so as to induce the final state of the target drone cluster to deviate from the expected state, thereby realizing an adversarial attack test on the target drone cluster. Therefore, by combining the predictor to determine the optimal attack time range and the optimal attack target during the test process to optimize the generation of the attack strategy, not only the attack range is reduced from the two dimensions of shortening the attack time step length and reducing the number of attacked drones while ensuring the effectiveness of the attack, thereby improving the concealment of the adversarial attack, but also avoiding the use of reinforcement learning algorithms to generate adversarial strategies, realizing lightweight black box adversarial attack testing, and reducing computational costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction is given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 This is one of the flow charts of the UAV cluster testing method provided by the present invention.

[0020] Figure 2 This is the second flow chart of the UAV cluster testing method provided by the present invention.

[0021] Figure 3 Schematic diagram of the workflow of the predictor provided by the present invention.

[0022] Figure 4 It is a schematic diagram of the performance evaluation results of the single-step prediction of the predictor provided by the present invention.

[0023] Figure 5 Schematic diagram of the performance evaluation results of the global prediction of the predictor provided by the present invention.

[0024] Figure 6 It is a structural schematic diagram of the UAV cluster testing system provided by the present invention.

[0025] Figure 7 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0026] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0027] Drone swarms are playing an increasingly important role in both real life and the military. When performing missions, drone swarms often face dynamic and complex environments, involving the collaboration and confrontation of multiple drones. Therefore, improving the stability, adaptability, and defense capabilities of drone swarms in adversarial environments has become a bottleneck in the development of drone swarm technology. For example, in military confrontation missions, drone swarms must not only complete their missions but also withstand interference and attacks from the enemy. Enemies can influence the behavior of drone swarms through electronic warfare, cyberattacks, and the dissemination of disinformation, thereby weakening their combat capabilities. Therefore, developing effective drone swarm testing methods and evaluating their performance (such as vulnerability and robustness) is crucial to ensuring the safe and stable operation of drone swarms.

[0028] Reinforcement learning (RL) is a class of machine learning methods for solving sequential decision-making problems. It enables intelligent agents to gradually learn how to select optimal action strategies to maximize cumulative rewards through interaction with their environment. In RL, agents interact with their environment to obtain reward signals and then adjust their decision-making strategies by optimizing these reward signals. In multi-agent systems (MAS), RL is widely used for autonomous decision-making and collaboration among agents. Each agent uses RL algorithms to learn how to make appropriate decisions in different situations to achieve mission objectives. In multi-agent reinforcement learning (MARL), each agent must consider not only its own actions but also the behavior of other agents to achieve overall system optimization. Drone swarms, as a manifestation of multi-agent systems, also use RL algorithms for policy learning. This means that each drone can make autonomous decisions based on perceived information and predetermined goals, and collaborate with other drones in the swarm.

[0029] Therefore, some researchers have explored techniques for testing drone swarms in multi-agent systems. However, on the one hand, the adversarial attacks in such testing methods often cover the entire testing time step, making the attack behavior easy to detect, resulting in poor concealment and mismatching with actual application scenarios. On the other hand, such testing methods are usually based on reinforcement learning-based adversarial attacks, which are essentially white-box attacks and result in high computational costs. Therefore, how to design a covert, efficient, and adaptive adversarial attack testing technology for drone swarms in multi-agent systems that can adapt to different strategies is a current research focus.

[0030] In this regard, the present application provides a drone cluster testing method to discover the weaknesses of the drone cluster from the perspective of a real attacker and promote the enhancement of the robustness of the drone cluster. The method specifically constructs a predictor that can simulate the environment and interact with the drones in the drone cluster, and predicts the effect of the target drone cluster executing the attack strategy within different time step ranges based on the predictor to determine the target time step range (also known as the optimal attack time range), and combines the predictor to predict the effect of different drones executing the attack strategy within the target time step range, to determine at least one target drone (also known as the optimal attack target range), and based on the optimal attack time range and the optimal attack target range, In a real test environment, an attack strategy is executed on the target drone cluster to induce the final state of the target drone cluster to deviate from the expected state, thereby realizing the adversarial attack test on the target drone cluster. Therefore, during the test process, the generation of attack strategies is optimized by automatically determining the optimal attack time range and the optimal attack target range by combining the predictor and the rule-based heuristic search algorithm. This not only reduces the attack range from the two dimensions of shortening the attack time step length and reducing the number of attacked drones, thereby achieving the purpose of improving the concealment of the adversarial attack, but also avoids the use of reinforcement learning algorithms to generate adversarial strategies, realizing lightweight black-box adversarial attack testing and reducing computational costs.

[0031] Figure 1 This is one of the flow charts of the drone cluster testing method provided by the present invention; this method can be applied to drone cluster testing in various mission scenarios, such as positioning and search mission scenarios, combat mission scenarios, reconnaissance mission scenarios, rescue mission scenarios, etc., and this embodiment does not specifically limit this.

[0032] like Figure 1 As shown, the method includes step 110 , step 120 , step 130 and step 140 .

[0033] Step 110 , jointly executing a first attack strategy on all drones in the target drone cluster at different time step ranges, and predicting a first global operating state sequence of the target drone cluster after executing the first attack strategy each time based on a predictor.

[0034] The target drone cluster here is the drone cluster on which the adversarial attack test is required, which includes multiple drones.

[0035] The predictor here is used to provide feedback on the running actions performed by the drone cluster in a given running state by simulating a real test environment, so as to predict the global running state sequence of the drone cluster after the running action is performed. The global running state sequence refers to the complete state sequence formed by the combination of the running states of the drone cluster at multiple different time steps after the running action is performed.

[0036] Optionally, when it is determined that the test task of the target drone cluster is triggered, such as when a test instruction of the target drone cluster input by the user is received, or when it is detected that the target drone cluster enters a specific combat area or performs a specific task, multiple different time step ranges are obtained.

[0037] Here, the multiple different time step ranges have different starting time steps and / or different time step lengths, which can be pre-set and generated, that is, the starting time and length of each time step range can be manually set according to task requirements; or a given time step sequence can be generated by multiple sliding windows according to multiple different window sizes, such as traversing each time step in a given time step sequence, and intercepting m time steps with each time step as the starting point to form a time step range, where m can dynamically change from small to large within a certain numerical range, and this embodiment does not specifically limit this.

[0038] After obtaining multiple different time step ranges, the multiple time step ranges can be traversed, and a first attack strategy can be jointly executed on all drones in the target drone cluster within each time step range. The first attack strategy can be a strategy of executing m-step attacks on all drones in the target drone cluster, where m is the time step length of each time step range. The attack in the first attack strategy can be achieved by adding disturbances (such as environmental noise, communication interference, etc.) to the environment in which all drones in the target drone cluster operate, or directly replacing the running actions output by the target policy network corresponding to the target drone cluster for all drones with attack actions. The specific attack strategy can be adaptively determined based on the type of task performed by the target drone cluster. For example, when the target drone cluster is performing a confrontation combat task, the running actions output by the target policy network corresponding to the target drone cluster for all drones can be replaced with attack actions to attack all drones in the target drone cluster. For another example, when the target drone cluster is performing a positioning search task, the running actions output by the target policy network corresponding to the target drone cluster for all drones can be replaced with attack actions to attack all drones in the target drone cluster.

[0039] The target policy network is obtained by reinforcement learning using the sample states, sample actions corresponding to the sample states, and the next sample states and reward values ​​corresponding to the sample actions in the training data generated by the interaction between the drone and the real environment. The policy network is used to output corresponding action decisions based on the state.

[0040] Then, for each executed first attack strategy, the following steps are performed: Get the current running status of the target drone cluster after executing the first attack strategy and the current running action The current operating action is used to represent the joint action of all drones in the target drone cluster. It is a joint action (such as the operating actions of each drone in different directions) determined by the target policy network corresponding to the target drone cluster based on the current operating state, or a joint action formed by replacing the actions of all drones in the joint action determined by the target policy network with attack actions. The specific operation can be adaptively determined based on the implementation method of the first attack strategy. The current operating state is used to represent the joint state of all drones in the target drone cluster, including but not limited to the operating state of each drone, such as the environment, position, distance, direction, and posture. This embodiment does not specifically limit this.

[0041] Then, the current operating status of the target drone cluster after executing the first attack strategy and the current running action Input into the predictor, which performs a multi-step prediction operation to predict the first global operating state sequence of the target drone cluster after executing the first attack strategy. The first global operating state sequence is the operating state of the target drone cluster under multiple time steps after executing the first attack strategy. The complete state sequence of the prediction operation can be adaptively determined according to the pre-configured maximum prediction step size or the task goal.

[0042] Based on the above prediction steps, by traversing the first attack strategy executed each time, the first global operating state sequence after the target drone cluster executes each first attack strategy can be obtained.

[0043] Step 120: Determine a target time step range from among the multiple time step ranges according to the first global operating state sequence.

[0044] Optionally, after obtaining the first global operating state sequence after the target drone cluster executes each first attack strategy, the attack impact when the first attack strategy is executed on the target drone cluster within different time step ranges can be evaluated based on each first global operating state sequence, so as to determine the time step range with the best attack effect in multiple time step ranges as the target time step range (also called the critical time step range) based on the attack impact.

[0045] Step 130: Within the target time step range, execute the second attack strategy on different drones in the target drone cluster in sequence, and based on the predictor, predict the second global operating state sequence of the target drone cluster after executing the second attack strategy each time.

[0046] Optionally, after obtaining the target time step range, the target UAV cluster can be traversed sequentially within the target time step range to execute the second attack strategy for each UAV in the target UAV cluster, while executing the normal strategy (i.e., the standardized strategy without attack perturbation) for other UAVs. For example, assuming there is UAVs, then within the target time step range, traverse Each time, the second attack strategy is executed against only one drone in the target drone cluster, while the other drones follow the normal strategy. This second attack strategy can be an m-step attack against a single drone in the target drone cluster (the drone executing the second attack strategy), where m is the length of the target time step range. The attack in this second attack strategy can be achieved by adding disturbances (such as ambient noise or communication interference) to the operating environment of a single drone in the target drone cluster, or by directly replacing the operational actions output by the target strategy network for a single drone in the target drone cluster with attack actions. The specific attack strategy can be adaptively determined based on the type of mission being performed by the target drone cluster.

[0047] Then, for each execution of the second attack strategy, the following steps are performed: Get the current running status of the target drone cluster after executing the second attack strategy and the current running action The current operating action is used to represent the joint action of all drones in the target drone cluster. This is a joint action (such as the operating actions of each drone in different directions) determined by the target policy network corresponding to the target drone cluster based on the current operating state, or a joint action formed by replacing the action of a single drone in the joint action determined by the target policy network with an attack action. The specific operation can be adaptively determined based on the implementation method of the second attack strategy. The current operating state is used to represent the joint state of all drones in the target drone cluster, including but not limited to the operating state of each drone, such as the environment, position, distance, direction, and posture. This embodiment does not specifically limit this.

[0048] Then, the current operating status of the target drone cluster after executing the second attack strategy and the current running action The prediction is input into the predictor, which performs a multi-step prediction operation to predict the second global operating state sequence of the target UAV cluster after executing the second attack strategy. The second global operating state sequence is the operating state of the target UAV cluster under multiple time steps after executing the second attack strategy. The complete state sequence of the prediction operation can be adaptively determined according to the pre-configured maximum prediction step size or the task goal.

[0049] Based on the above prediction steps, by traversing each execution of the second attack strategy, the second global operating state sequence after the target drone cluster executes each second attack strategy can be obtained.

[0050] Step 140: Determine at least one target drone in the target drone cluster based on the second global operating state sequence, and perform an adversarial attack test on the target drone cluster in a real test environment based on the target time step range and each target drone to obtain an adversarial attack test result of the target drone cluster.

[0051] Optionally, after obtaining the second global operating state sequence after the target drone cluster executes each second attack strategy, the attack impact when executing the second attack strategy on different drones in the target drone cluster within the target time step range can be evaluated based on each second global operating state sequence, so as to determine the drone with better attack effect in the target drone cluster as the target drone (also called key drone or key attack target) based on the attack impact.

[0052] Figure 2 This is the second flow chart of the UAV cluster testing method provided by the present invention. Figure 2 As shown in the figure, after searching the critical time step range and the critical drones, a third attack strategy can be executed against all target drones in the target drone cluster within the target time step range in a real test environment. This strategy is used to induce the target drone cluster's final operating state to deviate from the expected result (also known as the baseline state), ultimately leading to mission failure or a significant decrease in long-term reward value. This third attack strategy can be implemented by performing an m-step attack on all target drones in the target drone cluster. This can be achieved by adding perturbations to the operating environment of all target drones in the target drone cluster (such as environmental noise or communication interference), or by replacing the operating actions output by the target policy network corresponding to the target drone cluster with attack actions. The specific strategy can be adaptively determined based on the type of mission performed by the target drone cluster. m is the time step length of the target time step range.

[0053] The steps for obtaining the baseline state here include: in a real test environment, under a given initial operating state, inferring and outputting the third global operating state sequence after all drones in the target drone cluster jointly execute the normal strategy, and taking the terminal operating state in the third global operating state sequence as the baseline state.

[0054] Then, obtain the current running status of the target drone cluster after executing the third attack strategy and the current running action The current operating action is used to represent the joint action of all drones in the target drone cluster. This is a joint action (such as the operating actions of each drone in different directions) determined by the target policy network corresponding to the target drone cluster based on the current operating state, or a joint action formed by replacing the target drone's action in the joint action determined by the target policy network with an attack action. The specific operation can be adaptively determined based on the implementation method of the third attack strategy. The current operating state is used to represent the joint state of all drones in the target drone cluster, including but not limited to the operating state of each drone, such as the environment, position, distance, direction, and posture. This embodiment does not specifically limit this.

[0055] Then, the current operating status of the target drone cluster after executing the third attack strategy and the current running action On this basis, the drones in the target drone cluster interact with the real test environment multiple times to obtain the fourth global operating state sequence after the target drone cluster executes the third attack strategy. The fourth global operating state sequence is the operating state of multiple time steps after the target drone cluster executes the third attack strategy. The complete state sequence of the prediction operation can be adaptively determined according to the pre-configured maximum prediction step size or the task goal.

[0056] Then, the running state of the last time step in the fourth global running state sequence after the target UAV cluster executes the third attack strategy (also called the final running state) is used as the fourth terminal running state and the fourth terminated operation state Compare with the baseline state to determine the adversarial attack test results of the target drone cluster based on the comparison results.

[0057] The method provided in this embodiment constructs a predictor that can simulate an environment and interact with drones in a drone cluster. Based on the predictor, the predictor predicts the effects of the target drone cluster executing the attack strategy within different time step ranges to determine the target time step range. The predictor is then used to predict the effects of different drones executing the attack strategy within the target time step range, and at least one target drone is determined. Based on the target time step range and the at least one target drone, the attack strategy is executed on the target drone cluster in a real test environment to induce the final state of the target drone cluster to deviate from the expected state, thereby achieving an adversarial attack test on the target drone cluster. Therefore, during the test process, the predictor is used to determine the optimal attack time range and the optimal attack target to optimize the generation of the attack strategy. This not only reduces the attack range by shortening the attack time step length and reducing the number of attacked drones while ensuring the effectiveness of the attack, thereby improving the concealment of the adversarial attack, but also avoids the use of reinforcement learning algorithms to generate adversarial strategies, achieving lightweight black-box adversarial attack testing and reducing computational costs.

[0058] In some embodiments, the predictor is constructed based on the following steps: Obtaining a sample operating state of a sample UAV cluster, a sample operating action corresponding to the sample operating state, a true result of the next operating state corresponding to the sample operating action, and a true result of the global operating state sequence corresponding to the sample operating action; Inputting the sample running state and the sample running action into the initial neural network model for single-step prediction to obtain a next running state prediction result corresponding to the sample running action; Inputting the sample running state and the sample running action into the initial neural network model for global prediction, and obtaining a global running state sequence prediction result corresponding to the sample running action; Calculating a first loss value according to the actual result of the next operating state and the predicted result of the next operating state; Calculating a second loss value according to the actual result of the global operating state sequence and the predicted result of the global operating state sequence; The initial neural network model is iteratively trained according to the first loss value and the second loss value to obtain the predictor.

[0059] Optionally, when constructing a predictor, an initial neural network model (such as a deep neural network, a recurrent neural network, etc.) can be obtained first; the initial neural network model can be a pre-set network model, or a network model dynamically determined based on the task type of the task performed by the target drone cluster, etc. This embodiment does not specifically limit this.

[0060] In one possible implementation, the initial neural network model is constructed based on the task type of the task performed by the target drone cluster, and the task type includes a positioning search task or a confrontation combat task.

[0061] Optionally, when constructing the initial neural network model, it can be dynamically constructed based on the task type of the target drone cluster. Specifically, the corresponding network model can be dynamically selected from network models such as deep neural networks and recurrent neural networks based on the task type of the target drone cluster. The internal structure of the selected network model can be adjusted based on the task type to obtain an initial neural network model adapted to the task type, thereby improving the model's prediction accuracy and adaptability to specific tasks. For example, for adversarial combat tasks, a deep neural network containing two hidden layers can be selected to construct the initial neural network model.

[0062] In addition, a large amount of sample data is collected to form a training dataset; this sample data specifically includes the sample operating state of the sample drone cluster, the sample operating actions corresponding to the sample operating state, the actual result of the next operating state corresponding to the sample operating action, and the actual result of the global operating state sequence corresponding to the sample operating action. This sample data can be obtained from the training dataset of the target policy network corresponding to the target drone cluster, or it can be obtained from the historical interaction data generated during the historical interaction process between the target drone cluster and the real test environment, so that the predictor trained based on this sample data can better fit the environment and complete the computational feedback of state transition. In addition, to prevent overfitting of the model, the sample data can include both correctly executed actions and incorrectly executed actions.

[0063] Figure 3 Schematic diagram of the workflow of the predictor provided by the present invention; Figure 3 As shown in the figure, after obtaining the training data set, the sample running status and sample running action in the training data set can be input into the initial neural network model for single-step prediction to obtain the next running state prediction result corresponding to the sample running action. The specific implementation formula is as follows: ; in, is the model function of the initial neural network model; for The sample running status at the moment; for Sample running action at the moment; for The sample running status at the moment (that is, the prediction result of the next running status).

[0064] In addition, the sample operating states and sample operating actions in the training data set can be input into the initial neural network model for global prediction (i.e., multi-step prediction) to obtain the global operating state sequence prediction results corresponding to the sample operating actions.

[0065] Then, the prediction accuracy of the single prediction step is evaluated based on the predicted next state result and the actual next state result. That is, whether the predicted next state result is consistent with the actual next state result is evaluated, thereby obtaining a first loss value. Specifically, the prediction deviation of the single prediction step can be measured by calculating the Euclidean distance between the predicted next state result and the actual next state result, thereby evaluating the accuracy of the single-step prediction of the initial neural network model, thereby obtaining the first loss value.

[0066] In addition, the accuracy of the global prediction of the initial neural network model can be evaluated based on the prediction deviation between the terminal running state in the actual result of the global running state sequence predicted in the global prediction and the terminal running state in the next running state prediction result, that is, whether the initial neural network model can accurately predict a complete state sequence, thereby obtaining the second loss value.

[0067] After obtaining the first loss value and the second loss value, the first loss value and the second loss value can be fused (such as weighted addition or direct addition, etc.) to obtain a total loss value. The initial neural network model is iteratively trained based on the total loss value to obtain a predictor that can accurately simulate the environment to provide feedback on the operation actions of the drone cluster, and thereby predict the global operation state sequence corresponding to different operation actions.

[0068] The method provided in this embodiment obtains sample data based on the historical interaction data generated during the historical interaction process between the target drone cluster and the real test environment, as well as the training data of the target policy network corresponding to the target drone cluster. Based on the sample data, the initial neural network model is iteratively trained in combination with the loss value of the single-step prediction and the loss value of the global prediction, thereby constructing a predictor that can accurately predict the global operating state sequence corresponding to the operating actions of the drone cluster through a precise simulation environment, so that an effective adversarial test strategy can be dynamically explored based on the predictor constructed thereby. In the process of generating the adversarial test strategy, only the training data of the target policy network and the historical interaction data between the drones and the real test environment are obtained. There is no need to pay attention to the network structure of the target policy network, nor is there any need to use a reinforcement learning algorithm to generate the adversarial test strategy. This effectively implements black box attack testing, reduces computational costs, and increases the practicality of the test.

[0069] In some embodiments, step 120 specifically includes: Based on the predictor, in the initial operating state, predicting a third global operating state sequence after all drones in the target drone cluster jointly execute a normal strategy; Calculating, based on each first termination operating state and a reference state, a first attack impact degree after the target UAV cluster executes each of the first attack strategies; each first termination operating state is a termination operating state in each of the first global operating state sequences, and the reference state is a termination operating state in the third global operating state sequence; The target time step range is determined within the plurality of time step ranges according to each of the first attack impacts.

[0070] Optionally, during the target time step range search, for a given initial run state , can be based on the predictor to simulate the real test environment for multi-step prediction, to infer the output of the target UAV cluster after all UAVs jointly execute the normal strategy after multiple time steps of the running state, to form a third global running state sequence The running state of the last time step in the third global running state sequence (also called the third terminal running state) is used as the reference state. , to provide a reference for subsequent status comparison.

[0071] Next, the first global running state sequence after the UAV cluster executes the first attack strategy in different time steps based on the predictor simulation real test environment reasoning output is obtained, and the first termination running state is obtained in each first global running state sequence. .

[0072] Each first terminated operating state is then compared with a baseline state to calculate the first attack impact after the target drone cluster executes each of the first attack strategies. Based on each first attack impact, the time step range with the best attack effect is determined as the target time step range within multiple time step ranges. Specifically, because a larger gap (i.e., first attack impact) indicates a better attack effect, the time step range corresponding to the maximum first attack impact within the multiple time step ranges can be determined as the target time step range.

[0073] The attack impact here can be calculated based on a pre-set evaluation model, or it can be calculated based on an evaluation model dynamically determined based on the task type of the task performed by the target drone cluster, etc. This embodiment does not make specific restrictions on this.

[0074] Compared with the existing technology that uses empirical thresholds to select the optimal attack time step range, which makes it difficult to ensure the accuracy of attack performance testing, the method provided in this embodiment automatically explores the time step range that achieves the best attack effect through predictor-based and rule-based heuristic search. This not only reduces the number of attack steps while ensuring the attack effect, effectively enhancing the attack concealment in the time step dimension, but also reduces the dependence on empirical thresholds, effectively improving the accuracy of anti-attack testing.

[0075] In some embodiments, the calculating, based on each first termination operation state and the reference state, the first attack impact degree after the target drone cluster executes each first attack strategy includes: Determining a target evaluation function according to the task type of the task performed by the target UAV cluster; Based on the target evaluation function, respectively evaluating the task completion degree corresponding to each of the first terminated running states and the task completion degree corresponding to the reference state; A difference calculation is performed based on the task completion degree corresponding to each of the first termination operation states and the task completion degree corresponding to the baseline state to obtain the first attack impact degree after the target drone cluster executes each of the first attack strategies.

[0076] It should be noted that different mission types correspond to different completion targets. Therefore, in order to ensure the effectiveness and accuracy of the attack impact evaluation, it is necessary to configure corresponding evaluation functions for mission completion evaluation (hereinafter referred to as T-value functions) for different mission types. For example, in the positioning and search mission scenario, the T-value function reflects the distance between the target drone cluster and the target point. In the confrontational combat mission scenario, the T-value function is used to reflect the difference in the number of surviving individuals between the two drones.

[0077] Therefore, when evaluating the attack impact of drone clusters in different mission scenarios, the corresponding evaluation function can be dynamically determined as the target evaluation function based on the task type performed by the target drone cluster. The target evaluation function reflects the completion of the overall task of the target drone cluster and is used to evaluate the quality of the termination operation status of the target drone cluster.

[0078] Then, based on the target evaluation function, the first termination operation state is calculated respectively. Corresponding task completion , and baseline state Corresponding task completion ; and calculate each first termination state Corresponding task completion Compared with the baseline state Corresponding task completion The difference between them is used to determine the first attack impact of the target UAV cluster after executing each first attack strategy. , the specific calculation formula is as follows: .

[0079] The method provided in this embodiment accurately quantifies and evaluates the impact of different attack strategies on task completion by dynamically selecting an evaluation function based on the task type, thereby effectively screening out the optimal attack strategy and improving the success rate of attack testing.

[0080] In some embodiments, determining at least one target drone in the target drone cluster according to the second global operating state sequence in step 140 includes: Based on the predictor, in the initial operating state, predicting a third global operating state sequence after all drones in the target drone cluster jointly execute a normal strategy; Calculating, according to each second termination operating state and a reference state, a second attack impact degree after the target UAV cluster executes each second attack strategy; each second termination operating state is a termination operating state in each second global operating state sequence, and the reference state is a termination operating state in the third global operating state sequence; Determining an attack priority level corresponding to each of the drones according to each of the second attack impacts and the total number of time steps of each of the second global operating state sequences; At least one target drone is determined in the target drone cluster according to the attack priority level corresponding to each drone.

[0081] Optionally, during the target drone search process, for a given initial operating state, a multi-step prediction can be performed based on the predictor simulating a real test environment to infer the operating states of all drones in the target drone cluster after they jointly execute the normal strategy (i.e., the standard mission execution strategy when there is no attack), so as to construct a third global operating state sequence. The operating state of the last time step in the third global operating state sequence (also known as the third terminal operating state) is used as the baseline state. , to provide a reference for subsequent status comparison.

[0082] Next, the second global operation state sequence after the UAV cluster executes the second attack strategy on each UAV within the target time step range based on the predictor simulation real test environment reasoning output is obtained, and each second termination operation state is obtained in each second global operation state sequence. .

[0083] And based on the target evaluation function corresponding to the task type of the task performed by the target UAV cluster, the task completion degree corresponding to each second terminal operation state is calculated respectively. , and the task completion corresponding to the baseline state , and calculate the task completion degree corresponding to each second termination running state Task completion relative to the baseline state The difference between them determines the second attack impact after the target UAV cluster executes each second attack strategy. , the specific calculation formula is as follows: .

[0084] Furthermore, according to the second attack impact and the total time steps of each second global operating state sequence (that is, the total time steps of the task test), the attack priority level corresponding to each UAV is determined, and according to the attack priority level corresponding to each UAV, a rule-based heuristic search method is used in the target UAV cluster to search and determine one or more UAVs with better attack effects as target UAVs, so as to obtain the optimal attack target range.

[0085] The attack priority here can be obtained by performing multi-layer logical judgment on the second attack impact and total time steps corresponding to each drone, or it can be obtained by jointly scoring the attack effect of each drone based on the second attack impact and total time steps corresponding to each drone, and then dividing the drone into levels according to the score. This embodiment does not make specific restrictions on this.

[0086] Compared with the existing technology that uses empirical thresholds to select the optimal attack target range, which makes it difficult to ensure the accuracy of attack performance testing, the method provided in this embodiment automatically explores the attack target range that achieves the best attack effect through predictor-based and rule-based heuristic search. It not only reduces the number of attacking drones while ensuring the attack effect, effectively achieving the enhancement of attack concealment in the quantity dimension, but also reduces the dependence on empirical thresholds, effectively improving the accuracy of anti-attack testing.

[0087] In some embodiments, determining the attack priority level corresponding to each of the drones based on each of the second attack impacts and the total number of time steps of each of the second global operating state sequences includes: For each drone, when the second attack impact degree corresponding to the drone is greater than the second attack impact degree corresponding to any other drone, the attack priority level corresponding to the drone is determined to be higher than the attack priority level corresponding to any other drone; Alternatively, when the second attack influence degree corresponding to the drone is the same as the second attack influence degree corresponding to any other drone, and the total number of time steps of the second global operating state sequence corresponding to the drone is greater than the total number of time steps of the second global operating state sequence corresponding to any other drone, it is determined that the attack priority level corresponding to the drone is higher than the attack priority level corresponding to any other drone.

[0088] Optionally, since the attack priority of a drone is proportional to the second attack impact and the total number of time steps, the steps for determining the attack priority of each drone are as follows: If the second attack influence corresponding to this drone is greater than the second attack influence corresponding to any other drone, it means that the attack effect of this drone is more significant than that of any other drone. Therefore, the attack priority corresponding to this drone is higher than that of any other drone, and it should be attacked first.

[0089] If the second attack influence degree corresponding to the UAV is the same as the second attack influence degree corresponding to any other UAV, the total time steps corresponding to the UAV and the total time steps corresponding to any other UAV are further compared. If the total time steps corresponding to the UAV are greater than the total time steps corresponding to any other UAV, it means that the UAV has a greater impact on the task completion. Therefore, the attack priority level corresponding to the UAV is higher than the attack priority level corresponding to any other UAV, and it should be attacked first.

[0090] According to the above attack priority determination steps, each drone can be traversed to obtain the attack priority corresponding to each drone.

[0091] The method provided in this embodiment accurately determines the attack priority of the drone by comprehensively evaluating the attack impact and task completion time, thereby effectively screening out the optimal attack strategy and improving the success rate of the attack test.

[0092] The following specific application examples are used to verify the effectiveness of the method provided in this embodiment.

[0093] During the validation process, a virtual combat mission was selected as the test scenario. Within this scenario, a multi-agent reinforcement learning experiment graph was used as the implementation scenario. In this scenario, the friendly and enemy combat groups, each equipped with five drones, battled each other, with victory (or mission completion) achieved by eliminating all of the enemy's drones. The agents corresponding to the friendly drones were trained using the Monotone Value Function Decomposition (QMIX) algorithm from deep multi-agent reinforcement learning. After training, the agents achieved a 97.8% success rate, which served as the target policy network for attack.

[0094] During the verification process, you can follow Figure 2 The process shown here verifies the method provided in this embodiment. The specific process is as follows: The first step is to build a predictor: for the adversarial combat mission, build an initial neural network model and train the initial neural network model to build a predictor. The specific implementation process is as follows: For adversarial combat missions, a deep neural network with two hidden layers is constructed as the initial neural network model. 10,000 rounds of sample data are randomly sampled from the training and test datasets of the target policy network to serve as the training dataset for the initial neural network model. The sample data consists of the sample operating states of a sample drone cluster and the sample operating actions corresponding to the sample operating states, with the actual next operating state results and the actual global operating state results corresponding to the sample operating actions as labels. The sample data is then fed into the initial neural network model for single-step and multi-step predictions to obtain single-step and global prediction losses. These single-step and global prediction losses are then combined to train the initial neural network model to obtain a predictor. Specifically, the mean square error loss function is used to calculate the single-step and global prediction losses, and the adaptive moment estimation (Adam) optimizer is employed for optimization training.

[0095] In order to verify the performance of the predictor trained in this way, the following performance evaluation was performed on the predictor: (1) In a single prediction step, given the sample running state and sample running action of the sample UAV cluster, the predictor predicts the next running state prediction result corresponding to the sample running action; and the deviation of the single-step prediction result is measured by calculating the Euclidean distance between the next running state prediction result predicted by the predictor and the next running state actual result corresponding to the sample running action, as follows: Figure 4 As shown in Figure 2, the prediction deviation between the two is small, which shows that the predictor trained based on this has good single-step prediction performance.

[0096] (2) The predictor is used to simulate the real test environment for 100 rounds of testing, and the T value function is defined as the difference between the number of surviving drones of the own side and the number of surviving drones of the enemy side, that is, If the number of drones on one side reaches zero, the battle ends and the other side wins. Figure 5 As shown in the figure, in 100 rounds of testing, the winning rate of our drone reached 92%, which is slightly lower than the normal winning rate of 97%, further demonstrating that the predictor can effectively exert its global prediction performance.

[0097] The second step is the key time step range search: for a given initial state, the key time step range is determined based on the predictor and traversal search. The specific implementation process is as follows: For a given drone swarm’s initial operating state, the predictor is used to infer the final operating state (i.e., the baseline state) of the drone swarm after executing the normal strategy. The result is that all five friendly drones survive and all five enemy drones are defeated. At this time, the T value in the baseline state is 5.

[0098] Based on different m values, multiple time step ranges are determined, and the first attack strategy is jointly executed on all drones in the target drone cluster within different time step ranges. The final operating state (also called the first terminal operating state) of the target drone cluster after executing each first attack strategy is predicted based on the predictor, and each first terminal operating state is compared with the baseline state to determine the attack effect corresponding to each time step range, so as to search and determine the key time step range.

[0099] As shown in Table 1, when m=3, the winning rate drops from 97.8% to about 50%, when m=5, the winning rate drops to about 10%, and when m=7, the game is almost 100% lost. The traversal search results show that starting the attack from the 11th step can achieve the best attack effect. Based on this, the time step range determined when m=7 is Defined as the critical time step range.

[0100] Table 1 Test results corresponding to different attack time steps

[0101] The third step is to search the range of key attack targets: based on the determined key time step range, five loop tests are carried out, each time only one of the drones is attacked, and the remaining drones execute normal strategies. The final state of the drone cluster (also known as the second terminal operation state) is inferred through the predictor, and the T-value function is used to evaluate the attack effect corresponding to each drone, thereby determining the attack priority level of each drone.

[0102] As shown in Table 2, the test results show that Drone 2 has the smallest T value, which is the largest difference from the baseline T value of 5. Therefore, this drone is determined to have the highest attack priority, that is, the first priority. Drones 1 and 3 have the same T value, but Drone 0 has a longer total time step than the baseline state. Therefore, Drone 0 has a higher attack priority than Drone 1.

[0103] Table 2 Attack priority configuration table

[0104] The fourth step is to conduct an attack test in a real test scenario: Based on the determined critical time step range and the attack priority of the drones to be attacked, an adversarial attack test against the drone swarm is conducted in a real test scenario. As shown in Table 3, the test results show that within the critical time step range, attacking only drone 2 reduces the win rate from 97.8% to 61%. Attacking drones 2 and 4 reduces the win rate to 24%, and attacking drones 2, 4, and 1 reduces the win rate to 10%. In other words, when attacking only three drones and the critical time step range is 7 time steps, the win rate can be reduced from 97.8% to 10%, thus demonstrating the effectiveness of the attack test method provided in this embodiment.

[0105] Table 3. Distribution of anti-attack test results

[0106] The drone cluster testing system provided by the present invention is described below. The drone cluster testing system described below and the drone cluster testing method described above can be referenced to each other.

[0107] Figure 6 Schematic diagram of the structure of the UAV cluster test system provided by the present invention; Figure 6 As shown, the system includes: The first prediction unit 610 is configured to jointly execute a first attack strategy once on all drones in the target drone cluster in different time step ranges, and predict, based on a predictor, a first global operating state sequence of the target drone cluster after executing each of the first attack strategies; The search unit 620 is configured to determine a target time step range from among the multiple time step ranges according to the first global operating state sequence; The second prediction unit 630 is configured to execute the second attack strategy on each drone in the target drone cluster in sequence within the target time step range, and predict, based on the predictor, a second global operating state sequence of the target drone cluster after each execution of the second attack strategy; The testing unit 640 is used to determine at least one target drone in the target drone cluster according to the second global operating state sequence, and perform an adversarial test on the target drone cluster in a real test environment according to the target time step range and each target drone, to obtain an adversarial test result of the target drone cluster.

[0108] The system provided in this embodiment constructs a predictor that can simulate an environment and interact with drones in a drone cluster. Based on the predictor, it predicts the effects of attack strategies executed by the target drone cluster within different time step ranges to determine the target time step range. The system then combines the predictor to predict the effects of attack strategies executed by different drones within the target time step range to determine at least one target drone. Based on the target time step range and the at least one target drone, the system executes the attack strategy on the target drone cluster in a real test environment to induce the final state of the target drone cluster to deviate from the expected state, thereby achieving adversarial attack testing on the target drone cluster. Therefore, during the test process, the system optimizes the generation of attack strategies by combining the predictor to determine the optimal attack time range and the optimal attack target. This not only reduces the attack range by shortening the attack time step length and reducing the number of attacked drones while ensuring the effectiveness of the attack, thereby improving the concealment of the adversarial attack, but also avoids the use of reinforcement learning algorithms to generate adversarial strategies, achieving lightweight black-box adversarial attack testing and reducing computational costs.

[0109] The system provided by the present invention is used to execute the above-mentioned method embodiments. Please refer to the above-mentioned embodiments for the specific processes and detailed contents, which will not be repeated here.

[0110] Figure 7 An example of a physical structure diagram of an electronic device is shown below. Figure 7As shown, the electronic device may include: a processor (processor) 710 , a communication interface (Communications Interface) 720 , a memory (memory) 730 and a communication bus 740 , wherein the processor 710 , the communication interface 720 and the memory 730 communicate with each other via the communication bus 740 . The processor 710 can call the logic instructions in the memory 730 to execute the drone cluster testing method, which includes: jointly executing a first attack strategy on all drones in the target drone cluster within different time step ranges, and predicting a first global operating state sequence after the target drone cluster executes the first attack strategy each time based on a predictor; determining a target time step range in multiple time step ranges based on the first global operating state sequence; within the target time step range, sequentially executing a second attack strategy on different drones in the target drone cluster, and predicting a second global operating state sequence after the target drone cluster executes the second attack strategy each time based on the predictor; determining at least one target drone in the target drone cluster based on the second global operating state sequence, and performing an adversarial attack test on the target drone cluster in a real test environment based on the target time step range and each target drone, to obtain an adversarial attack test result of the target drone cluster.

[0111] Furthermore, the logic instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product, stored in a storage medium, includes instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0112] On the other hand, the present invention also provides a computer program product, which includes a computer program, which can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the drone cluster testing method provided by the above methods, the method including: jointly executing a first attack strategy on all drones in the target drone cluster in different time step ranges, and predicting a first global operating state sequence after the target drone cluster executes the first attack strategy each time based on a predictor; determining a target time step range in multiple time step ranges based on the first global operating state sequence; within the target time step range, sequentially executing a second attack strategy on different drones in the target drone cluster, and predicting a second global operating state sequence after the target drone cluster executes the second attack strategy each time based on the predictor; determining at least one target drone in the target drone cluster based on the second global operating state sequence, and performing an adversarial attack test on the target drone cluster in a real test environment based on the target time step range and each target drone, to obtain an adversarial attack test result of the target drone cluster.

[0113] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the drone cluster testing method provided by the above methods, the method comprising: jointly executing a first attack strategy on all drones in the target drone cluster within different time step ranges, and predicting, based on a predictor, a first global operating state sequence after the target drone cluster executes the first attack strategy each time; determining a target time step range in multiple time step ranges based on the first global operating state sequence; within the target time step range, sequentially executing a second attack strategy on different drones in the target drone cluster, and predicting, based on the predictor, a second global operating state sequence after the target drone cluster executes the second attack strategy each time; determining at least one target drone in the target drone cluster based on the second global operating state sequence, and performing an adversarial attack test on the target drone cluster in a real test environment based on the target time step range and each target drone, to obtain an adversarial attack test result of the target drone cluster.

[0114] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0115] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A drone cluster testing method, characterized in that: include: Within different time step ranges, jointly executing a first attack strategy on all drones in the target drone cluster, and predicting a first global operating state sequence of the target drone cluster after each execution of the first attack strategy based on a predictor; Determining a target time step range among the plurality of time step ranges according to the first global operating state sequence; Within the target time step range, sequentially executing a second attack strategy on different drones in the target drone cluster, and predicting, based on the predictor, a second global operating state sequence of the target drone cluster after each execution of the second attack strategy; According to the second global operating state sequence, at least one target drone is determined in the target drone cluster, and based on the target time step range and each target drone, an adversarial attack test is performed on the target drone cluster in a real test environment to obtain an adversarial attack test result of the target drone cluster.

2. The UAV cluster testing method according to claim 1, characterized in that: The predictor is constructed based on the following steps: Obtaining a sample operating state of a sample UAV cluster, a sample operating action corresponding to the sample operating state, a true result of the next operating state corresponding to the sample operating action, and a true result of the global operating state sequence corresponding to the sample operating action; Inputting the sample running state and the sample running action into the initial neural network model for single-step prediction to obtain a next running state prediction result corresponding to the sample running action; Inputting the sample running state and the sample running action into the initial neural network model for global prediction, and obtaining a global running state sequence prediction result corresponding to the sample running action; Calculating a first loss value according to the actual result of the next operating state and the predicted result of the next operating state; Calculating a second loss value according to the actual result of the global operating state sequence and the predicted result of the global operating state sequence; The initial neural network model is iteratively trained according to the first loss value and the second loss value to obtain the predictor.

3. The UAV cluster testing method according to claim 2, characterized in that: The initial neural network model is constructed according to the task type of the task performed by the target drone cluster, and the task type includes a positioning search task or a confrontation combat task.

4. The drone cluster testing method according to any one of claims 1 to 3, characterized in that: Determining a target time step range from among the plurality of time step ranges according to the first global operating state sequence includes: Based on the predictor, in the initial operating state, predicting a third global operating state sequence after all drones in the target drone cluster jointly execute a normal strategy; Calculating, based on each first termination operating state and a reference state, a first attack impact degree after the target UAV cluster executes each of the first attack strategies; each first termination operating state is a termination operating state in each of the first global operating state sequences, and the reference state is a termination operating state in the third global operating state sequence; The target time step range is determined within the plurality of time step ranges according to each of the first attack impacts.

5. The UAV cluster testing method according to claim 4, characterized in that: The calculating, based on each first termination operation state and the reference state, a first attack impact degree after the target UAV cluster executes each of the first attack strategies, includes: Determining a target evaluation function according to the task type of the task performed by the target UAV cluster; Based on the target evaluation function, respectively evaluating the task completion degree corresponding to each of the first terminated running states and the task completion degree corresponding to the reference state; A difference calculation is performed based on the task completion degree corresponding to each of the first termination operation states and the task completion degree corresponding to the baseline state to obtain the first attack impact degree after the target drone cluster executes each of the first attack strategies.

6. The drone cluster testing method according to any one of claims 1 to 3, characterized in that: The step of determining at least one target drone in the target drone cluster according to the second global operating state sequence includes: Based on the predictor, in the initial operating state, predicting a third global operating state sequence after all drones in the target drone cluster jointly execute a normal strategy; Calculating, according to each second termination operating state and a reference state, a second attack impact degree after the target UAV cluster executes each second attack strategy; each second termination operating state is a termination operating state in each second global operating state sequence, and the reference state is a termination operating state in the third global operating state sequence; Determining an attack priority level corresponding to each of the drones according to each of the second attack impacts and the total number of time steps of each of the second global operating state sequences; At least one target drone is determined in the target drone cluster according to the attack priority level corresponding to each drone.

7. The UAV cluster testing method according to claim 6, characterized in that: Determining the attack priority level corresponding to each of the UAVs according to each of the second attack impacts and the total number of time steps of each of the second global operating state sequences includes: For each drone, when the second attack impact degree corresponding to the drone is greater than the second attack impact degree corresponding to any other drone, the attack priority level corresponding to the drone is determined to be higher than the attack priority level corresponding to any other drone; Alternatively, when the second attack influence degree corresponding to the drone is the same as the second attack influence degree corresponding to any other drone, and the total number of time steps of the second global operating state sequence corresponding to the drone is greater than the total number of time steps of the second global operating state sequence corresponding to any other drone, it is determined that the attack priority level corresponding to the drone is higher than the attack priority level corresponding to any other drone.

8. A drone cluster testing system, characterized in that: include: A first prediction unit is configured to jointly execute a first attack strategy on all drones in a target drone cluster at different time steps, and predict, based on a predictor, a first global operating state sequence of the target drone cluster after each execution of the first attack strategy; A search unit, configured to determine a target time step range from among the plurality of time step ranges according to the first global operating state sequence; A second prediction unit is configured to sequentially execute a second attack strategy on different drones in the target drone cluster within the target time step range, and predict, based on the predictor, a second global operating state sequence of the target drone cluster after each execution of the second attack strategy; The testing unit is used to determine at least one target drone in the target drone cluster according to the second global operating state sequence, and perform an adversarial attack test on the target drone cluster in a real test environment according to the target time step range and each target drone, to obtain an adversarial attack test result of the target drone cluster.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the drone cluster testing method according to any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the drone cluster testing method according to any one of claims 1 to 7 is implemented.

11. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the drone cluster testing method according to any one of claims 1 to 7 is implemented.

Images