Data-driven unmanned aerial vehicle configuration parameter vulnerability fuzzy test method
By employing a data-driven fuzzy testing method for the vulnerability of UAV parameters, and utilizing SHAP and Pearson correlation analysis, key sensor features are identified and the search space is narrowed. This solves the problem of low efficiency in anomaly detection caused by combinations of UAV configuration parameters, and achieves efficient identification of abnormal states and discovery of unknown errors.
Patent Information
- Application Number
- CN202511223239.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-29
- Publication Date
- 2025-12-23
AI Technical Summary
Existing technologies are insufficient for efficiently identifying specific abnormal states caused by combinations of drone configuration parameters, and are also lacking in specificity and efficiency.
A data-driven approach is adopted to identify key sensor features through SHAP analysis, construct a dynamic fitness function, and use Pearson correlation analysis to screen a subset of configuration parameters. Combined with genetic algorithm mutation, the search space is narrowed.
It achieves accurate determination of specific abnormal states, improves detection efficiency, has the ability to discover unknown errors, and forms an adaptive closed-loop system.
Smart Images
Figure CN121187906A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of unmanned aerial vehicle (UAV) system security, and particularly relates to a data-driven unmanned aerial vehicle configuration parameter vulnerability fuzzing method. BACKGROUND
[0002] As a complex cyber-physical system, unmanned aerial vehicles (UAVs) control actuators to accomplish autonomous tasks by processing sensor inputs and user commands. Its applications have penetrated into many fields such as traffic monitoring, logistics distribution, infrastructure inspection, etc. To adapt to diverse flight environments, task requirements and aircraft types, UAV firmware (such as ArduPilot) usually contains a large number (e.g., more than 2600) of configurable control parameters. Users upload these parameters through the ground control station (GCS) according to the official guidelines.
[0003] However, this high degree of configurability introduces a vulnerability called "range specification bug". Even if a single parameter is within its official recommended range, improper parameter combinations can cause the UAV to fly unstably, or even cause serious abnormal states such as crash, yaw imbalance or potential thrust loss. Given the vastness of the parameter space, it is almost impossible to comprehensively test all parameter combinations.
[0004] Traditional software analysis methods, such as static program analysis, are difficult to effectively apply due to the complexity of UAV firmware. There have been some fuzzing tests for range specification bugs, for example, RVFuzzer uses binary search to detect abnormal parameter boundaries, but it is not effective when multiple parameters interact to cause abnormalities. LGDFuzzer and APFuzzer and other methods have made some progress in configuration testing, but LGDFuzzer lacks specificity for specific abnormal state categories, and APFuzzer's detection of abnormal states is uneven, making it difficult to effectively search for specific problem configurations. Previous research has largely failed to effectively and specifically search for configurations that cause specific serious flight abnormalities. Therefore, how to efficiently identify configurations that cause specific abnormal states among numerous parameter combinations, and effectively connect the "cyber space" of configuration parameters with the "physical space" of actual UAV operation states, is a key challenge currently faced. SUMMARY
[0005] The present application aims to address the problem that existing UAV configuration parameter vulnerability detection methods are inefficient, lack specificity, and are difficult to effectively reduce the search space in identifying specific parameter combinations for specific abnormal states, and proposes a data-driven, correlation analysis-based UAV parameter vulnerability fuzzing method (ASFuzzer).
[0006] The object of the application is achieved by the technical solutions below: a data-driven unmanned aerial vehicle configuration parameter vulnerability fuzzing method, comprising:
[0007] Performing a flight task in a simulator to pre-collect unmanned aerial vehicle flight logs;
[0008] Classifying various abnormal states of the unmanned aerial vehicle until the number of abnormal state logs reaches a preset threshold,
[0009] Iterative optimization: merging flight logs of abnormal states and randomly sampled normal logs, and performing SHAP analysis on the merged data set;
[0010] Updating the fitness function of the genetic algorithm according to the SHAP analysis results, and performing Pearson correlation analysis on the relevant sensor feature joint configuration parameters obtained by SHAP analysis to obtain an updated population range of the genetic algorithm;
[0011] Based on the fitness function and the updated population range, a new round of configuration parameters is generated using genetic algorithm mutation and input into the simulator to generate mutated flight logs, and the fitness function value of this flight is calculated for the next iteration of population mutation, and the obtained flight logs are classified according to abnormal types;
[0012] Iterating until the relevant configuration parameters are not updated or the maximum number of evolutions is reached, and outputting the test results.
[0013] Further, the pre-collection of the unmanned aerial vehicle flight logs includes obtaining codes related to abnormal state determination from the source code of the unmanned aerial vehicle control program, and defining an initial fitness function to guide the rapid collection of abnormal state logs; the flight log data includes flight attitude, acceleration, GPS position, speed, motor throttle output, internal monitoring variables, and a list of current flight configuration parameters.
[0014] Further, the classification of various abnormal states of the unmanned aerial vehicle includes: according to the performance of the unmanned aerial vehicle in physical behavior and the safety definition in the firmware, the abnormal states are divided into: crash, potential thrust loss, yaw imbalance, or unintended autonomous rotation;
[0015] Further, the SHAP analysis on the combined data set of the flight logs of the abnormal state and the randomly sampled normal logs specifically includes: combining B abnormal logs and B randomly sampled normal logs, where B is a threshold of the number of abnormal states; training a random forest classifier model to learn the relationship between sensor features and abnormal states, using the SHAP method to assign a contribution score to each sensor, quantifying the influence of the feature on the abnormal state, and outputting a list of key sensor features most relevant to the current target abnormal state, including the abnormal threshold of the sensor and the contribution score.
[0016] Further, the updating of the fitness function of the genetic algorithm according to the SHAP analysis result specifically includes:
[0017] First, the fitness function of the genetic algorithm is updated, and the identified key sensor features, their abnormal thresholds and contribution scores are used to construct the fitness function.
[0018] The function is in the form of: f n =∑W i (S n,i -threshold i ); where W i is the weight of the i-th key sensor, and its absolute value is equal to the contribution score of the sensor; S n,i is the value of the sensor in the flight segment n; threshold i is the abnormal threshold of the sensor; the sign of the weight W i is determined according to whether the sensor value is high or low in the abnormal state, so as to ensure that the abnormal state always obtains a positive fitness score.
[0019] Further, the relevant sensor feature joint configuration parameters obtained by the SHAP analysis are subjected to Pearson correlation analysis.
[0020] Pearson correlation analysis is performed, and according to the key sensor features, the corresponding sensor values S and the configuration parameters P of the log are extracted from the previous log, and the log is cut into segments F, in the form of: F=(S,P); a representative sensor evaluation score (S score ) is calculated for each flight log, and then Pearson correlation analysis is used to calculate the correlation coefficient between each configuration parameter and the S score , and finally, all configuration parameters with a correlation greater than a threshold value are combined to obtain a relevant configuration parameter set P c for the abnormal state.
[0021] Further, the relevant configuration parameters are used to update the population range of the genetic algorithm.
[0022] According to the obtained abnormal state related configuration parameter set P c , update the range of the genetic algorithm population, and in the subsequent genetic algorithm mutation process, only the related configuration parameter set P c will be modified and mutated, and all non-selected irrelevant parameters will be set to their default values when generating new individuals.
[0023] Further, in the generation of a new round of configuration parameters using genetic algorithm mutation, the mutation strategy adopts a differential evolution algorithm.
[0024] On the other hand, the specification also provides a data-driven unmanned aerial vehicle configuration parameter vulnerability fuzz testing device, comprising a memory and one or more processors, the memory stores executable code, and the processor executes the executable code to realize the data-driven unmanned aerial vehicle configuration parameter vulnerability fuzz testing method.
[0025] On the other hand, the specification also provides a computer readable storage medium having a program stored thereon, which, when executed by a processor, implements the data-driven unmanned aerial vehicle configuration parameter vulnerability fuzz testing method.
[0026] The beneficial effects of the present application: a data-driven, correlation analysis-based unmanned aerial vehicle parameter vulnerability fuzz testing method is proposed, which effectively solves the problem of low detection efficiency and lack of targeting caused by the huge parameter space in the prior art. This method accurately identifies key sensor features related to specific abnormal states (such as crash, rotation anomaly, etc.) through SHAP analysis, and constructs a dynamic fitness function, thereby achieving accurate judgment of specific risks. At the same time, Pearson correlation analysis is used to select a subset of configuration parameters that are strongly related to these features, and the mutation range of the genetic algorithm is concentrated in this subset, greatly reducing the search space and improving the test efficiency. The entire process forms a self-adaptive closed-loop system that continuously learns new flight logs to optimize the model, not only practical, but also has the ability to discover unknown error types. BRIEF DESCRIPTION OF DRAWINGS
[0027] Figure 1 A flowchart of the vulnerability detection method for the configurable parameters in the flight control firmware of the unmanned aerial vehicle of the present application;
[0028] Figure 2 A SHAP analysis result graph of the vulnerability detection method for the configurable parameters in the flight control firmware of the unmanned aerial vehicle of the present application;
[0029] Figure 3 A sensor abnormal threshold box plot of the vulnerability detection method for the configurable parameters in the flight control firmware of the unmanned aerial vehicle of the present application;
[0030] Figure 4 A Pearson correlation analysis heat map of configuration parameters for a vulnerability detection method for configurable parameters in unmanned aerial vehicle flight control firmware is configured for the present application.
[0031] Figure 5 A schematic diagram of a vulnerability detection device for configurable parameters in unmanned aerial vehicle flight control firmware is configured for the present application. DETAILED DESCRIPTION
[0032] The specific embodiments of the present application are further described in detail below with reference to the accompanying drawings.
[0033] Figure 1 A vulnerability detection method for configurable parameters in unmanned aerial vehicle flight control firmware is configured for the present application in the embodiment of the present application, which comprises the following steps:
[0034] Step 1: Perform flight tasks in the simulator to pre-collect unmanned aerial vehicle flight logs. This step is the pre-processing stage of the method. The purpose is to accumulate enough diverse flight data before the main fuzzy test starts. Specifically, this step uses a genetic algorithm framework and defines an initial fitness function based on the safety check code logic built into the unmanned aerial vehicle firmware ArduPilot. The form of the fitness function is as follows:
[0035]
[0036] where W i is the weight of the ith key sensor; S n,i is the value of the sensor in flight segment n; threshold i is the abnormal threshold of the sensor. The key sensors and the corresponding initial weights and thresholds are obtained from the safety check code logic. A variety of configuration parameter combinations are generated through the initial genetic algorithm, and these parameters are input into the unmanned aerial vehicle simulator (for example, APM simulator of IRIS+ quadrotor model with ArduPilot4.3.0 firmware) to perform flight tasks. The simulator will record the complete flight log, including sensor readings, internal state variables, applied configuration parameters and timestamps, which constitute the basic data set for subsequent analysis. Specifically, the flight log data includes but is not limited to: flight attitude (roll, pitch, yaw angle and its rate), acceleration, GPS position, speed, motor throttle output, internal monitoring variables (such as RMSYAWI), current flight configuration parameter list, data sampling frequency is 10Hz.
[0037] Step 2, classify the multiple abnormal states of the UAV. The flight logs collected in step 1 are sorted and classified, and the abnormal states are divided into at least four types according to the performance of the UAV in physical behavior and the safety definition in the firmware: 1) crash, the UAV loses control and collides with an object or the ground; 2) potential thrust loss, the UAV cannot generate enough thrust to maintain stable flight or attitude, even if the throttle is full output; 3) yaw imbalance, the UAV control system overuses the control ability to maintain yaw stability, which may cause yaw control saturation and even out-of-control rotation; 4) a new error type discovered by the invention, which refers to the unexpected autonomous rotation of the UAV during task execution, affecting the accuracy of the heading and trajectory. All flight logs will be labeled and classified according to these abnormal states for subsequent targeted analysis.
[0038] Step 3, determine whether the number of abnormal logs of the UAV to be found is greater than the threshold value. This step is the trigger condition for starting the core analysis. For each abnormal state classified in step 2, the system will monitor the number of logs, and when the number of logs of a specific abnormal state reaches a preset threshold B, the subsequent step 4 analysis process will be started for that abnormal state. The threshold B is initially set to ensure that there is enough data for robust statistical analysis, while also considering computational efficiency. If the number of logs does not reach the threshold B, the process returns to step 2.
[0039] Step 4, SHAP analysis of the attitude, sensor, and other states in the collected UAV abnormal state logs. This step is the core of sensor feature selection. First, initialize a sensor feature list, then when the condition of step 3 is met, the system will combine B abnormal logs and B randomly sampled normal logs. In the data merging process, we extract the configuration parameters and candidate sensor readings at each time from the abnormal and normal logs, and label the data in the normal logs as "0" and the data in the abnormal logs as "1" for easy classification. Then use the SHAP method to analyze the merged data set. SHAP (SHapley Additive exPlanations) is a model-agnostic explanation method based on cooperative game theory, which assigns a contribution score (i.e. Shapley value) to each sensor feature (such as roll angle, pitch angle, etc.) to quantify the influence of the feature on the model prediction (i.e. judging as an abnormal state). To perform SHAP analysis, we use the Xgboost open source library to train a random forest classifier model to learn the relationship between sensor features and abnormal states in this invention. The training set and test set are divided according to the 7:3 ratio, and the evaluation indicators of the classifier are set as precision, recall and F1 score. SHAP analysis calculates the contribution score for each sensor feature, such as Figure 2As shown, the contribution of the RMSYAWI parameter to the abnormal state is positively correlated, indicating that the larger the parameter value, the more likely the UAV will be in an abnormal state. Then, select the sensors with higher contribution and analyze the abnormal threshold of these sensors. The present application uses a box plot to analyze the abnormal threshold, as shown in Figure 3 . The final output of this step is a list of key sensor features most relevant to the current target abnormal state, which includes the abnormal threshold of the sensor and the contribution score.
[0040] Step 5, update the fitness function according to the results of SHAP analysis, and perform Pearson correlation analysis on the relevant sensor features and joint configuration parameters obtained by SHAP analysis. This step contains two parallel core steps, both based on the analysis results of step 4:
[0041] 1. First, update the fitness function of the genetic algorithm, use the key sensor features and their abnormal thresholds identified by S4 and the contribution scores to construct a dynamic updated fitness function for a specific abnormal state. The form of this function can be: f n =∑W i (S n,i -threshold i ). Where W i is the weight of the ith key sensor, whose absolute value is equal to the contribution score of the sensor; S n,i is the value of the sensor in flight segment n; threshold i is the abnormal threshold of the sensor. The positive and negative values of the weight W i are determined according to whether the sensor value is high or low when it is abnormal, to ensure that the abnormal state always gets a positive fitness score.
[0042] 2. Perform Pearson correlation analysis, according to the key sensor features obtained in S4, extract the corresponding sensor values S and the configuration parameters P of the log from the previous log, and cut the log into 2-second segments F, in the form of: F=(S,P). In order to solve the problem of time scale mismatch between static parameters and dynamic sensor data, a representative sensor evaluation score (S score ) needs to be calculated for each flight log, and then Pearson correlation analysis is used to calculate the correlation coefficient between each configuration parameter and the S score . The results are shown in Figure 4 . Finally, combine all configuration parameters with a correlation greater than the threshold to obtain a set of relevant configuration parameters P c for this abnormal state.
[0043] Step 6, update the genetic algorithm population according to the correlation configuration parameter set obtained by Pearson correlation analysis. In this step, the abnormal state correlation configuration parameter set P c obtained in step 5 is used to update the range of the genetic algorithm population. In the subsequent genetic algorithm mutation process, only this correlation configuration parameter set P c will be modified and mutated. The specific implementation of updating the population is to set all non-selected non-correlated parameters to default values when generating new individuals. This design greatly reduces the search space and is the key to improving the efficiency of fuzz testing.
[0044] Step 7, generate a new round of configuration parameters according to the fitness function value and the latest population. The genetic algorithm uses the fitness function updated in step 5 as the evaluation standard and follows the mutation range determined in step 6. Through selection, recombination, and mutation operations, the population is iteratively optimized. The mutation strategy uses the differential evolution algorithm. Finally, the genetic algorithm generates a new round of candidate configuration parameter combinations that are more likely to trigger the target genetic state.
[0045] Step 8, input the generated configuration parameters into the simulator to generate flight logs. The candidate configuration parameter combinations generated in step 7 are input into the simulator to generate flight logs. The collected flight logs are returned to step 2 for abnormal state classification and added to the historical log database for future analysis (step 3), thereby realizing continuous enrichment of data and continuous optimization of the model.
[0046] Step 9, calculate the fitness function value of this flight according to the flight log generated by the current configuration parameters. The obtained flight log is classified according to the abnormal type, and the fitness function value is used to guide the mutation process of the genetic algorithm. Calculate the fitness function value of the flight log generated in step 8. We use the updated fitness function in step 5 to calculate the corresponding fitness function value for each candidate configuration parameter combination generated by the flight log, and feed it back to the genetic algorithm as key information for selection in the next iteration (step 7). In step 5, we define to ensure that the larger the fitness function value, the more the system tends to be in an abnormal state. We set the target in the genetic algorithm to maximize the fitness function value, that is, individuals with higher fitness function values have a higher probability of being retained and reproduced during the iteration process, thereby guiding the search direction of the genetic algorithm.
[0047] Step 10, if the population range is no longer updated, end this round of fuzz testing, otherwise continue. This step is the termination condition of the entire fuzz testing. If the correlation configuration parameter set P cIf the updating is no longer performed or a preset maximum evolution number is reached, a termination condition is met, the test ends, and a set of configuration parameter combinations, key sensors, and related configuration parameters that cause the abnormal state are obtained; otherwise, a new round of iterative optimization is performed from step 4.
[0048] In the above manner, the unmanned aerial vehicle parameter vulnerability fuzzy test method based on data driving and correlation analysis can effectively solve the problems of low detection efficiency and lack of pertinence caused by a huge parameter space in the prior art. The method accurately identifies key sensor features related to a specific abnormal state (such as crash, rotation anomaly, etc.) through SHAP analysis and constructs a dynamic fitness function, thereby realizing accurate determination of specific risks. Meanwhile, Pearson correlation analysis is used to screen a subset of configuration parameters that are strongly related to these features, and the mutation range of the genetic algorithm is concentrated in this subset, greatly reducing the search space and improving the test efficiency. The entire process forms a self-adaptive closed-loop system that continuously optimizes the model by learning newly generated flight logs, and not only has strong practicality but also has the ability to discover unknown error types.
[0049] The above steps are compared with experimental analysis as follows:
[0050] In order to verify the effectiveness of the method of the present application, experiments were carried out in a specific environment.
[0051] Experimental setup: The experiment mainly uses an IRIS+ (quadrotor) virtual unmanned aerial vehicle equipped with ArduPilot 4.3.0 firmware, and runs on an APM simulator. All experiments were performed on an Ubuntu 20.04 desktop configured with a 10-core 2.5GHz Intel Core i5 CPU and 32GB RAM. The genetic algorithm part was implemented using the Geatpy toolkit, with a population size of 25 and a maximum evolution number of 100.
[0052] Verification of the accuracy of the fitness function: We identified sensor features that significantly affect specific abnormal states through SHAP analysis, as shown in Figure 2 , and determined their abnormal threshold values using box plots, as shown in Figure 3 . The detection performance of the fitness function constructed based on this was excellent under four abnormal states, with F1 scores of more than 85% for three states ("thrust loss", "yaw imbalance", and "rotation") except for the "crash" state with an F1 score of 67.86%, verifying the effectiveness of the fitness function in determining the abnormal state of the unmanned aerial vehicle. The specific results are shown in the following table.
[0053]
[0054] Search efficiency verification: In 2500 search iterations, the search efficiency of the method of the application for "yaw imbalance" and "rotation" anomalies is more than 75%. For "thrust loss" and "crash" anomalies with higher feature dimensions, the efficiencies are 66.1% and 38.9% respectively as shown in the following table. This shows that although the efficiency of the method in the high-dimensional feature space has decreased, the search efficiency of the method is still significantly superior to other fuzzy testers that have not been optimized.
[0055]
[0056] Correlation parameter effectiveness verification: In order to verify the effectiveness of the parameter subset selected by the Pearson correlation analysis, a set of comparative experiments were conducted. The experimental results shown in the following table show that the efficiency of inducing abnormal states by varying the selected strongly correlated parameters is significantly higher than that of varying all parameters. For example, in the "yaw imbalance" test, the success rate of the partial parameter variation strategy is 62.8%, while the success rate of the all parameter variation strategy is only 53.6%. This result strongly proves that the strategy of the application for reducing the test space through correlation analysis is efficient and practical.
[0057]
[0058] Corresponding to the foregoing embodiment of the data-driven unmanned aerial vehicle configuration parameter vulnerability fuzzy test method, the application also provides an embodiment of a data-driven unmanned aerial vehicle configuration parameter vulnerability fuzzy test device.
[0059] Referring to Figure 5 , the embodiment of the application provides a data-driven unmanned aerial vehicle configuration parameter vulnerability fuzzy test device, which comprises a memory and one or more processors, the memory stores executable code, and the processor executes the executable code to implement the data-driven unmanned aerial vehicle configuration parameter vulnerability fuzzy test method in the foregoing embodiment.
[0060] The embodiment of the data-driven unmanned aerial vehicle configuration parameter vulnerability fuzzy test device provided by the application can be applied to any device with data processing capability, which can be a device or apparatus such as a computer. The device embodiment can be realized by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a logical device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory and running by the processor of the device with data processing capability where it is located. From the hardware level, as shown in Figure 5 , it is a hardware structure diagram of the device with data processing capability where the data-driven unmanned aerial vehicle configuration parameter vulnerability fuzzy test device provided by the application is located, in addition to theFigure 5 In addition to the processor, the memory, the network interface, and the nonvolatile memory, any data processing capable device in which the apparatus of the embodiments is implemented can also include other hardware according to the actual functions of the data processing capable device, which will not be described herein.
[0061] The implementation processes of the functions and roles of the units in the apparatus are specifically described in the implementation processes of the corresponding steps in the above method, which will not be described herein.
[0062] For the apparatus embodiments, since they basically correspond to the method embodiments, the related parts can be referred to the part of the method embodiments. The apparatus embodiments described above are merely illustrative, and the units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. Some or all of the modules can be selected to achieve the purposes of the present application according to actual needs. Those skilled in the art can understand and implement them without creative labor.
[0063] The embodiments of the present application also provide a computer readable storage medium, which stores a program, and the program is executed by a processor to implement the data-driven unmanned aerial vehicle configuration parameter vulnerability fuzz testing method in the above embodiments.
[0064] The computer readable storage medium can be an internal storage unit of any data processing capable device, such as a hard disk or a memory. The computer readable storage medium can also be an external storage device of any data processing capable device, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. Further, the computer readable storage medium can include both the internal storage unit and the external storage device of any data processing capable device. The computer readable storage medium is used to store the computer program and other programs and data required by the data processing capable device, and can also be used to temporarily store data that has been output or will be output.
[0065] The present application also provides a computer program product, which includes a computer program, and the computer program is executed by a processor to implement the data-driven unmanned aerial vehicle configuration parameter vulnerability fuzz testing method.
[0066] Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the specification and examples be considered as exemplary only, with the true scope and spirit of the application being indicated by the following claims.
[0067] It is to be understood that the above general description and the subsequent detailed description are merely exemplary and explanatory and are not restrictive of the application. The application is not restricted to the precise construction and combinations of parts and steps described above and shown in the accompanying drawings, as such variations and modifications are intended to be included within the scope of the application. The scope of the application is limited only by the claims that follow.
Claims
1. A data-driven fuzzy testing method for the vulnerability of UAV configuration parameters, characterized in that, The method includes: Perform flight missions in the simulator to pre-collect drone flight logs; The system categorizes various abnormal states of the drone until the number of abnormal state logs reaches a preset threshold. Perform iterative optimization: merge flight logs from abnormal states with randomly sampled normal logs, and perform SHAP analysis on the merged dataset; The fitness function of the genetic algorithm is updated based on the SHAP analysis results. At the same time, Pearson correlation analysis is performed on the joint configuration parameters of the relevant sensor features obtained from the SHAP analysis to obtain the relevant configuration parameters and update the population range of the genetic algorithm. Based on the fitness function and the updated population range, a new round of configuration parameters is generated by mutation using a genetic algorithm. These parameters are then input into the simulator to generate mutated flight logs. The fitness function value for this flight is calculated and used for the next iteration of population mutation. The obtained flight logs are classified according to the anomaly type. Iterate until the relevant configuration parameters stop updating or the maximum number of evolutions is reached, then output the test results.
2. The data-driven fuzzy testing method for the vulnerability of UAV configuration parameters according to claim 1, characterized in that, The pre-collection of UAV flight logs includes: obtaining code related to abnormal state determination from the source code of the UAV control program, defining a corresponding initial fitness function, and guiding the rapid collection of abnormal state logs through an initial genetic algorithm; the initial fitness function is obtained through the values of key sensors during flight and the abnormal threshold, wherein the key sensors and their corresponding weights are obtained through the safety check code logic; the flight log data includes: flight attitude, acceleration, GPS position, speed, motor throttle output, internal monitoring variables, and a list of current flight configuration parameters.
3. The data-driven fuzzy testing method for the vulnerability of UAV configuration parameters according to claim 1, characterized in that, The classification of various abnormal states of the UAV includes: based on the UAV's physical behavior and the security definitions in the firmware, the abnormal states are divided into: crash, potential thrust loss, yaw imbalance or unexpected autonomous rotation.
4. The data-driven fuzzy testing method for the vulnerability of UAV configuration parameters according to claim 1, characterized in that, The process of merging flight logs from anomalous states and randomly sampled normal logs, and then performing SHAP analysis on the merged dataset, specifically includes: merging B anomalous logs and B randomly sampled normal logs, where B is the threshold for the number of anomalous states; training a random forest classifier model to learn the relationship between sensor features and anomalous states; using the SHAP method to assign a contribution score to each sensor, quantifying the influence of this feature on the anomalous state; and outputting a list of key sensor features most relevant to the current target anomalous state, including the sensor's anomalous threshold and contribution score.
5. The data-driven fuzzy testing method for the vulnerability of UAV configuration parameters according to claim 1, characterized in that, The step of updating the fitness function of the genetic algorithm based on the SHAP analysis results specifically includes: First, update the fitness function of the genetic algorithm by using the identified key sensor features, their anomaly thresholds, and contribution scores to construct the fitness function. The function has the following form: f n =∑W i (S n,i -threshold i ); where W i S is the weight of the i-th key sensor, and its absolute value is equal to the contribution score of that sensor; n,i It is the sensor value in flight segment n; threshold i This is the abnormal threshold of the sensor; weight W i The positive and negative values are used to ensure that abnormal states always obtain positive fitness scores.
6. The data-driven fuzzy testing method for the vulnerability of UAV configuration parameters according to claim 1, characterized in that, The Pearson correlation analysis is performed on the joint configuration parameters of the relevant sensor features obtained from the SHAP analysis. Perform Pearson correlation analysis. Based on the key sensor characteristics, extract the corresponding sensor values S and the configuration parameters P of the previous log from the previous log, and slice the log into segments F, in the form of F = (S, P). Calculate a representative sensor evaluation score (S) for each flight log. score Subsequently, Pearson correlation analysis was used to calculate the relationship between each configuration parameter and the S. score The correlation coefficients between the parameters are calculated, and finally, all configuration parameters with a correlation greater than the threshold are merged to obtain a set P of relevant configuration parameters for this abnormal state. c .
7. The data-driven fuzzy testing method for the vulnerability of UAV configuration parameters according to claim 1, characterized in that, The genetic algorithm population range is updated with the relevant configuration parameters obtained in the following ways: Based on the obtained set of abnormal state related configuration parameters P c This updates the scope of the genetic algorithm population, ensuring that subsequent genetic algorithm mutations only target this relevant configuration parameter set P. c Modifications and mutations will be performed, and all unselected irrelevant parameters will be set to their default values when a new individual is generated.
8. The data-driven fuzzy testing method for the vulnerability of UAV configuration parameters according to claim 1, characterized in that, In the process of generating a new round of configuration parameters using a genetic algorithm, the mutation strategy adopted is the differential evolution algorithm.
9. A data-driven fuzz testing device for the vulnerability of UAV configuration parameters, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it implements a data-driven fuzz testing method for the vulnerability of UAV configuration parameters as described in any one of claims 1-8.
10. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements a data-driven fuzzy testing method for the vulnerability of UAV configuration parameters as described in any one of claims 1-8.