A communication security protection method and system of a UAV

Abstract

The application discloses a kind of communication security protection method and system of unmanned plane, applied to unmanned plane technical field, method includes obtaining the historical communication telemetry data of unmanned plane to be protected;Periodic communication data is obtained by cyclic coding processing based on the time stamp information extracted from historical communication telemetry data to historical communication telemetry data;Characteristic information extracted from periodic communication data is analyzed, and normal communication characteristics are obtained by screening characteristic information with analysis result;Determine disturbance amount;Invasion sample data set is constructed by processing historical communication telemetry data based on disturbance amount;Real-time state data obtained is input into stacked intrusion detection model and processed, to obtain the invasion detection result of unmanned plane to be protected;Control unmanned plane to be protected to execute security protection strategy.The communication security protection method and system of unmanned plane provided by the application effectively improve the detection accuracy of unmanned plane facing various invasions.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and in particular to a communication security protection method and system for UAVs. Background Technology

[0002] Drones have been widely used in key areas such as logistics delivery, emergency rescue, agricultural plant protection, and urban air transportation. However, with their increasing networking and intelligence, building an efficient communication security protection mechanism has become an urgent need to ensure the safe and reliable operation of drones in complex electromagnetic and network environments.

[0003] Existing communication security protection methods typically utilize known raw state data of drones, whether normal or compromised, such as telemetry information, control commands, and network traffic logs. These data are matched or compared based on known attack patterns to identify abnormal behavior in the communication link, thereby achieving security protection. However, current attacks against drone communication systems are becoming increasingly intelligent and covert. Attackers no longer rely on obvious abnormal traffic or protocol violations, but instead manipulate legitimate communication behavior subtly. For example, in urban logistics flight scenarios, attackers finely modulate the packet length, transmission interval, or sequence timing of legitimate control commands, making malicious communication highly approximate normal mission traffic in terms of statistical characteristics and protocol behavior. Such adversarial perturbations can effectively circumvent traditional communication security protection mechanisms based on fixed thresholds, rule bases, or shallow communication characteristics, leading to a significant increase in the system's false negative rate. This makes it difficult to detect covert communication hijacking and command injection in a timely manner, thus seriously threatening the operational safety and public safety of low-altitude airspace. Summary of the Invention

[0004] This invention provides a communication security protection method and system for unmanned aerial vehicles (UAVs) to solve the technical problem of existing technologies that rely on existing intrusion detection methods to detect intrusions and then implement protection, thereby achieving the effect of accurately detecting various network intrusions and then implementing protection.

[0005] To address the aforementioned technical problems, this invention provides a communication security protection method and system for unmanned aerial vehicles (UAVs), the method comprising: Acquire historical communication telemetry data of the drone to be protected; Based on the timestamp information extracted from the historical communication telemetry data, the historical communication telemetry data is cyclically encoded to obtain periodic communication data; The feature information extracted from the periodic communication data is analyzed, and the analysis results are used to filter the feature information to obtain normal communication features. The feature information is processed using a feature reconstruction mechanism built based on the normal communication characteristics to determine the perturbation amount; Based on the disturbance amount, the historical communication telemetry data is processed to construct an intrusion sample dataset; The real-time status data of the drone to be protected is input into the stacked intrusion detection model trained based on the intrusion sample dataset for processing, so as to obtain the intrusion detection result of the drone to be protected. Control the drone to be protected to execute the security protection strategy generated based on the intrusion detection results.

[0006] Preferably, the step of processing the historical communication telemetry data based on the disturbance amount to construct the intrusion sample dataset includes: The historical communication telemetry data is divided into black-box communication sample datasets and white-box communication sample datasets. Perform a binary search on the black-box communication sample dataset to determine the perturbation strength parameters; The black-box communication sample dataset is processed based on the disturbance intensity parameter and the disturbance amount to determine the first intrusion sample dataset; Based on the disturbance intensity parameter and the disturbance amount, the white-box communication sample dataset and the corresponding white-box sample data in the second feature information after removing the normal communication features are processed to determine the second intrusion sample dataset; The intrusion sample dataset is constructed based on the first intrusion sample dataset and the second intrusion sample dataset.

[0007] Preferably, the step of performing a binary search on the black-box communication sample dataset to determine the perturbation strength parameter includes: The black-box communication sample dataset is trained based on a pre-built basic detection model to obtain the output result corresponding to each sample data. For each sample data in the black box communication sample dataset, the binary search algorithm is used to optimize the control disturbance intensity coefficient within a preset disturbance interval until the corresponding output result is reversed, thereby obtaining each minimum control disturbance intensity coefficient value, and thus obtaining the minimum control disturbance intensity coefficient set. The disturbance intensity parameter is obtained based on the minimum control disturbance intensity coefficient set.

[0008] Preferably, the step of processing the feature information using a feature reconstruction mechanism constructed based on the normal communication features to determine the perturbation amount includes: Based on the aforementioned normal communication characteristics, a feature reconstruction mechanism is constructed; The feature information is reconstructed using the aforementioned feature reconstruction mechanism to obtain a reconstructed feature vector; The reconstructed feature vector and the feature information are subjected to difference processing, and the perturbation amount is determined based on the reconstruction difference.

[0009] Preferably, the step of performing cyclic encoding on the historical communication telemetry data based on the timestamp information extracted from the historical communication telemetry data to obtain periodic communication data includes: Based on the timestamp information, an absolute time series representing the sampling time is obtained; The absolute time series is processed by trigonometric function mapping using a cyclic encoding function to obtain sine-cosine feature pairs. The sine-cosine feature pairs are concatenated with the non-time fields in the historical communication telemetry data in a time-aligned manner to obtain the periodic communication data.

[0010] Another embodiment of the present invention provides a communication security protection system for unmanned aerial vehicles (UAVs), comprising: The acquisition module is used to acquire historical communication telemetry data of the UAV to be protected; The loop module is used to perform cyclic encoding processing on the historical communication telemetry data based on the timestamp information extracted from the historical communication telemetry data to obtain periodic communication data. The filtering module is used to analyze the feature information extracted from the periodic communication data, and to filter the feature information based on the analysis results to obtain normal communication features. The disturbance module is used to process the feature information using a feature reconstruction mechanism built based on the normal communication features to determine the disturbance amount; The construction module is used to process the historical communication telemetry data based on the disturbance amount to construct an intrusion sample dataset; The processing module is used to input the real-time status data of the drone to be protected into the stacked intrusion detection model trained based on the intrusion sample dataset for processing, so as to obtain the intrusion detection result of the drone to be protected. The control module is used to control the drone to be protected to execute the security protection strategy generated based on the intrusion detection results.

[0011] Preferably, the building module includes: A partitioning unit is used to divide the historical communication telemetry data into a black-box communication sample dataset and a white-box communication sample dataset. The search unit is used to perform a binary search process on the black-box communication sample dataset to determine the perturbation strength parameters. The first unit is used to process the black box communication sample dataset based on the disturbance intensity parameter and the disturbance amount to determine the first intrusion sample dataset; The second unit is used to process the white-box communication sample dataset and the corresponding white-box sample data in the second feature information after removing the normal communication features based on the disturbance intensity parameter and the disturbance amount, so as to determine the second intrusion sample dataset. An intrusion unit is used to construct the intrusion sample dataset based on the first intrusion sample dataset and the second intrusion sample dataset.

[0012] Preferably, the step of performing a binary search on the black-box communication sample dataset to determine the perturbation strength parameter includes: The black-box communication sample dataset is trained based on a pre-built basic detection model to obtain the output result corresponding to each sample data. For each sample data in the black box communication sample dataset, the binary search algorithm is used to optimize the control disturbance intensity coefficient within a preset disturbance interval until the corresponding output result is reversed, thereby obtaining each minimum control disturbance intensity coefficient value, and thus obtaining the minimum control disturbance intensity coefficient set. The disturbance intensity parameter is obtained based on the minimum control disturbance intensity coefficient set.

[0013] Preferably, the disturbance module includes: A construction unit is used to construct a feature reconstruction mechanism based on the normal communication features; The reconstruction unit is used to reconstruct the feature information using the feature reconstruction mechanism to obtain a reconstructed feature vector; The determining unit is used to perform difference processing on the reconstructed feature vector and the feature information, and determine the perturbation amount based on the reconstruction difference.

[0014] Preferably, the loop module includes: An absolute time unit is used to obtain an absolute time series representing the sampling time based on the timestamp information. The mapping unit is used to perform trigonometric function mapping on the absolute time series using a cyclic coding function to obtain sine-cosine feature pairs. The splicing unit is used to splice the sine-cosine feature pairs with the non-time fields in the historical communication telemetry data according to time alignment to obtain the periodic communication data.

[0015] Compared with the prior art, the beneficial effects of the present invention are at least one of the following: This invention first acquires historical communication telemetry data of the UAV to be protected, and then performs cyclic encoding on the historical communication telemetry data based on the timestamp information extracted from the data to obtain periodic communication data. This explicitly captures the mission cycle pattern of the UAV on a daily / weekly scale, overcoming the misjudgment caused by neglecting the temporal semantics of traditional methods. Next, the feature information extracted from the periodic communication data is analyzed, and the analysis results are used to filter the feature information to obtain normal communication features, achieving decoupling of key dimensions. Based on this, a feature reconstruction mechanism built based on the normal communication features is used to process the feature information and determine the perturbation amount. This perturbation amount essentially reflects the degree to which the current communication behavior deviates from the normal cycle pattern, accurately characterizing the structural bias introduced by the attacker's fine modulation. Furthermore, based on the perturbation amount, the historical communication telemetry data is processed to construct an intrusion sample dataset, generating adversarial samples with realistic attack semantics for training a more generalized detection model. Finally, the real-time status data of the drone to be protected is input into a stacked intrusion detection model trained based on the intrusion sample dataset for processing, resulting in the intrusion detection result of the drone to be protected. The drone is then controlled to execute a security protection strategy generated based on the intrusion detection result. This invention no longer relies on fixed thresholds or known attack signatures, but instead learns the periodic normal communication baseline of individual drones and utilizes a residual-driven disturbance perception mechanism to significantly improve sensitivity to low-amplitude, highly concealed adversarial disturbances. Simultaneously, the constructed stacked intrusion detection model, having incorporated intrusion samples generated based on real disturbance distributions during the training phase, possesses stronger robustness, effectively reducing the false negative rate and promptly blocking advanced persistent threats such as communication hijacking and command injection, thereby effectively ensuring the safety of low-altitude airspace operations and public safety. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for protecting the communication security of a drone in one embodiment of the present invention. Figure 2 This is a schematic diagram of the communication security protection system for a drone in one embodiment of the present invention; Figure label: Among them, 11. Acquisition module; 12. Loop module; 13. Filtering module; 14. Disturbance module; 15. Construction module; 16. Processing module; 17. Control module. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] In the description of this invention, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0019] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0020] In the description of this invention, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0021] Drones are widely used in critical scenarios such as logistics, emergency response, agriculture, and urban air traffic, posing serious challenges to their communication security. Existing protection methods rely on matching telemetry, command, or traffic logs with known attack patterns, which is insufficient to cope with new and covert attacks. Attackers finely modulate legitimate communication behavior, making malicious traffic highly mimic normal tasks at the statistical and protocol levels, thereby bypassing traditional mechanisms based on fixed thresholds or shallow features, resulting in high false negative rates and seriously threatening low-altitude airspace safety.

[0022] One embodiment of the present invention provides a communication security protection method for unmanned aerial vehicles (UAVs). For details, please refer to [link to specific documentation]. Figure 1 , Figure 1 The diagram shown is a flowchart illustrating a communication security protection method for a drone according to one embodiment of the present invention, including: S1. Obtain historical communication telemetry data of the drone to be protected; S2. Based on the timestamp information extracted from historical communication telemetry data, the historical communication telemetry data is cyclically encoded to obtain periodic communication data. S3. Analyze the feature information extracted from periodic communication data, and use the analysis results to filter the feature information to obtain normal communication features; S4. Process the feature information using a feature reconstruction mechanism based on normal communication features to determine the perturbation amount; S5. Based on the disturbance amount, process historical communication telemetry data to construct an intrusion sample dataset; S6. Input the real-time status data of the drone to be protected into the stacked intrusion detection model trained based on the intrusion sample dataset for processing to obtain the intrusion detection result of the drone to be protected. S7. Control the drone to be protected to execute the security protection strategy generated based on the intrusion detection results.

[0023] First, acquire the historical communication and telemetry data of the drone to be protected. Acquiring this data is a fundamental step in any drone intrusion detection scheme. Its core is collecting various communication data and telemetry parameters generated during the drone's past flights, providing the initial basis for subsequent data processing and model training. The drone to be protected refers to drone equipment requiring intrusion detection and security protection mechanisms, typically used in civilian aerial photography, power line inspection, agricultural plant protection, or military reconnaissance scenarios. Its communication and telemetry data encompasses command transmission data, status feedback data, location information data, sensor data, and equipment operating parameter data between the drone and the ground control station. Historical communication and telemetry data refers to the collection of all communication and telemetry data recorded by the drone during normal flight, abnormal flight, and potential intrusion attacks over a past period. This data must include complete time-series information and corresponding status labels to support subsequent analysis. The specific implementation method involves using a data acquisition module to capture data packets transmitted in real time between the UAV and the ground control station via a wireless communication link. This is combined with local data stored in the UAV's onboard data logger. Historical data is extracted from flash memory through a data read interface. A data synchronization protocol is used to align the time and standardize the format of the multi-source data. Data cleaning techniques are employed to remove noisy, missing, and redundant data. The processed raw data is then transmitted to a local server or cloud storage node via a file transfer protocol. A structured query language is used to classify, store, and index the data. A distributed file system is leveraged to achieve efficient storage and rapid retrieval of large-scale historical data. This approach provides high-quality data support for subsequent data encoding analysis and model training, ensuring the intrusion detection model has the ability to accurately identify intrusion behaviors. Furthermore, the accumulated historical data covers various flight scenarios and potential intrusion patterns, improving the versatility and reliability of the detection solution.

[0024] Secondly, based on the timestamp information extracted from historical communication telemetry data, the historical communication telemetry data is subjected to cyclic encoding to obtain periodic communication data. Based on the timestamp information, an absolute time series representing the sampling time is obtained; a cyclic encoding function is used to perform trigonometric function mapping on the absolute time series to obtain sine-cosine feature pairs; these sine-cosine feature pairs are then concatenated with non-time fields in the historical communication telemetry data according to time alignment to obtain the periodic communication data. The core of obtaining periodic communication data by cyclic encoding based on the timestamp information extracted from historical communication telemetry data is to transform discrete time information into periodic numerical features, strengthening the temporal correlation of the data to adapt to subsequent feature analysis and model training. Timestamp information is the core information in historical communication telemetry data that marks the time of each data acquisition, mostly high-precision time records at the millisecond level. The absolute time series representing the sampling time is a one-dimensional time series formed by arranging the extracted timestamp information in an ordered manner according to the data acquisition sequence, which can intuitively reflect the temporal evolution pattern of the data. Cyclic coding functions are a set of mathematical functions that implement time period mapping. They rely on the periodicity of trigonometric functions to transform time features. Trigonometric function mapping refers to the process of mapping the values ​​of an absolute time series to the interval [-1, 1] using a combination of sine and cosine functions. The sine-cosine feature pair is a set of complementary numerical features obtained after this processing. Sine and cosine features can jointly characterize the periodic changes in time, avoiding the loss of feature information from a single function. Non-time fields are all business parameter fields in historical communication telemetry data except for the timestamp, including the UAV's flight altitude, flight speed, attitude angle, and the signal strength, data packet length, and transmission rate of the communication link.

[0025] First, the structured historical communication telemetry data is read using the `read_csv` or `read_parquet` methods of the Pandas library in Python. The timestamp field is extracted and converted to an absolute timestamp value of type `int64` using the `astype` method. Then, the `sort_values` method is used to sort the data in ascending order of timestamps to generate an absolute time series representing the sampling time. Next, the fundamental period `T` is calculated based on the actual sampling frequency of the UAV communication telemetry data. Common sampling frequencies are 10Hz (T = 0.1 seconds) and 20Hz (T = 0.05 seconds). A cyclic encoding function is then defined based on the NumPy library. The algorithm constructs sine and cosine mapping functions, respectively. Each value in the absolute time series is sequentially substituted into these two functions for vectorization calculation, generating equal-dimensional sine-cosine feature pairs, which are added as new feature columns to the original dataset. Then, using the timestamp field as a unique key, the `merge` method is called to concatenate the sine-cosine feature pairs with non-time fields in the original dataset, aligning them row-wise. After concatenation, the `drop_duplicates` method is called to remove duplicate data, and the `dropna` method is called to remove invalid data with missing timestamps. Finally, the processed dataset is categorized and stored according to communication scenarios, resulting in standardized periodic communication data. This approach quantifies the periodic characteristics of time into computable numerical features, making subsequent feature analysis more likely to capture temporal patterns. It also improves the adaptability of the data to machine learning models and enhances the intrusion detection model's ability to identify time-related intrusion behaviors.

[0026] Next, the feature information extracted from the periodic communication data is analyzed, and the results are used to filter the feature information to obtain normal communication features. Analyzing the feature information extracted from the periodic communication data and filtering the results to obtain normal communication features is to eliminate invalid and redundant features, focus on core features to improve the accuracy of subsequent intrusion detection, and lay the foundation for feature reconstruction and model training. Feature information refers to various quantitative indicators extracted from periodic communication data that characterize the communication status of the UAV, including time-series statistical indicators, communication link features, and UAV operational status features. Time-series statistical indicators include the mean, variance, and trend change rate of the sine-cosine feature pair; communication link features include signal strength fluctuation values, data packet transmission delay, and packet loss rate; and UAV operational status features include attitude angle changes, flight speed stability, and battery voltage fluctuation range.

[0027] First, a feature extraction algorithm is used to extract the aforementioned multi-dimensional feature information from periodic communication data. The feature extraction module is invoked using the Python programming language, and statistical methods are combined to calculate the mean, variance, range, and correlation coefficient of each feature. Simultaneously, the local anomaly factor algorithm from unsupervised learning is used to perform anomaly correlation analysis on the feature information, identifying feature fluctuation patterns related to potential intrusion behaviors. Then, the mutual information algorithm is used to calculate the correlation degree between each feature and the communication status label, setting a correlation threshold of 0.3 to 0.5 to filter low-correlation features. Subsequently, a recursive feature elimination algorithm combined with cross-validation is used to gradually eliminate features with low contribution to the classification results. At the same time, by comparing the feature distribution differences between normal communication scenarios and simulated intrusion scenarios, features that are consistently stable in normal scenarios and unrelated to intrusion behaviors are retained as normal communication features. During the screening process, feature normalization is simultaneously used to unify feature dimensions, ensuring the reliability and consistency of the feature screening results. This approach simplifies feature dimensions, reduces model computational complexity, makes subsequent feature reconstruction and model training more efficient, and accurately distinguishes the feature differences between intrusion and normal communication, improving the recognition accuracy and generalization ability of the intrusion detection model.

[0028] Then, a feature reconstruction mechanism based on normal communication characteristics is used to process the feature information and determine the perturbation amount. A feature reconstruction mechanism is constructed based on normal communication characteristics; the feature information is reconstructed using this mechanism to obtain a reconstructed feature vector; the difference between the reconstructed feature vector and the original feature information is processed, and the perturbation amount is determined based on the reconstruction difference. A benchmark model is established using normal communication characteristics to quantify the deviation between the feature information and the normal benchmark to locate potential intrusion interference. The feature reconstruction mechanism is a feature restoration model trained based on normal communication characteristics. It can simulate the feature distribution patterns under normal communication scenarios, achieving normalized reconstruction of input feature information. The reconstructed feature vector is a feature vector output by the feature reconstruction mechanism after processing the feature information, conforming to the patterns of normal communication characteristics. Its dimension is consistent with the input feature information. The perturbation amount is a quantitative indicator characterizing the degree of deviation between the input feature information and the reconstructed feature vector, reflecting the strength of the feature information being affected by intrusion interference.

[0029] Based on the selected normal communication features, an autoencoder is used to construct a feature reconstruction mechanism. The autoencoder model is built using Python and the TensorFlow framework. The dimensions of the input and output layers match the dimensions of the normal communication features. An activation function is used in the hidden layers to reduce the risk of gradient vanishing. Mean squared error is used as the loss function. The model is trained using a preprocessed normal communication feature dataset, and the model parameters are iteratively optimized until the loss function converges to a stable range. After training, a feature reconstruction mechanism with normal feature reconstruction capabilities is obtained. The selected feature information is then input into this mechanism for reconstruction, resulting in corresponding reconstructed feature vectors. Subtraction operations are used to subtract the reconstructed feature vectors from the original feature information dimension by dimension, resulting in a reconstructed difference sequence. The L2 norm is used to calculate the overall deviation of the reconstructed difference sequence. Simultaneously, the weighted summation of the feature weights for each dimension is used to determine the final perturbation. The weights are obtained through feature importance analysis, assigning higher weights to features that contribute significantly to intrusion detection. This approach accurately quantifies the abnormal deviation of feature information, providing a core basis for subsequently constructing an intrusion sample dataset, while simultaneously improving the sensitivity to covert intrusion behavior and ensuring the accuracy of intrusion detection.

[0030] Furthermore, an intrusion sample dataset is constructed by processing historical communication telemetry data based on the disturbance amount. The historical communication telemetry data is divided into a black-box communication sample dataset and a white-box communication sample dataset. A binary search is performed on the black-box communication sample dataset to determine the disturbance intensity parameter. Based on the disturbance intensity parameter and the disturbance amount, the black-box communication sample dataset is processed to determine the first intrusion sample dataset. Based on the disturbance intensity parameter and the disturbance amount, the corresponding white-box sample data in the white-box communication sample dataset and the second feature information after removing the normal communication features are processed to determine the second intrusion sample dataset. An intrusion sample dataset is constructed based on the first and second intrusion sample datasets. The black-box communication sample dataset is trained using a pre-built basic detection model to obtain the output result corresponding to each sample data. For each sample data in the black-box communication sample dataset, a binary search algorithm is used to optimize the control disturbance intensity coefficient within a preset disturbance interval until the corresponding output result is reversed, obtaining each minimum control disturbance intensity coefficient value, and thus obtaining the minimum control disturbance intensity coefficient set. Based on the minimum control disturbance intensity coefficient set, the disturbance intensity parameter is obtained.

[0031] This study constructs an intrusion sample dataset by processing historical communication telemetry data based on perturbation. The core idea is to simulate various intrusion scenarios through precise perturbation, generating intrusion samples covering different modes. This provides sufficient and realistic sample support for subsequent training of stacked intrusion detection models. Black-box communication sample datasets refer to collections of communication samples whose internal structure and parameters are not publicly disclosed. These are mostly data collected in real flight environments without explicit state labels, covering communication telemetry data under normal drone flight, unknown minor intrusion, and covert intrusion scenarios. White-box communication sample datasets refer to collections of communication samples whose internal structure, parameters, and state labels are publicly disclosed. They include clearly labeled normal communication samples, known intrusion type samples, and corresponding feature parameter configurations. The perturbation intensity parameter is the core quantitative value controlling the degree of perturbation application, obtained through statistical optimization from the minimum control perturbation intensity coefficient set. The minimum control perturbation intensity coefficient is used to ensure that the samples meet the basic detection requirements. The minimum perturbation coefficient at which the output result in the model is reversed is the critical coefficient at which a sample originally judged as normal becomes abnormal or a sample originally abnormal becomes normal. The second feature information after removing the normal communication features is the intrusion-related feature. The first intrusion sample dataset is the set of intrusion samples generated after processing black-box samples with perturbation intensity parameters and perturbation amounts. The second intrusion sample dataset is the set of intrusion samples generated after directional perturbation of samples corresponding to intrusion-related features in white-box samples. The basic detection model is a pre-built simplified intrusion detection model used to provide the basis for output result judgment for optimizing the perturbation coefficient of black-box samples.

[0032] First, historical communication telemetry data is divided into black-box and white-box communication sample datasets based on sample transparency and label integrity, with a 7:3 ratio to ensure sample diversity and representativeness. The black-box dataset selects unlabeled real flight data, covering communication telemetry data from different flight scenarios and time periods. The white-box dataset selects fully labeled samples, including normal communication samples as well as three types of known intrusion samples: signal interference, command tampering, and data forgery. Then, a basic detection model is built, employing Logistic Regression (LR). The Logistic Regression (LR) algorithm was constructed using Python to configure model parameters, setting the regularization coefficient to 0.01 and the number of iterations to 1000. Using previously filtered feature information as input, the model was trained with normal samples and known intrusion samples from a white-box dataset. Parameters were iteratively adjusted until the model accuracy stabilized above 85% and the recall was not less than 80%. After training, a black-box communication sample dataset was input to obtain the output result for each sample, representing a preliminary judgment of normal or abnormality. Then, a binary search algorithm was used to optimize the control perturbation intensity coefficient for each sample in the black-box dataset within a preset perturbation interval of 0 to 2. The upper limit of the interval was initialized to 2 and the lower limit to 0. The iteration termination condition was set as the interval accuracy reaching 0.001 and the output results being consistent for three consecutive iterations. The median value of the interval was calculated as the current perturbation coefficient, and this coefficient was compared with the perturbation amount... The weights are applied to the sample features after being superimposed. The weights are allocated according to the importance of the features. The weight of intrusion-related features is set to 0.7, and the weight of other features is set to 0.3. After superposition, the perturbed sample features are obtained and input into the basic detection model to obtain new output results. The original output results are compared to determine whether the perturbation has been reversed. If the perturbation has not been reversed, the interval is adjusted according to the results. If the original is normal and it is still normal after perturbation, the lower limit of the interval is updated to the median value. If the original is abnormal and it is still abnormal after perturbation, the upper limit of the interval is updated to the median value. The iterative process is repeated until the output results are reversed, and the minimum control perturbation strength coefficient value of the sample is obtained. All samples in the black-box dataset are traversed to obtain the minimum control perturbation strength coefficient set. Outliers (coefficients that deviate from the mean by more than 3 times the standard deviation) are first removed from this coefficient set. Then, the mean of the remaining coefficients is taken and combined with the standard deviation for correction to obtain the final perturbation strength parameter. The correction formula is: perturbation strength parameter = mean + 0.Using a 3× standard deviation parameter and perturbation amount, the black-box communication sample dataset is perturbed sample by sample. The feature values ​​of each sample are adjusted according to the perturbation coefficient, while controlling the fluctuation of feature values ​​to not exceed 20% of the normal range to ensure sample rationality, generating the first intrusion sample dataset. Simultaneously, for the white-box communication sample dataset, white-box sample data corresponding to intrusion-related features are extracted, and targeted perturbations are performed using the same perturbation strength parameter and amount, focusing on enhancing the abnormal performance of intrusion-related features. Only slight perturbations are applied to non-intrusion-related features to simulate real intrusion scenarios, generating the second intrusion sample dataset. Finally, the first and second intrusion sample datasets are merged. Duplicate samples are removed based on timestamp and core feature combinations, and invalid samples missing key features are deleted. The merged dataset is then subjected to feature normalization processing to unify the units of measurement, completing the construction of the intrusion sample dataset. This approach generates diverse intrusion samples covering both black-box and white-box scenarios, improving the model's adaptability to different types of intrusion behaviors. Furthermore, determining parameters based on the minimum perturbation coefficient ensures the authenticity and validity of the samples.

[0033] Preferably, the real-time status data of the drone to be protected is input into a stacked intrusion detection model trained on an intrusion sample dataset for processing to obtain the intrusion detection results of the drone to be protected. Relying on a maturely trained model to accurately analyze real-time data, potential intrusion behaviors are quickly identified, providing a basis for subsequent security protection strategy formulation. Real-time status data refers to various data collected and transmitted by the drone to be protected during its current flight, including real-time location information, flight attitude parameters, communication link status data, and onboard equipment operating parameters, covering flight altitude, speed, heading angle, signal strength, data packet transmission rate, and battery voltage. The stacked intrusion detection model is an ensemble learning model that integrates the output results of multiple basic models. It improves detection accuracy through a hierarchical stacking structure, combining the advantages of each basic model while overcoming the limitations of a single model. Its core consists of a basic model layer and a meta-classifier layer. The basic model layer is responsible for preliminary detection, and the meta-classifier layer is responsible for integrating the results and outputting the final detection conclusion.

[0034] First, a stacked intrusion detection model architecture is built. Three models can be used as the base model layers: LR, SVM (Support Vector Machine), and RF (Random Forest). The parameters of each model are configured using Python libraries. The LR model's regularization coefficient is set to 0.01 and the number of iterations to 1000. The SVM model uses a radial basis function kernel with a penalty coefficient of 1.0. The RF model has 100 decision trees and a maximum depth of 10. Then, the constructed intrusion sample dataset is divided into training and validation sets in an 8:2 ratio. The three base models are trained on the training set, and the parameters are iteratively optimized until the accuracy of each model on the validation set reaches above 88%. The outputs of each base model are then used as input to the meta-classifier layer. A gradient boosting tree is used as the meta-classifier, and iterative training is performed on the training set to learn and integrate the outputs of each base model. After training is complete... The model was then optimized to ensure an overall accuracy of no less than 92% on the validation set and a false positive rate of less than 5%. A real-time data acquisition channel was then established to capture the status data of the drone under protection in real time using onboard sensors and communication modules. Following the same preprocessing steps described earlier, the real-time data was cleaned, normalized, and feature extracted to generate real-time feature vectors consistent with the model's input dimensions. These real-time feature vectors were then input into a well-trained stacked intrusion detection model. The basic model layers detected the intrusions and output preliminary results. The meta-classifier layer weighted and integrated these preliminary results, outputting the final intrusion detection results. The results included four categories: normal communication, signal interference intrusion, command tampering intrusion, and data forgery intrusion, along with labeled detection confidence levels. This approach balances detection accuracy and speed, improves the ability to identify various intrusion behaviors, reduces the probability of false positives and false negatives, and provides reliable support for real-time drone security protection.

[0035] Finally, the system controls the drone to be protected to execute the security protection strategy generated based on the intrusion detection results. This translates the intrusion detection conclusions into specific protective actions, quickly curbing the spread of intrusion and reducing communication security risks and the probability of flight accidents for the drone. The security protection strategy refers to a customized response plan based on different intrusion detection results, covering three categories of measures: communication link protection, flight status management, and data security assurance. It is precisely matched and executed according to different intrusion types to ensure the effectiveness and relevance of the protection.

[0036] First, a detection result parsing module was built. A Python program was used to identify and classify the results output by the stacked intrusion detection model, clarifying the intrusion type, confidence level, and scope of impact. For normal communication results, a strategy to maintain the current state was generated; for signal interference intrusions, a communication link switching strategy was generated; for command tampering intrusions, a command verification and rollback strategy was generated; and for data forgery intrusions, a data verification and discard strategy was generated. Then, a strategy distribution channel was constructed. A secure communication link between the ground control station and the UAV to be protected was established using a message queue telemetry transmission protocol to ensure the real-time performance and security of strategy command transmission. Simultaneously, a command encryption mechanism was set up, using advanced encryption standards to encrypt the protection strategy commands to prevent them from being intercepted. Secondary tampering prevention involves the drone receiving encrypted commands and decrypting and verifying them via its onboard encryption module. Once the command's legitimacy is confirmed, corresponding protective actions are triggered. For signal interference protection, the drone immediately switches to a backup communication frequency band, shuts down the interfered frequency band, and strengthens the backup band's signal strength. For command tampering protection, a command verification mechanism is activated, comparing the current command with historical legitimate command characteristics, rolling back to the most recent legitimate command state, and locking command reception permissions. For data forgery protection, forged data is discarded, local backup data is used to fill gaps, and a data verification process is initiated. During the execution of these protective actions, drone status data is collected in real time and fed back to the ground control station to confirm the protection effect. If the protection fails to meet expectations, an upgraded protection strategy is triggered until the intrusion risk is eliminated. This approach achieves closed-loop management of intrusion detection and protection actions, enabling rapid response to various intrusion attempts and ensuring drone flight safety and communication data integrity.

[0037] Another embodiment of the present invention provides a general intrusion detection model that is independent of specific UAV hardware platforms or mission types, possessing strong cross-platform generalization capabilities. To achieve this goal and lay the foundation for subsequently improving the adversarial robustness of non-gradient models, the system employs a stacked ensemble learning strategy, fusing multiple base learners with different inductive biases—covering both traditional machine learning models and deep learning models—to effectively address the complexity and uncertainty brought about by heterogeneous UAV data distribution. The entire technical process comprises four key stages.

[0038] First, time-aware feature engineering is performed. Considering that many network attack behaviors (such as periodic port scans or timed data infiltration) exhibit clear temporal regularities, and that different UAV platforms differ in clock synchronization accuracy and data sampling frequency, directly using raw timestamps would cause the model to become overly sensitive to absolute time, reducing generalization performance. Therefore, the system extracts timestamp information from the raw state data and performs cyclic encoding on periodic units such as "hours" and "minutes." This method maps periodic time quantities to a continuous two-dimensional feature space using sine and cosine functions, making "23:00" and "1:00" naturally adjacent in the feature space, thus preserving the periodic structure of time while eliminating the influence of clock deviations between platforms. Furthermore, the system synchronously constructs derived time-series features such as time intervals, event durations, and cumulative runtime. These features together constitute a robust feature set that is platform-independent, attack-mode-sensitive, and comprehensively characterizes the dynamic evolution of network behavior.

[0039] Secondly, key feature screening is performed. Not all features have equal discriminative value across various attack scenarios or different drone platforms. To identify the most universally applicable intrusion-related signals, the system employs two tree-based ensemble models—random forest and extreme gradient boosting—as evaluation tools. Based on intrinsic mechanisms such as Gini impurity reduction and split gain, respectively, each model calculates an importance score for each feature on a comprehensive training set integrating data from multiple platforms. Subsequently, the scores of the two models are normalized, and their average is used as the global importance index. On this basis, a dynamic threshold strategy is adopted: features are sorted in descending order of importance, retaining the top N features whose cumulative contribution reaches 90% of the total importance, ensuring that the selected subset statistically covers the vast majority of discriminative information, while adapting to different data scales. To further ensure feature stability, the system repeats the above evaluation process on multiple hierarchical subsets—including division by drone platform type (e.g., quadcopter, fixed-wing) and by attack type (e.g., denial-of-service, GPS spoofing). The features ultimately retained must stably lie within the high importance range determined by the dynamic threshold in more than 80% of the subsets. This mechanism is equivalent to equipping the model with a "smart filter," enabling it to focus on common discrimination signals that are cross-platform, cross-attack type, and highly stable.

[0040] The third step is to construct and train a stacked ensemble model. A single model is susceptible to noise interference or overfitting, making it difficult to balance diversity and robustness. Therefore, the system abandons a single architecture and instead adopts a layered fusion stacking strategy: first, seven types of base learners with complementary inductive biases are trained in parallel, including logistic regression, K-nearest neighbors, decision trees, random forests, multilayer perceptrons, one-dimensional convolutional neural networks, and TabNet; the hyperparameters of each model are optimized through grid search or random search on the validation set to maximize the F1 score. Subsequently, a lightweight meta-learner (using logistic regression) is introduced, whose input is the attack probability vector output by all base learners for the same sample. To prevent information leakage and overfitting, the training data for the meta-learner is generated through five-fold cross-validation—that is, the predicted probability of each sample is generated from datasets not involved in the training of its base models. L2 regularization and early stopping mechanisms are also incorporated during training to further improve generalization ability. This two-layer architecture not only significantly enhances the model's adaptability to unknown platforms and attacks, but its non-gradient components (such as decision trees and random forests) also provide key support for subsequent adversarial hardening.

[0041] The fourth step is to conduct rigorous cross-platform generalization capability verification. To objectively assess the model's versatility, the system adopts an industry-recognized stringent testing paradigm: in a dataset containing three typical UAV configurations (quadcopter, fixed-wing, and VTOL), training is performed using data from only two of these platforms, while the data from the third platform is entirely reserved for testing. This setting ensures that the test platform is completely invisible during the training phase. Only when the model achieves a harmonic mean score of at least 0.85 for overall accuracy and recall, and an accuracy of at least 0.90 under these conditions, is it considered to possess qualified cross-platform versatility. These thresholds stem from in-depth analysis of actual UAV security operation requirements; falling below these standards will result in excessively high risks of false negatives or missed detections, failing to meet deployment requirements. Models that pass this verification constitute the basic general-purpose intelligent intrusion detection system.

[0042] After obtaining a base model with strong generalization capabilities, the system enters the adversarial robustness enhancement phase. This module is specifically designed to address the vulnerability of non-gradient models (such as decision trees and random forests) to adversarial attacks, constructing a complete technical closed loop of "threat modeling—attack generation—sample validation—adversarial training." Its core innovation lies in proposing a model-independent adversarial sample generation method, completely eliminating the dependence on gradient information of the target model.

[0043] Specifically, the system first defines two types of attacker models: a white-box attacker, who possesses all the internal details of the model, is used to assess the lower bound of robustness in the worst-case scenario; and a black-box attacker, who can only probe the model through input-output interaction, which is closer to real-world threats. For these two scenarios, the system utilizes an autoencoder to learn the potential data distribution of normal network traffic (i.e., the "normal manifold"). This autoencoder is trained on clean, normal samples and can reconstruct data that conforms to legitimate communication patterns. For any input sample, the system calculates the difference between its autoencoder reconstruction result and the original value, i.e., the "reconstruction residual." This residual is not random noise, but rather reflects the structural deviation of the sample from the normal pattern.

[0044] In black-box attack scenarios, the system directly superimposes the residual onto the original sample with a certain strength to generate adversarial examples; the perturbation strength is adaptively determined through binary search to achieve prediction reversal with minimal modification. In white-box scenarios, the system combines feature importance ranking and applies perturbations only to the top-ranked key features (cumulative importance contribution exceeding 70%), achieving "precision-guided" attacks that efficiently mislead the model with minimal overall modification. Because the perturbation originates from the data's inherent distribution characteristics, its concealment is far greater than that of Gaussian noise.

[0045] To ensure the physical plausibility of adversarial examples, the system introduces a 3σ filtering mechanism: for each feature, its mean and standard deviation are calculated based on all normal training data to construct a reasonable value range. Any adversarial examples exceeding this range are eliminated, thus ensuring that subsequent training and evaluation are based on realistic and feasible attack hypotheses.

[0046] Finally, the system performs adversarial training: the filtered high-quality adversarial examples are proportionally mixed with their original labels (keeping them consistent with clean examples) to form an augmented training set, and the entire stacked ensemble model is retrained from scratch using this set. Since the base learners contain non-differentiable components, parameter fine-tuning is impossible; therefore, all base and meta-models must be completely reconstructed. This process forces the model to actively adapt to malicious perturbations driven by the data manifold while learning attack discrimination capabilities, thereby learning more robust decision boundaries that are less susceptible to being fooled by small input changes.

[0047] Through this entire process of construction and reinforcement, the final system not only possesses excellent cross-platform general detection capabilities, but also effectively resists complex adversarial attacks against non-gradient models, providing highly reliable, plug-and-play intelligent protection for the safe operation of heterogeneous drone swarms.

[0048] The system's workflow is clearly divided into two major stages: "offline training and hardening" and "online deployment and detection," forming a complete closed loop from model building to practical application.

[0049] During the offline training and hardening phase, the system first conducts multi-source heterogeneous data collection, extensively acquiring telemetry data and network communication traffic from various UAV platforms (such as quadcopters, fixed-wing aircraft, and VTOL drones) under normal operation and typical network attack conditions (such as GPS spoofing and denial-of-service attacks), constructing a comprehensive dataset covering diverse hardware configurations and attack types. Subsequently, this dataset is fed into the generalization enhancement module, undergoing core steps such as time-aware feature engineering, key feature selection, and stacked ensemble model training, and is evaluated through a rigorous cross-platform verification mechanism. Only models that pass this verification are recognized as basic general-purpose intelligent intrusion detection systems with strong generalization capabilities. Based on this, the system further conducts adversarial vulnerability assessment: using the autoencoder reconstruction residual method proposed in this invention—a model-independent attack generation technique specifically designed for non-gradient models—adversarial samples are constructed in both black-box and white-box scenarios, and a 3σ rule is introduced to filter the samples for physical plausibility, eliminating unrealistic disturbances that exceed the natural fluctuation range of real traffic. Finally, the filtered high-quality adversarial samples are mixed with the original clean samples in a preset ratio to form an enhanced training set. This set is then used to perform a complete adversarial retraining on the base model, thereby outputting a final model that combines excellent cross-platform versatility with high-order adversarial robustness—that is, an adversarial robust general intelligent intrusion detection system.

[0050] Once in the online deployment and detection phase, the system collects real-time traffic data streams from drones or their networks in the urban air traffic environment. This data undergoes the same feature engineering pipeline processing as in the offline phase (including the same timestamp cyclic encoding strategy), is transformed into standardized feature vectors, and input into the deployed adversarial robust model. The model then outputs a judgment result of the current state (normal or attack). Upon detecting intrusion, the system immediately triggers an alarm and can coordinate with external security facilities such as firewalls and flight controllers to execute automatic or manual intervention measures, achieving an integrated security closed loop of "perception-decision-response."

[0051] This invention achieves truly cross-platform, plug-and-play, high-precision detection. Unlike traditional solutions that require customized models for a single platform, leading to high deployment costs and migration difficulties, this solution effectively decouples detection performance from specific hardware or tasks through a time-aware cyclic feature encoding and a stacked integration architecture that fuses multiple models. On a test set including five types of heterogeneous UAVs (hexacoros, quadcopters, fixed-wing aircraft, tail-sitting VTOL, and standard VTOL), a rigorous Leave-One-Platform-Out validation strategy was employed. The model achieved stable performance on completely new platforms, with a harmonic mean score exceeding 0.87 for overall accuracy and recall, and an accuracy exceeding 0.92. This fundamentally solves the bottleneck of needing to re-collect data and retrain models when deploying on new platforms, providing a rapidly deployable standardized solution for the safety management of large-scale, multi-model UAV formations.

[0052] Meanwhile, this solution is the first to systematically overcome the robustness challenge of non-gradient models in adversarial environments. Existing research mostly focuses on differentiable models such as deep learning, neglecting the defense needs of non-gradient models such as decision trees and random forests widely deployed in edge devices. This invention successfully extends efficient adversarial training capabilities to such models through an adversarial example generation mechanism based on data manifold reconstruction. Experiments show that under white-box attacks, the probability of the system being successfully compromised is significantly reduced from approximately 60% to below 15%; in more realistic black-box scenarios, the attack success rate is also greatly reduced from approximately 40% to below 10%. Even in the worst-case scenario where the attacker has complete knowledge of the model's internal details, the system can still maintain a high detection rate, greatly improving the reliability of security protection.

[0053] To avoid a disconnect between "laboratory robustness" and real-world scenarios, this solution specifically introduces an adversarial example filtering mechanism based on the 3σ rule. This mechanism uses the statistical distribution of all normal training data as a benchmark, sets a reasonable physical value range for each feature, and retains only adversarial examples falling within this range for training. This "physical rationality check" ensures that the defense strategy is based on realistic and feasible attack assumptions, significantly enhancing the reliability of the technical solution's migration from simulation environments to engineering deployments.

[0054] Furthermore, the system supports flexible resource adaptation and deployment. Based on a unified technical framework, it offers two configurations: a lightweight version with a core of efficient traditional machine learning models, characterized by low computational overhead and fast inference speed, suitable for airborne embedded devices and meeting the real-time detection needs of edge devices; and an enhanced version integrating high-capacity models such as deep neural networks, suitable for deployment at ground control stations, supporting high-precision situational awareness and deep analysis. Users can flexibly select or combine different modules to achieve optimal deployment results based on the specific application scenario's trade-off between real-time performance, computing resources, and detection accuracy.

[0055] Another embodiment of the present invention provides a communication security protection system for unmanned aerial vehicles (UAVs). For details, please refer to [link to relevant documentation]. Figure 2 , Figure 2 The diagram shown illustrates the structure of a communication security protection system for a drone according to one embodiment of the present invention, comprising: Module 11 is used to acquire historical communication telemetry data of the UAV to be protected; The loop module 12 is used to perform cyclic encoding on the historical communication telemetry data based on the timestamp information extracted from the historical communication telemetry data to obtain periodic communication data. The filtering module 13 is used to analyze the feature information extracted from periodic communication data, and to filter the feature information based on the analysis results to obtain normal communication features. The disturbance module 14 is used to process feature information using a feature reconstruction mechanism based on normal communication features to determine the disturbance amount; Module 15 is used to process historical communication telemetry data based on the disturbance amount to construct an intrusion sample dataset; The processing module 16 is used to input the real-time status data of the drone to be protected into the stacked intrusion detection model trained based on the intrusion sample dataset for processing, so as to obtain the intrusion detection result of the drone to be protected. The control module 17 is used to control the drone to be protected to execute the security protection strategy generated based on the intrusion detection results.

[0056] Preferably, the building module 15 includes: The partitioning unit is used to divide historical communication telemetry data into black-box communication sample datasets and white-box communication sample datasets. The search unit is used to perform a binary search on the black-box communication sample dataset to determine the perturbation strength parameters. The first unit is used to process the black box communication sample dataset based on the disturbance intensity parameter and the disturbance amount to determine the first intrusion sample dataset; The second unit is used to process the white-box communication sample dataset and the corresponding white-box sample data in the second feature information after removing normal communication features based on the disturbance intensity parameter and the disturbance amount, so as to determine the second intrusion sample dataset. Intrusion unit, used to construct an intrusion sample dataset based on the first intrusion sample dataset and the second intrusion sample dataset.

[0057] Preferably, a binary search process is performed on the black-box communication sample dataset to determine the perturbation strength parameters, including: The black-box communication sample dataset is trained based on a pre-built basic detection model to obtain the output result corresponding to each sample data. For each sample data in the black-box communication sample dataset, the binary search algorithm is used to optimize the control disturbance intensity coefficient within the preset disturbance interval until the corresponding output result is reversed, thereby obtaining the minimum control disturbance intensity coefficient value and thus obtaining the minimum control disturbance intensity coefficient set. The disturbance intensity parameters are obtained based on the minimum control disturbance intensity coefficient set.

[0058] Preferably, the disturbance module 14 includes: The building unit is used to construct a feature reconstruction mechanism based on normal communication characteristics; The reconstruction unit is used to reconstruct feature information using a feature reconstruction mechanism to obtain a reconstructed feature vector; The determination unit is used to perform difference processing between the reconstructed feature vector and the feature information, and to determine the perturbation amount based on the reconstruction difference.

[0059] Preferably, the loop module 12 includes: Absolute time units are used to obtain absolute time series representing sampling moments based on timestamp information. The mapping unit is used to perform trigonometric function mapping on the absolute time series using a cyclic coding function to obtain sine-cosine feature pairs. The splicing unit is used to splice the sine-cosine feature pairs with the non-time fields in the historical communication telemetry data according to time alignment, so as to obtain periodic communication data.

[0060] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0061] Accordingly, embodiments of the present invention provide a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform steps in the communication security protection method for unmanned aerial vehicles as described in the above embodiments, for example... Figure 1 Steps S1 to S7 as described above.

[0062] This invention overcomes the reliance of traditional communication security methods on known attack patterns by synergistically enhancing data generation mechanisms and model robustness, effectively addressing detection blind spots caused by intelligent and covert attacks. Unlike existing technologies that directly input raw telemetry data into the model for discrimination, this solution first transforms timestamps into features with a periodic topological structure through cyclic encoding, constructing periodic communication data so that the system can understand the task context of when and what kind of communication behavior should occur. Based on this, by separating normal communication features and constructing a feature reconstruction mechanism solely based on the former, any subtle tampering with legitimate communication flows is amplified into a quantifiable perturbation during the reconstruction process. This perturbation is not only used to identify anomalies but is also actively used to construct an intrusion sample dataset, thus injecting adversarial samples derived from real data distribution deviations during the training phase, rather than relying on manual rules or random noise. The resulting stacked intrusion detection model possesses intrinsic immunity to structural deception and can accurately distinguish malicious communication disguised as normal tasks when real-time state data is input. Therefore, this solution realizes a paradigm shift from passively matching known attacks to actively modeling normal behavior and sensing hidden disturbances, which significantly improves the ability to detect advanced covert attacks, reduces the risk of false negatives, and provides reliable support for UAVs to make autonomous security decisions in complex network environments.

[0063] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A communication security protection method for unmanned aerial vehicles (UAVs), characterized in that, include: Acquire historical communication telemetry data of the drone to be protected; Based on the timestamp information extracted from the historical communication telemetry data, the historical communication telemetry data is cyclically encoded to obtain periodic communication data; The feature information extracted from the periodic communication data is analyzed, and the analysis results are used to filter the feature information to obtain normal communication features. The feature information is processed using a feature reconstruction mechanism built based on the normal communication features to determine the perturbation amount. The perturbation amount is a quantitative index characterizing the degree of deviation between the feature information and the reconstructed feature vector. The feature reconstruction mechanism is constructed by an autoencoder. Based on the disturbance amount, the historical communication telemetry data is processed to construct an intrusion sample dataset; The real-time status data of the drone to be protected is input into a stacked intrusion detection model trained based on the intrusion sample dataset for processing to obtain the intrusion detection result of the drone to be protected. The stacked intrusion detection model consists of a base model layer and a meta-classifier layer. The training method of the stacked intrusion detection model includes: dividing the constructed intrusion sample dataset into a training set and a validation set; training each base model in the base model layer with the training set; using the output results of each base model as the input of the meta-classifier layer; and training the meta-classifier layer with the training set. The input of the stacked intrusion detection model is the feature vector of the real-time status data, and the output is the intrusion detection result. The intrusion detection result includes normal communication, signal interference intrusion, instruction tampering intrusion, and data forgery intrusion. The system controls the drone to be protected to execute a security protection strategy generated by the ground control station based on the intrusion detection results. The security protection strategy includes: generating a strategy to maintain the current state for normal communication results; generating a communication link switching strategy for signal interference intrusions; generating a command verification and rollback strategy for command tampering intrusions; and generating a data verification and discard strategy for data forgery intrusions. The process of processing the historical communication telemetry data based on the disturbance amount to construct the intrusion sample dataset includes: The historical communication telemetry data is divided into black-box communication sample datasets and white-box communication sample datasets. The black-box communication sample dataset refers to a set of communication samples whose internal structure and parameters are not disclosed, while the white-box communication sample dataset refers to a set of communication samples whose internal structure, parameters, and status labels are disclosed. Perform a binary search on the black-box communication sample dataset to determine the perturbation strength parameters; The black-box communication sample dataset is processed based on the disturbance intensity parameter and the disturbance amount to determine the first intrusion sample dataset; Based on the disturbance intensity parameter and the disturbance amount, the white-box sample data corresponding to the second feature information in the white-box communication sample dataset is processed to determine the second intrusion sample dataset. The second feature information is obtained by removing the normal communication feature from the feature information. The second feature information is an intrusion-related feature. The white-box sample data corresponding to the second feature information is the white-box sample data in the white-box sample dataset that has the intrusion-related feature. The intrusion sample dataset is constructed based on the first intrusion sample dataset and the second intrusion sample dataset; The step of processing the feature information using a feature reconstruction mechanism based on the normal communication features to determine the perturbation amount includes: Based on the aforementioned normal communication characteristics, a feature reconstruction mechanism is constructed; The feature information is reconstructed using the aforementioned feature reconstruction mechanism to obtain a reconstructed feature vector; The reconstructed feature vector and the feature information are subjected to dimension-wise difference processing to obtain a reconstructed difference sequence; The overall deviation value of the reconstructed difference sequence is calculated using the L2 norm, and the overall deviation value is weighted and summed by combining the weight coefficients of each dimension feature to determine the perturbation amount.

2. The communication security protection method for unmanned aerial vehicles as described in claim 1, characterized in that, The step of performing a binary search on the black-box communication sample dataset to determine the perturbation strength parameters includes: The black-box communication sample dataset is trained based on a pre-built basic detection model to obtain the output result corresponding to each sample data. For each sample data in the black box communication sample dataset, the binary search algorithm is used to optimize the control disturbance intensity coefficient within a preset disturbance interval until the corresponding output result is reversed, thereby obtaining each minimum control disturbance intensity coefficient value, and then obtaining the minimum control disturbance intensity coefficient set. The output result reversal means that a sample that was originally judged as normal becomes abnormal or a sample that was originally abnormal becomes normal. The disturbance intensity parameter is obtained based on the minimum control disturbance intensity coefficient set, including: first removing outliers (coefficients that deviate from the mean by more than 3 times the standard deviation) from the minimum control disturbance intensity coefficients, and then taking the mean of the remaining coefficients and correcting it with the standard deviation to obtain the final disturbance intensity parameter.

3. The communication security protection method for unmanned aerial vehicles as described in claim 1, characterized in that, The process of cyclically encoding the historical communication telemetry data based on the timestamp information extracted from the historical communication telemetry data to obtain periodic communication data includes: Based on the timestamp information, an absolute time series representing the sampling time is obtained; The absolute time series is processed by trigonometric function mapping using a cyclic encoding function to obtain sine-cosine feature pairs. The sine-cosine feature pairs are concatenated with the non-time fields in the historical communication telemetry data in a time-aligned manner to obtain the periodic communication data.

4. A communication security protection system for unmanned aerial vehicles (UAVs), characterized in that, A communication security protection method for a drone as described in any one of claims 1 to 3 includes: The acquisition module is used to acquire historical communication telemetry data of the UAV to be protected; The loop module is used to perform cyclic encoding processing on the historical communication telemetry data based on the timestamp information extracted from the historical communication telemetry data to obtain periodic communication data. The filtering module is used to analyze the feature information extracted from the periodic communication data, and to filter the feature information based on the analysis results to obtain normal communication features. The disturbance module is used to process the feature information using a feature reconstruction mechanism built based on the normal communication features to determine the disturbance amount; The construction module is used to process the historical communication telemetry data based on the disturbance amount to construct an intrusion sample dataset; The processing module is used to input the real-time status data of the drone to be protected into the stacked intrusion detection model trained based on the intrusion sample dataset for processing, so as to obtain the intrusion detection result of the drone to be protected. The control module is used to control the drone to be protected to execute the security protection strategy generated based on the intrusion detection results.

5. The communication security protection system for unmanned aerial vehicles as described in claim 4, characterized in that, The building module includes: A partitioning unit is used to divide the historical communication telemetry data into a black-box communication sample dataset and a white-box communication sample dataset. The search unit is used to perform a binary search process on the black-box communication sample dataset to determine the perturbation strength parameters. The first unit is used to process the black box communication sample dataset based on the disturbance intensity parameter and the disturbance amount to determine the first intrusion sample dataset; The second unit is used to process the white-box communication sample dataset and the corresponding white-box sample data in the second feature information after removing the normal communication features based on the disturbance intensity parameter and the disturbance amount, so as to determine the second intrusion sample dataset. An intrusion unit is used to construct the intrusion sample dataset based on the first intrusion sample dataset and the second intrusion sample dataset.

6. The communication security protection system for unmanned aerial vehicles as described in claim 5, characterized in that, The step of performing a binary search on the black-box communication sample dataset to determine the perturbation strength parameters includes: The black-box communication sample dataset is trained based on a pre-built basic detection model to obtain the output result corresponding to each sample data. For each sample data in the black box communication sample dataset, the binary search algorithm is used to optimize the control disturbance intensity coefficient within a preset disturbance interval until the corresponding output result is reversed, thereby obtaining each minimum control disturbance intensity coefficient value, and thus obtaining the minimum control disturbance intensity coefficient set. The disturbance intensity parameter is obtained based on the minimum control disturbance intensity coefficient set.

7. The communication security protection system for unmanned aerial vehicles as described in claim 4, characterized in that, The disturbance module includes: A construction unit is used to construct a feature reconstruction mechanism based on the normal communication features; The reconstruction unit is used to reconstruct the feature information using the feature reconstruction mechanism to obtain a reconstructed feature vector; The determining unit is used to perform difference processing on the reconstructed feature vector and the feature information, and determine the perturbation amount based on the reconstruction difference.

8. The communication security protection system for unmanned aerial vehicles as described in claim 4, characterized in that, The loop module includes: An absolute time unit is used to obtain an absolute time series representing the sampling time based on the timestamp information. The mapping unit is used to perform trigonometric function mapping on the absolute time series using a cyclic coding function to obtain sine-cosine feature pairs. The splicing unit is used to splice the sine-cosine feature pairs with the non-time fields in the historical communication telemetry data according to time alignment to obtain the periodic communication data.

Images