An end-side detection method for unmanned aerial vehicle navigation attack

Abstract

This application provides an edge-side detection method for UAV navigation attacks, belonging to the field of UAV technology. The method includes acquiring time-series data of UAV navigation states to construct navigation state window samples; constructing a normal prototype discrimination branch and an attack prototype discrimination branch based on the low-dimensional embedding vectors of all navigation state window samples; inputting the low-dimensional embedding vector of the navigation state window sample to be tested into the normal prototype discrimination branch and the attack prototype discrimination branch respectively, to obtain a first attribution metric relative to the normal state of the UAV and a second attribution metric relative to the attack state of the UAV; if the first attribution metric or the second attribution metric is greater than a preset threshold, the navigation state window sample to be tested is determined to belong to an unknown navigation attack. This application achieves effective detection of unknown navigation attacks not seen during the training phase through a dual-prototype discrimination branch and meets the requirements for real-time edge deployment.

Description

Technical Field

[0001] This application belongs to the field of unmanned aerial vehicle (UAV) technology, and in particular relates to an end-side detection method for UAV navigation attacks. Background Technology

[0002] The rapid popularization of drone technology has led to its widespread application in various fields such as logistics delivery, emergency rescue, smart agriculture, and power line inspection, becoming a crucial support for the low-altitude economy. Drones heavily rely on Global Navigation Satellite System (GNSS) signals for positioning and navigation during operation. However, civilian GNSS signals reach the ground with a power of only -160dBm, resulting in extremely weak signal strength. Furthermore, they are transmitted in plaintext and lack encryption mechanisms and authentication capabilities. Attackers can suppress satellite signals with only milliwatt-level transmission power, preventing drones from locking onto satellites. They can also inject fake satellite signals to generate false latitude, longitude, altitude, and speed navigation information, causing the drone's flight control system to misinterpret the forged signals as valid, thus leading to deviations from the intended flight path. These navigation signal suppression and deception attacks are characterized by low implementation costs, low technical barriers, and high defense difficulties, becoming prominent problems hindering the healthy development of the low-altitude economy and threatening the safe operation of drones.

[0003] Existing detection methods for drone navigation attacks typically rely on machine learning or deep learning to model and identify navigation state data. The process includes: First, collecting navigation and telemetry data during drone flight, including fields such as position, velocity, attitude, and satellite status, and labeling the collected flight states (e.g., normal flight, navigation spoofing, navigation interference, etc.); second, preprocessing the raw data, such as parsing, field alignment, and standardization, and extracting features reflecting the navigation state; then, training a classification or anomaly detection model using the labeled training data, enabling the model to learn the feature boundaries of different navigation states; finally, during deployment, real-time collected navigation data is input into the trained model, which outputs a judgment result on the current navigation state, thereby achieving the identification and alerting of navigation interference, spoofing, and other attack behaviors.

[0004] However, the aforementioned existing technologies largely rely on the closed-set assumption in their model training steps. This means that the labeled data used in the training phase is assumed to cover all possible attack types during deployment, and the model only learns the feature boundaries of normal flight and known attack categories. Because they do not model unknown navigation attacks outside the training categories, when the model encounters novel interference or deception patterns not present in the training data during deployment, it will forcibly classify them into a known category, leading to missed detections and false positives, making it difficult to detect and alert on unknown navigation attacks in a timely manner. Furthermore, existing methods often rely on a large number of features or complex network structures in the feature extraction and model building steps, resulting in a large number of parameters and computational costs. This makes it difficult to directly deploy the trained models on UAV accompanying computing modules or low-power edge devices, failing to meet the requirements of real-time edge detection. Summary of the Invention

[0005] This application provides an edge-side detection method for UAV navigation attacks, which can solve the technical problems of existing UAV navigation attack detection methods that cannot identify unknown attack types not seen during the training phase due to their reliance on the closed set assumption, and whose large model parameter count makes real-time deployment on the edge difficult. This application provides an edge detection method for drone navigation attacks, including: Acquire the navigation state time-series data of the UAV to construct a navigation state window sample; Extract the low-dimensional embedding vectors of the navigation state window samples; Construct normal prototype discrimination branch and attack prototype discrimination branch based on the low-dimensional embedding vectors of all navigation state window samples; The low-dimensional embedding vector of the navigation state window sample to be tested is input into the normal prototype discrimination branch and the attack prototype discrimination branch respectively to obtain the first attribution metric of the navigation state window sample to be tested relative to the normal state of the UAV and the second attribution metric relative to the attack state of the UAV. If either the first attribution metric or the second attribution metric is greater than a preset threshold, the sample navigation state window to be tested is determined to be an unknown navigation attack.

[0006] Optionally, the navigation state time-series data of the UAV is acquired to construct a navigation state window sample, including: Obtain navigation status logs from the UAV flight control side and telemetry side, and parse and standardize them into multi-dimensional navigation status feature sequences; The multidimensional navigation state feature sequence is divided into time blocks to construct multiple navigation state window samples.

[0007] Optionally, navigation status logs from the UAV flight control side and telemetry side are acquired, and parsed and standardized into a multi-dimensional navigation status feature sequence, including: Multiple navigation state features during UAV flight are extracted from the navigation state logs of the flight control side and the telemetry side respectively to construct a multi-dimensional navigation state feature. Using a unified sampling time grid as a benchmark, the multi-dimensional navigation state features extracted from the flight control side and the telemetry side are aligned according to timestamps and merged into the same time series; The mean and standard deviation of each dimension of navigation state features under the same time series are statistically analyzed, and then standardized dimension by dimension using the z-score method to obtain a standardized multidimensional navigation state feature sequence.

[0008] Optionally, multiple navigation state features during UAV flight are extracted from the navigation state logs of the flight control side and the telemetry side, respectively, to construct multi-dimensional navigation state features, including: The total speed, vertical speed, ground speed, three-axis acceleration components, three-axis angular velocity components, and number of effective satellites during UAV flight are extracted from the navigation status logs of the flight control side and the telemetry side, respectively. Calculate the Euclidean magnitudes of the triaxial components of acceleration and the triaxial components of angular velocity respectively to obtain the acceleration magnitude and the angular velocity magnitude. The total speed, vertical velocity, ground speed, acceleration modulus, angular velocity modulus, and effective number of satellites are used as multi-dimensional navigation state features.

[0009] Optionally, the multidimensional navigation state feature sequence is divided into time blocks to construct multiple navigation state window samples, including: The standardized multidimensional navigation state feature sequence and the label sequence composed of the navigation state label of each sampling point are sequentially traversed along the time axis. If the navigation status labels or flight identifications of two adjacent sampling points are different, the two adjacent sampling points will be disconnected in order to divide the standardized multidimensional navigation status feature sequence into multiple time blocks with consistent navigation status and continuous time. Each time block is divided into training interval, validation interval, and test interval along the time axis; an interval band is set between adjacent intervals; the length g of the interval band satisfies g≥L-1; L is the sliding window length; Within each interval, starting from the beginning of the interval, windows of length L are sequentially extracted along the time axis with a sliding step size d. The window extraction stops when the right end of the window exceeds the end of the current interval. The navigation status label of the sampling point that appears most frequently in each window is used as the navigation status label of the corresponding window, resulting in multiple navigation status window samples.

[0010] Optionally, a low-dimensional embedding vector is extracted from the navigation state window sample, including: The navigation state window samples are rearranged from an initial dimension of batch size × window length × number of features into a temporal input tensor of batch size × number of features × window length; The rearranged temporal input tensor is sequentially fed into a cascaded three-layer one-dimensional convolution module to output multi-scale short-term variation features layer by layer; wherein, each one-dimensional convolution module sequentially performs one-dimensional convolution, batch normalization, nonlinear activation and max pooling operations. The short-term variation features output by the last layer's one-dimensional convolutional module are sequentially subjected to global average pooling, fully connected mapping, and embedding normalization to obtain a low-dimensional embedding vector with a fixed dimension.

[0011] Optionally, extracting the low-dimensional embedding vector of the navigation state window sample also includes: The low-dimensional embedding vector is input into the linear classification head and mapped to a K-dimensional original score vector through a linear transformation; where K is the total number of navigation state categories; each dimension of the original score vector represents the probability that the current navigation state window sample belongs to the corresponding known navigation state category; Based on the original score vector and the label of each navigation state window sample, the cross-entropy loss is calculated. With the cross-entropy loss as the optimization objective, the loss gradient is backpropagated layer by layer through the backpropagation algorithm. All trainable parameters of the three-layer one-dimensional convolutional module and the linear classification head are iteratively updated until the probability of correctly identifying the navigation state category in each validation interval reaches a preset threshold. Training stops when this happens and the current parameters of the three-layer one-dimensional convolutional module and the linear classification head are saved.

[0012] Optionally, a normal prototype discrimination branch and an attack prototype discrimination branch are constructed based on the low-dimensional embedding vectors of all navigation state window samples, including: The current three-layer one-dimensional convolutional module is used as a temporal embedding network to extract low-dimensional embedding vectors for all navigation state window samples. Each low-dimensional embedding vector is divided into the normal embedding set and the attack embedding set according to the navigation state label to which the corresponding navigation state window sample belongs; Calculate the mean vectors of the normal embedding set and the attack embedding set respectively, and use them as the normal prototype and the attack prototype. Calculate the distance from each embedding vector in the normal embedding set to the normal prototype, and the distance from each embedding vector in the attack embedding set to the attack prototype, and use them as the first distance set and the second distance set. Preset quantiles of the first distance set and the second distance set are selected respectively as the normal boundary radius and the attack boundary radius; wherein, the normal prototype and the normal boundary radius constitute the normal prototype discrimination branch; the attack prototype and the attack boundary radius constitute the attack prototype discrimination branch.

[0013] Optionally, the low-dimensional embedding vector of the navigation state window sample to be tested is input into the normal prototype discrimination branch and the attack prototype discrimination branch, respectively, to obtain the first attribution metric of the navigation state window sample to be tested relative to the normal state of the UAV and the second attribution metric relative to the attack state of the UAV, including: The ratio of the distance between the low-dimensional embedding vector of the navigation state window sample under test and the normal prototype to the normal boundary radius is used as the first attribution metric. The ratio of the distance from the low-dimensional embedding vector of the navigation state window sample under test to the attack prototype to the radius of the attack boundary is used as the second attribution metric.

[0014] Optionally, if either the first attribution metric or the second attribution metric is greater than a preset threshold, the sample of the navigation state window to be tested is determined to belong to an unknown navigation attack, including: The minimum value between the first attribution metric and the second attribution metric is used as the unknown attack risk metric. If the unknown attack risk metric exceeds a preset threshold, the sample navigation state window to be tested is determined to be an unknown navigation attack.

[0015] Beneficial effects: By constructing a dual-branch discrimination architecture of normal prototype and attack prototype, and introducing an unknown determination mechanism based on attribution metric, this application effectively solves the technical problem that existing models cannot identify unknown navigation attack types not seen during the training phase under the closed set assumption. Specifically, this application no longer forcibly classifies the samples of the navigation state window to be tested into a certain known category. Instead, it maps their low-dimensional embedding vectors to two prototype spaces, normal state and attack state, respectively. By calculating the attribution metric and comparing it with a preset threshold, any sample deviating from the known distribution boundary can be identified as an unknown attack, thereby significantly improving the detection sensitivity and generalization ability for new interference or deception patterns. At the same time, this application performs discrimination based on low-dimensional embedding vectors, resulting in a simple network structure and a small number of parameters. This avoids the computational redundancy caused by complex feature engineering and deep networks, enabling the entire detection process to meet the stringent real-time requirements of UAV-side accompanying computing modules or low-power edge devices. Furthermore, the dual-prototype discrimination branch architecture design gives the model good scalability. When a new attack type is subsequently discovered, only the prototype and boundary radius need to be incrementally updated, without retraining the entire network, facilitating continuous iterative upgrades in actual deployments.

[0016] Based on this, this application provides a process for acquiring navigation state time-series data and constructing window samples. By extracting multi-dimensional navigation state features from the flight control side and telemetry side respectively and performing timestamp alignment and z-score standardization, the integrity and dimensional consistency of the data source are guaranteed. In particular, by dividing the time blocks based on navigation state labels and sortie identifiers, and setting an interval band of no less than L-1 between the training interval, validation interval, and test interval, the contamination of window samples by sampling points across states or intervals is effectively avoided, ensuring the strict temporal independence of training, validation, and test data, and providing a reliable data foundation for model training and generalization performance evaluation.

[0017] This application provides a cascaded structure and training strategy for a one-dimensional convolutional module. It uses three layers to extract multi-scale short-term changing features layer by layer for the convolutional module, and obtains a fixed-dimensional embedding vector through global average pooling and fully connected mapping. It uses cross-entropy loss as the optimization objective and stops training early when the preset accuracy is reached in the validation interval. This effectively prevents overfitting while ensuring the discriminative power of the features, and further improves the robustness of the model under small sample attack data conditions.

[0018] This application provides a method for constructing normal prototypes and attack prototypes, as well as a specific method for calculating the attribution metric. By calculating the distance from the embedded vectors in the two sets to the corresponding prototypes and selecting a preset quantile as the boundary radius, the ratio of the distance from the test sample to the prototype to the boundary radius is used as the attribution metric. The minimum value of the two attribution metrics is used as the unknown attack risk metric. This judgment mechanism is intuitive and computationally efficient. It can quickly output detection results in the inference stage without additional training, which is fully in line with the application scenario of real-time detection on the edge. Attached Figure Description

[0019] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating an end-side detection method for drone navigation attacks provided in this application embodiment; Figure 2 A schematic diagram of the construction process of the navigation status window sample provided in the embodiments of this application; Figure 3 This is a schematic diagram illustrating the boundary discrimination for dual-prototype unknown navigation attack detection provided in an embodiment of this application. Figure 4 This is a schematic diagram of the structure of an end-side detection system for unmanned aerial vehicle (UAV) navigation attacks, provided as an embodiment of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0022] Example 1 like Figure 1 As shown, this embodiment provides an edge detection method for drone navigation attacks, including: Step 101: Obtain the navigation status time series data of the UAV to construct a navigation status window sample.

[0023] The experimental platform in this embodiment includes a drone, a Pixhawk 4 flight controller, PX4 flight controller firmware, a global satellite navigation receiver, a telemetry radio, a remote control receiver, a MAVLink communication link, edge computing devices, and a controlled experimental environment. Logs are collected through simulation experiments and real flight experiments. Satellite navigation spoofing generates fake navigation signals through global satellite navigation signal generation software, software radio equipment, and antennas, and transmits them to the navigation receiver, causing the receiver to lock onto the fake signals and output a false positioning. Satellite navigation interference generates noise signals through radio tools and acts on the navigation receiver, suppressing real satellite signals. During data collection, telemetry-side logs and flight controller-side logs are recorded simultaneously. Sampling points are marked as normal flight, satellite navigation spoofing, or satellite navigation interference according to the experimental stage and attack injection period, and the logs are exported as comma-separated value files. For example, the processed satellite navigation spoofing log contains 3622 records, of which 3124 are normal records and 498 are abnormal records; the satellite navigation interference log contains 6445 records, of which 4985 are normal records and 1460 are abnormal records.

[0024] Since the flight control side logs and telemetry side communication logs come from different sources and have different field names and sampling rates, they need to be unified into feature sequences under the same time reference. This step first parses the comma-separated value logs from both sides (flight control side and telemetry side) to extract the raw measurements related to the navigation status. Considering that quantities such as acceleration and angular velocity are usually recorded in the logs in the form of three-axis components, in order to eliminate the differences caused by the aircraft attitude orientation and retain only the information on the intensity of motion, this step calculates the Euclidean modulus of the acceleration three-axis component and the angular velocity three-axis component respectively, to obtain the acceleration modulus and angular velocity modulus. Together with the total rate (velocity), vertical velocity, ground speed, and number of effective satellites, they constitute the six-dimensional navigation status features of total rate, vertical velocity, ground speed, acceleration modulus, angular velocity modulus, and number of effective satellites.

[0025] Since the sampling times on both sides are not strictly aligned (i.e., their respective sampling frequencies (number of samples recorded per second or per time unit) are different), this step uses a unified sampling time grid as a benchmark. The six-dimensional navigation state features extracted from both sides are aligned according to the nearest timestamp, and missing parts are filled in using previous value preservation or linear interpolation, merging them into a single multi-dimensional time series arranged by time. Subsequently, explicit timestamps, experiment numbers, and manual labels, as well as fields related to specific collection batches that are not available during the deployment phase, are deleted, retaining only the aforementioned six-dimensional navigation state features.

[0026] During standardization, the mean and standard deviation of each feature dimension are calculated only in the training interval. The training interval, validation interval, test interval, and board-end inference input are standardized dimension by dimension using the z-score method, which is to subtract the mean of the dimension and then divide by the standard deviation of the dimension. This ensures that the dimensions of each navigation state feature are consistent, the mean is close to zero, and the variance is close to one, thus avoiding the improper weighting of features with significantly different dimensions, such as the number of effective satellites and velocity, in subsequent networks or models.

[0027] Drone navigation status logs are typically composed of multiple flight sorties, and within the same sortie, different navigation states such as normal flight and attack injection will occur sequentially. Therefore, it is not a single continuous flight, but rather a concatenation of several subsequences that are continuous in both time and navigation state. This step first traverses the six-dimensional navigation state features and their point-by-point (navigation state) label sequence along the time axis, comparing the navigation state labels of adjacent (two) sampling points with the sortie identifier. Once one of them changes, the sequence is broken at that point (between two adjacent sampling points), thus dividing the entire sequence into several continuous time blocks (i.e., strict time blocks) with consistent states and continuous time, preventing sampling points across sorties or navigation states from being included in the same window.

[0028] Subsequently, within each consecutive time block, training, validation, and test intervals are sequentially drawn according to chronological order rather than random sampling (e.g., in a 7:1.5:1.5 temporal ratio), ensuring that the three intervals (sets) do not overlap temporally. Considering that a sliding window of length L (e.g., L=15) may span L consecutive sampling points, if adjacent intervals are drawn close together, windows straddling the boundary will simultaneously capture sampling points belonging to different intervals, causing future information from the test interval to be prematurely included in the training. Therefore, this step leaves an additional g sampling points between every two adjacent intervals as an interval band, which does not participate in windowing for any interval, and the interval length g is set to g≥L-1, thus ensuring that no window of length L will simultaneously fall into two intervals.

[0029] Within each interval, starting from the interval's beginning, windows of length L are sequentially extracted along the time axis with a sliding step size d (e.g., d=1). Window extraction stops when the right edge of a window extends beyond the end of the interval. For each window, the mode of the labels for each sampling point (navigation status) within it is taken, and the label corresponding to the mode is used as the window's label, indicating normal flight, satellite navigation spoofing, or satellite navigation interference. The expression for the navigation status window sample and its label is: .

[0030] Among them, W k This represents the set of samples for the k-th navigation state window, which consists of L consecutive sampling points. This represents the six-dimensional navigation state feature vector corresponding to the starting sampling point of the k-th window; t represents the six-dimensional navigation state feature vector corresponding to the sampling point at the end of the k-th window; k This represents the index of the starting sampling point of the k-th window; numerically, t k =1+(k-1)d;y k This represents the set of labels consisting of the navigation status labels of each sampling point within the k-th window; The navigation status label represents the starting sampling point of the k-th window; Indicates the navigation status label of the sampling point at the end of the k-th window; The label represents the navigation status label corresponding to the k-th window; mode(·) represents the label that appears most frequently in the window; after the above processing, a total of 10037 navigation status window samples were obtained, including 8095 normal samples, 490 satellite navigation spoofing samples, and 1452 satellite navigation interference samples.

[0031] For example, this step includes: Obtain navigation status logs from the UAV flight control and telemetry sides, and parse and standardize them into multi-dimensional navigation status feature sequences, including: Multiple navigation state features during UAV flight are extracted from the navigation state logs of the flight control side and the telemetry side respectively to construct a multi-dimensional navigation state feature, including: The total speed, vertical speed, ground speed, three-axis acceleration components, three-axis angular velocity components, and number of effective satellites during UAV flight are extracted from the navigation status logs of the flight control side and the telemetry side, respectively. Calculate the Euclidean magnitudes of the triaxial components of acceleration and the triaxial components of angular velocity respectively to obtain the acceleration magnitude and the angular velocity magnitude. The total speed, vertical velocity, ground speed, acceleration modulus, angular velocity modulus, and effective number of satellites are used as multi-dimensional navigation state features.

[0032] Using a unified sampling time grid as a benchmark, the multi-dimensional navigation state features extracted from the flight control side and the telemetry side are aligned according to timestamps and merged into the same time series.

[0033] The mean and standard deviation of each dimension of navigation state features under the same time series are statistically analyzed, and then standardized dimension by dimension using the z-score method to obtain a standardized multidimensional navigation state feature sequence.

[0034] The multidimensional navigation state feature sequence is divided into time blocks to construct multiple navigation state window samples, including: The standardized multidimensional navigation state feature sequence and the label sequence consisting of the navigation state label of each sampling point are traversed sequentially along the time axis.

[0035] If the navigation status labels or flight identifiers of two adjacent sampling points are different, the two adjacent sampling points are disconnected in order to divide the standardized multidimensional navigation status feature sequence into multiple time blocks with consistent navigation status and continuous time.

[0036] Each time block is divided into training interval, validation interval and test interval along the time axis; an interval band is set between adjacent intervals; the length g of the interval band satisfies g≥L-1; L is the length of the sliding window.

[0037] Within each interval, starting from the interval's beginning, windows of length L are sequentially extracted along the time axis with a sliding step size d. Window extraction stops when the right edge of a window exceeds the end of the current interval. The navigation state label of the sampling point that appears most frequently within each window is used as the corresponding window's navigation state label, resulting in multiple navigation state window samples (as a window sample set). The above process is as follows: Figure 2 As shown.

[0038] Step 102: Extract the low-dimensional embedding vector of the navigation state window sample.

[0039] For each navigation state window sample obtained in step 101, it is rearranged from the initial dimension of batch number × window length × number of features into a temporal input tensor of batch number × number of features × window length and then input into the temporal embedding network. The temporal embedding network includes three stacked one-dimensional convolutional modules. Each one-dimensional convolutional module performs one-dimensional convolution, batch normalization, non-linear activation and pooling in sequence. Specifically, the navigation state window sample is passed through a one-dimensional convolutional module with 16 channels, a kernel size of 3 and padding of 1, and then batch normalization, non-linear activation and max pooling in sequence. Two more identical one-dimensional convolutional modules are stacked, with the number of channels being 32 and 64 respectively. Each layer uses a kernel size of 3 and padding of 1, and then performs batch normalization, non-linear activation and max pooling in sequence.

[0040] The short-term variation characteristics of multi-scale output layer by layer are as follows: .

[0041] Among them, H i The short-term variation features output by the i-th layer one-dimensional convolutional module; i=1,2,3; Pool(·) represents the max pooling operation; BN(·) represents a non-linear activation operation; BN(·) represents a batch normalization operation; Conv1D i (·) represents the one-dimensional convolution operation of the i-th layer one-dimensional convolution module.

[0042] The short-term variation features output by the last one-dimensional convolutional module are sequentially subjected to global average pooling (to obtain a global temporal representation), fully connected mapping, and embedding normalization to obtain a low-dimensional embedding vector z with fixed dimensions (32 dimensions): z=ρ(FC(GAP(H3).

[0043] Where H3 represents the short-term variation features output by the third-layer one-dimensional convolutional module; ρ(·) represents embedding normalization; FC(·) represents fully connected mapping; and GAP(·) represents global average pooling.

[0044] For example, this step also includes: The low-dimensional embedding vector z is input into the linear classification head and mapped to a K-dimensional original score vector through a linear transformation; where K is the total number of navigation state categories (K=3); each dimension of the original score vector represents the probability that the current navigation state window sample belongs to the corresponding known navigation state category.

[0045] Each dimension's value refers to the unnormalized score (logit) output by the linear classification head, corresponding to a known navigation state category. The magnitude of each dimension's value represents the confidence (or probability tendency) that the current input navigation state window sample belongs to the corresponding category. The higher the score, the greater the likelihood that the sample belongs to that category (these scores have not yet been normalized to probabilities by softmax, but the relative magnitude of the scores determines the classification result).

[0046] Based on the original score vector and the label (one-hot encoded) of each navigation state window sample, the cross-entropy loss is calculated. With the cross-entropy loss as the optimization objective, the loss gradient is backpropagated layer by layer through the backpropagation algorithm to iteratively update all trainable parameters of the three-layer one-dimensional convolutional module and the linear classifier head. Training stops when the probability of correctly identifying the navigation state category in each validation interval reaches a preset threshold. The current parameters of the three-layer one-dimensional convolutional module and the linear classifier head are saved, along with the standardized mean and standard deviation of the training interval.

[0047] Step 103: Construct normal prototype discrimination branch and attack prototype discrimination branch based on the low-dimensional embedding vectors of all navigation state window samples.

[0048] In this step, the current three-layer one-dimensional convolutional module is used as a temporal embedding network to extract low-dimensional embedding vectors for all navigation state window samples. Each low-dimensional embedding vector is then assigned to the normal embedding set Z according to the navigation state label to which the corresponding navigation state window sample belongs. N and attack embedding set Z A That is, according to the labels, the embedding vectors of normal flight samples are assigned to the normal embedding set, and the embedding vectors of known satellite navigation spoofing or known satellite navigation interference samples are assigned to the known attack embedding set; the mean of the embedding vectors of the two sets is calculated to obtain the normal prototype c. N and attack prototype c A : .

[0049] Calculate the distance from each embedding vector in the normal embedding set to the normal prototype, and the distance from each embedding vector in the attack embedding set to the attack prototype, and use them as the first distance set and the second distance set.

[0050] Preset quantiles (e.g., 0.95 quantiles) of the first and second distance sets are selected as the normal boundary radius and the attack boundary radius, respectively; where, for example, Figure 3 As shown, the normal prototype and normal boundary radius r N The normal prototype discrimination branch is formed, and the attack prototype and attack boundary radius r are also considered. A This constitutes the attack prototype discrimination branch.

[0051] Step 104: Input the low-dimensional embedding vector of the navigation state window sample to be tested into the normal prototype discrimination branch and the attack prototype discrimination branch respectively to obtain the first attribution metric of the navigation state window sample to be tested relative to the normal state of the UAV and the second attribution metric relative to the attack state of the UAV.

[0052] For example, the distance d from the low-dimensional embedding vector of the navigation state window sample to the normal prototype is... N With normal boundary radius r N The ratio of is used as the first attribution metric.

[0053] The distance d from the low-dimensional embedding vector of the navigation state window sample to the attack prototype. A With the attack boundary radius r A The ratio of is used as the second attribution metric.

[0054] Step 105: If the first attribution metric or the second attribution metric is greater than a preset threshold, determine that the navigation state window sample to be tested belongs to an unknown navigation attack.

[0055] For example, the minimum of the first attribution metric and the second attribution metric is used as the unknown attack risk metric.

[0056] If the unknown attack risk metric is greater than the preset threshold, the sample of the navigation state window to be tested is determined to be an unknown navigation attack and rejected. Otherwise, the maximum response category given by the linear classification head is output, which is one of normal flight, satellite navigation spoofing or satellite navigation interference, to obtain the final detection result.

[0057] The trained lightweight three-class classification model is exported as an Open Neural Network Exchange Format file, with a fixed input shape of 1×15×6, and loaded into the runtime environment of the embedded board. The board maintains a first-in-first-out feature window of length L. Upon receiving a new six-dimensional navigation state sampling point, the oldest sampling point is removed and added. When the window is full, the features within the window are first standardized dimension-by-dimensionally using the saved mean and standard deviation to construct a 1×6×15 input tensor. Then, forward inference of the model is performed to obtain the embedding vector and class output. Subsequently, the detection result is output based on the ratio of the distance to the boundary radius of the prototype and a preset threshold. The above steps are repeatedly executed on the real-time input navigation state sampling points to achieve online real-time detection of the UAV navigation state.

[0058] In summary, the edge detection method for UAV navigation attacks provided in this embodiment organically combines a dual-prototype discriminative branch architecture with a lightweight temporal embedding network. On the one hand, it utilizes low-dimensional embedding vectors to achieve efficient feature representation, and the three-layer one-dimensional convolutional module maintains a very small number of parameters, meeting the requirements for real-time deployment on the edge. On the other hand, through dual boundary constraints of normal prototypes and attack prototypes, it effectively rejects unknown attack samples that deviate from the known distribution boundary, solving the defect of traditional methods that cannot identify attack types not seen during the training phase due to the closed set assumption. This significantly improves the generalization and security of the model in open and dynamic environments.

[0059] Example 2 Based on the same inventive concept as Embodiment 1, this embodiment provides an end-side detection system for UAV navigation attacks. Since the principle of this system in solving the problem is similar to the end-side detection method for UAV navigation attacks described in Embodiment 1, the implementation of this system can refer to the implementation of the end-side detection method for UAV navigation attacks.

[0060] like Figure 4 As shown, this embodiment provides an edge detection system for drone navigation attacks, including: The first construction module 10 is used to acquire the navigation state timing data of the UAV in order to construct a navigation state window sample.

[0061] Vector extraction module 20 is used to extract low-dimensional embedding vectors of navigation state window samples.

[0062] The second building module 30 is used to construct normal prototype discrimination branch and attack prototype discrimination branch based on the low-dimensional embedding vectors of all navigation state window samples.

[0063] The attribution metric determination module 40 is used to input the low-dimensional embedding vector of the navigation state window sample to be tested into the normal prototype discrimination branch and the attack prototype discrimination branch respectively, so as to obtain the first attribution metric of the navigation state window sample to be tested relative to the normal state of the UAV and the second attribution metric relative to the attack state of the UAV.

[0064] The attack detection module 50 is used to determine that the navigation state window sample under test belongs to an unknown navigation attack when the first attribution metric or the second attribution metric is greater than a preset threshold.

[0065] For more detailed information on the working process of each of the above modules, please refer to the relevant content disclosed in Example 1, which will not be repeated here.

[0066] Example 3 This embodiment provides a computer device, including a processor and a memory; wherein, when the processor executes a computer program stored in the memory, it implements the steps of the end-side detection method for UAV navigation attacks described in Embodiment 1.

[0067] For a more detailed explanation of the above method, please refer to the relevant content disclosed in Example 1, which will not be repeated here.

[0068] Example 4 This embodiment provides a computer-readable storage medium for storing a computer program; when the computer program is executed by a processor, it implements the steps of the end-side detection method for UAV navigation attacks described in Embodiment 1.

[0069] For a more detailed explanation of the above method, please refer to the relevant content disclosed in Example 1, which will not be repeated here.

[0070] Example 5 This embodiment provides a computer program product, including computer-executable instructions or a computer program. When the computer-executable instructions or the computer program are executed by a processor, they implement the steps of the end-side detection method for UAV navigation attacks described in Embodiment 1.

[0071] For a more detailed explanation of the above method, please refer to the relevant content disclosed in Example 1, which will not be repeated here.

[0072] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems, devices, storage media, and computer program products disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

[0073] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0074] In some embodiments, computer-executable instructions may take the form of programs, software, software modules, scripts, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as stand-alone programs or as modules, components, subroutines, or other units suitable for use in a computing environment.

[0075] As an example, computer-executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple co-located files (e.g., files that store one or more modules, subroutines, or code sections).

[0076] As an example, computer-executable instructions can be deployed to execute on a single electronic device, or on multiple electronic devices located at one location, or on multiple electronic devices distributed across multiple locations and interconnected via a communication network.

[0077] The present application has been described in detail above with reference to specific embodiments and exemplary examples; however, these descriptions should not be construed as limiting the present application. Those skilled in the art will understand that various equivalent substitutions, modifications, or improvements can be made to the technical solutions and implementation methods of the present application without departing from the spirit and scope of the present application, and all such modifications and improvements fall within the scope of the present application. The scope of protection of the present application is determined by the appended claims.

Claims

1. An edge detection method for unmanned aerial vehicle (UAV) navigation attacks, characterized in that, include: Acquire the navigation state time-series data of the UAV to construct a navigation state window sample; Extract the low-dimensional embedding vectors of the navigation state window samples; Based on the low-dimensional embedding vectors of all navigation state window samples, construct normal prototype discrimination branches and attack prototype discrimination branches, including: The current cascaded three-layer one-dimensional convolutional modules are used as a temporal embedding network to extract low-dimensional embedding vectors for all navigation state window samples. Each low-dimensional embedding vector is divided into the normal embedding set and the attack embedding set according to the navigation state label to which the corresponding navigation state window sample belongs; Calculate the mean vectors of the normal embedding set and the attack embedding set respectively, and use them as the normal prototype and the attack prototype. Calculate the distance from each embedding vector in the normal embedding set to the normal prototype, and the distance from each embedding vector in the attack embedding set to the attack prototype, and use them as the first distance set and the second distance set. Preset quantiles of the first distance set and the second distance set are selected respectively as the normal boundary radius and the attack boundary radius; wherein, the normal prototype and the normal boundary radius constitute the normal prototype discrimination branch; the attack prototype and the attack boundary radius constitute the attack prototype discrimination branch; The low-dimensional embedding vector of the navigation state window sample to be tested is input into the normal prototype discrimination branch and the attack prototype discrimination branch, respectively, to obtain the first attribution metric of the navigation state window sample relative to the normal state of the UAV and the second attribution metric relative to the attack state of the UAV, including: The ratio of the distance between the low-dimensional embedding vector of the navigation state window sample under test and the normal prototype to the normal boundary radius is used as the first attribution metric. The ratio of the distance from the low-dimensional embedding vector of the navigation state window sample to the attack prototype to the radius of the attack boundary is used as the second attribution metric. The minimum value between the first attribution metric and the second attribution metric is used as the unknown attack risk metric. If the unknown attack risk metric exceeds a preset threshold, the sample navigation state window to be tested is determined to be an unknown navigation attack.

2. The end-side detection method for UAV navigation attacks according to claim 1, characterized in that, Acquire the timing data of the UAV's navigation state to construct a navigation state window sample, including: Obtain navigation status logs from the UAV flight control side and telemetry side, and parse and standardize them into multi-dimensional navigation status feature sequences; The multidimensional navigation state feature sequence is divided into time blocks to construct multiple navigation state window samples.

3. The end-side detection method for UAV navigation attacks according to claim 2, characterized in that, Obtain navigation status logs from the UAV flight control and telemetry sides, and parse and standardize them into multi-dimensional navigation status feature sequences, including: Multiple navigation state features during UAV flight are extracted from the navigation state logs of the flight control side and the telemetry side respectively to construct a multi-dimensional navigation state feature. Using a unified sampling time grid as a benchmark, the multi-dimensional navigation state features extracted from the flight control side and the telemetry side are aligned according to timestamps and merged into the same time series; The mean and standard deviation of each dimension of navigation state features under the same time series are statistically analyzed, and then standardized dimension by dimension using the z-score method to obtain a standardized multidimensional navigation state feature sequence.

4. The end-side detection method for UAV navigation attacks according to claim 3, characterized in that, Multiple navigation state features during UAV flight are extracted from the navigation state logs of the flight control side and the telemetry side respectively to construct a multi-dimensional navigation state feature, including: The total speed, vertical speed, ground speed, three-axis acceleration components, three-axis angular velocity components, and number of effective satellites during UAV flight are extracted from the navigation status logs of the flight control side and the telemetry side, respectively. Calculate the Euclidean magnitudes of the triaxial components of acceleration and the triaxial components of angular velocity respectively to obtain the acceleration magnitude and the angular velocity magnitude. The total speed, vertical velocity, ground speed, acceleration modulus, angular velocity modulus, and effective number of satellites are used as multi-dimensional navigation state features.

5. The end-side detection method for UAV navigation attacks according to claim 2, characterized in that, The multidimensional navigation state feature sequence is divided into time blocks to construct multiple navigation state window samples, including: The standardized multidimensional navigation state feature sequence and the label sequence composed of the navigation state label of each sampling point are sequentially traversed along the time axis. If the navigation status labels or flight identifications of two adjacent sampling points are different, the two adjacent sampling points will be disconnected in order to divide the standardized multidimensional navigation status feature sequence into multiple time blocks with consistent navigation status and continuous time. Each time block is divided into training interval, validation interval, and test interval along the time axis; an interval band is set between adjacent intervals; the length g of the interval band satisfies g≥L-1; L is the sliding window length; Within each interval, starting from the beginning of the interval, windows of length L are sequentially extracted along the time axis with a sliding step size d. The window extraction stops when the right end of the window exceeds the end of the current interval. The navigation status label of the sampling point that appears most frequently in each window is used as the navigation status label of the corresponding window, resulting in multiple navigation status window samples.

6. The end-side detection method for UAV navigation attacks according to claim 5, characterized in that, Extract the low-dimensional embedding vectors of the navigation state window samples, including: The navigation state window samples are rearranged from an initial dimension of batch size × window length × number of features into a temporal input tensor of batch size × number of features × window length; The rearranged temporal input tensor is sequentially fed into a cascaded three-layer one-dimensional convolution module to output multi-scale short-term variation features layer by layer; wherein, each one-dimensional convolution module sequentially performs one-dimensional convolution, batch normalization, nonlinear activation and max pooling operations. The short-term variation features output by the last layer's one-dimensional convolutional module are sequentially subjected to global average pooling, fully connected mapping, and embedding normalization to obtain a low-dimensional embedding vector with a fixed dimension.

7. The end-side detection method for UAV navigation attacks according to claim 6, characterized in that, Extracting the low-dimensional embedding vector of the navigation state window samples also includes: The low-dimensional embedding vector is input into the linear classification head and mapped to a K-dimensional original score vector through a linear transformation; where K is the total number of navigation state categories; each dimension of the original score vector represents the probability that the current navigation state window sample belongs to the corresponding known navigation state category; Based on the original score vector and the label of each navigation state window sample, the cross-entropy loss is calculated. With the cross-entropy loss as the optimization objective, the loss gradient is backpropagated layer by layer through the backpropagation algorithm. All trainable parameters of the three-layer one-dimensional convolutional module and the linear classification head are iteratively updated until the probability of correctly identifying the navigation state category in each validation interval reaches a preset threshold. Training stops when this happens and the current parameters of the three-layer one-dimensional convolutional module and the linear classification head are saved.

Images