A GNSS spoofing detection method based on adaptive RAIM and latent space diffusion model
By constructing a pseudorange residual feature extractor and a latent space diffusion model, the problems of detection accuracy and stability of GNSS spoofing detection under diverse attacks are solved, and high-precision, stable and reliable spoofing detection of GNSS receivers is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DALIAN MARITIME UNIVERSITY
- Filing Date
- 2026-01-09
- Publication Date
- 2026-05-29
AI Technical Summary
Existing GNSS spoofing detection algorithms have low detection accuracy and stability when facing diverse covert spoofing attacks, especially in complex environments where their generalization ability is insufficient, making it difficult to maintain stable and reliable detection performance.
A pseudorange residual feature extractor is constructed, which extracts multi-scale trend features through a weighted adaptive network and weighted least squares RAIM, and is trained in conjunction with a latent space diffusion model. Deception detection is performed using a state evaluation function.
It improves the detection performance of GNSS receivers against diverse spoofing attacks and enhances detection stability and accuracy in unknown environments.
Smart Images
Figure CN122110159A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of satellite navigation and artificial intelligence, and more specifically, to a deep learning-based GNSS receiver signal processing method, particularly a method for GNSS spoofing detection in satellite navigation systems. Background Technology
[0002] Global Navigation Satellite System (GNSS) provides users with global positioning and navigation services. In recent years, with the widespread deployment of GNSS in various navigation systems, its reliability and security issues have become increasingly prominent. Currently, most navigation systems rely primarily on GNSS for their positioning and navigation capabilities. However, in extreme scenarios such as military conflicts and cyberattacks, GNSS signals are highly susceptible to spoofing attacks due to their inherent vulnerabilities, leading to a decrease in the reliability and security of the navigation system. Therefore, proposing a high-precision and highly reliable GNSS spoofing detection algorithm has significant theoretical and engineering value.
[0003] Existing GNSS spoofing detection algorithms can be broadly categorized into spoofing detection methods based on satellite message encryption, spoofing detection methods based on external navigation source-assisted comparison, and spoofing detection methods based on receiver signal characteristics. However, spoofing detection methods based on satellite message encryption involve modifying signal protocols, sometimes requiring hardware or software modifications to the receiver, making these methods difficult to implement. Spoofing detection methods based on external navigation source-assisted comparison often use autonomous navigation sources such as Inertial Navigation Systems (INS) to perform consistency spoofing detection with GNSS positioning results. However, due to the accumulation of errors in INS, its positioning results become unreliable after a certain period. Furthermore, in integrated navigation systems, when the spoofed GNSS corrects the INS, it accelerates INS divergence. Therefore, spoofing detection methods based on external navigation source-assisted comparison have certain shortcomings in long-term spoofing detection scenarios. Spoofing detection methods based on receiver signal characteristics use features from the GNSS receiver, such as power, carrier-to-noise ratio, correlation peaks, and derived pseudorange and Doppler shift, for spoofing detection. Receiver Autonomous Integrity Monitoring (RAIM) can eliminate anomalous signals by checking the spatial consistency of all available GNSS signals. However, RAIM is primarily suitable for situations where spoofing signals are weak and only a few satellites are being deceived. If spoofing signals dominate or are locked onto by the receiver, RAIM may incorrectly reject genuine observations and accept spoofed ones.
[0004] With the development of artificial intelligence technology, spoofing detection methods based on receiver signal features often use these features to train machine learning and deep learning models. However, most existing methods still employ supervised learning to train spoofing detection models. These models rely on labeled sample data, and their detection performance is limited by the prior class information introduced during the training phase. Therefore, these methods can typically only effectively learn for a limited number of known spoofing attack scenarios. In complex environments, spoofing attack types are diverse, and even the same spoofing attack can have different propagation characteristics affected by environmental parameters, resulting in different spoofing effects. Therefore, supervised learning-based spoofing detection algorithms often face problems such as insufficient generalization ability and unstable detection performance in complex environments.
[0005] Therefore, how to design a GNSS spoofing detection algorithm that enables the receiver to maintain stable and reliable spoofing detection performance when facing unknown and diverse spoofing attacks has become a key problem that urgently needs to be solved in the field of GNSS spoofing detection. Summary of the Invention
[0006] In view of the shortcomings of existing technologies, this invention provides a GNSS spoofing detection method based on adaptive RAIM and latent space diffusion model. This invention constructs a pseudorange residual feature extractor, adaptively learns pseudorange observation weights and extracts multi-scale trend features, and combines this with a latent space diffusion model trained only on unspoofed GNSS data. Through a state evaluation function, it achieves high-precision detection against diverse covert GNSS spoofing attacks.
[0007] The technical means employed in this invention are as follows: A GNSS spoofing detection method based on adaptive RAIM and latent space diffusion model includes the following steps: A pseudorange residual feature extractor is constructed, which includes a weighted adaptive network and a weighted least squares RAIM. The weighted adaptive network is used to generate a pseudorange observation weight matrix that is adapted to the constellation structure based on the spatial geometric relationship between the satellite and the receiver. The weighted least squares RAIM is used to calculate pseudorange residual features based on the pseudorange observation weight matrix. Unaffected GNSS pseudorange measurement data, satellite coordinates, and the receiver's initial position coordinates are obtained. Based on these data, a weighted adaptive network in the pseudorange residual feature extractor is trained. The pseudorange observation weights output by the trained network, along with the GNSS pseudorange measurements, satellite coordinates, and the receiver's initial position coordinates, are used to calculate RAIM pseudorange residual features. These features are then standardized, and multi-scale trend features are obtained for training the latent space diffusion model. A latent space diffusion model is constructed, comprising a latent space encoder, a forward diffusion process, and a noise prediction network. The latent space encoder maps high-dimensional multi-scale trend features to latent space features suitable for modeling the diffusion model, thereby improving the stability of the noise prediction network training process. The forward diffusion process progressively injects Gaussian noise into the latent space features to generate noisy latent space features. The noise prediction network performs noise prediction on the noisy latent space features. The latent space diffusion model is trained using the multi-scale trend features as the original input to learn the statistical distribution of the multi-scale trend features under normal GNSS conditions. The process involves acquiring GNSS pseudorange measurement data, satellite coordinates, and the receiver's initial position coordinates to train a weighted adaptive network in the pseudorange residual feature extractor. Based on the pseudorange observation weights output by the trained weighted adaptive network, and the GNSS pseudorange measurements, satellite coordinates, and receiver's initial position coordinates, RAIM pseudorange residual features are calculated. These pseudorange residual features are then standardized, and multi-scale trend features of the data to be detected are obtained. Noise prediction is performed on these multi-scale trend features based on the trained latent space diffusion model, yielding the corresponding noise prediction results. Finally, a noise consistency check is performed on the noise prediction results using a state evaluation function and an adaptive threshold, resulting in the GNSS spoofing detection result.
[0008] Furthermore, the weighted adaptive network includes a first graph convolutional neural network and a second graph convolutional neural network; the first graph convolutional neural network is used to process the input adjacency matrix and feature matrix sequentially by performing adjacency relationship normalization, feature aggregation, feature mapping and ReLU activation function processing, and outputting an inter-layer feature matrix; the adjacency matrix is an adjacency matrix of a fully connected undirected graph constructed based on the satellites in the current epoch, and the feature matrix is constructed based on the satellite elevation angle vector, the satellite-receiver three-dimensional line-of-sight vector and the unit constant bias vector; The second convolutional neural network is used to perform feature aggregation and weight mapping on the adjacency matrix and the inter-layer feature matrix after adjacency relationship normalization, and output the pseudorange observation weights of each satellite.
[0009] Furthermore, the fully connected undirected graph is constructed as follows: each satellite is treated as a graph node, and an edge is connected between any two satellites, with the weight of each edge set to an equal weight constant.
[0010] Furthermore, statistical processing is performed on the standardized pseudorange residual features to obtain multi-scale trend features, including: The mean, infinite norm, and 2-norm of the pseudorange residuals of individual satellites are statistically analyzed as the single-satellite-scale trend characteristics of the residuals. The infinite norm, 2-norm, mean standard deviation, and maximum standard deviation of the global pseudorange residuals are used as multi-satellite scale trend characteristics of the residuals. The logarithmic mean, logarithmic maximum, logarithmic slope, and logarithmic 2-norm within the time window are used as geometric consistency features for extracting RAIM residual sums. The single-satellite-scale trend features, multi-satellite-scale trend features, and geometric consistency features are concatenated to construct a multi-scale trend feature vector.
[0011] Furthermore, the weight-adaptive network is trained according to the following loss function:
[0012] in, The pseudorange observation weight vector is the output of the weighted adaptive network. The pseudorange observation weights output by the weighted adaptive network are exponentially reparameterized. Let be the 2-norm of the RAIM pseudorange residuals. The geometric matrix is obtained by linearizing the GNSS pseudorange observation equation. Used to transform a weight vector into a diagonal matrix. and This is the penalty coefficient.
[0013] Furthermore, the state evaluation function is defined as follows:
[0014] in, The latent space feature vector is the output of the latent space encoder, and the number of diffusion steps in the forward diffusion process is... , and These represent the odd-numbered diffusion steps in the forward diffusion process. and even-numbered diffusion steps The corresponding noisy latent space feature vectors and Representing noise prediction networks right and The noise prediction results are used. Based on the comparison between the state function value and the adaptive threshold, samples with state function values lower than the preset threshold are identified as deceptive samples.
[0015] Compared with the prior art, the present invention has the following advantages: This invention discloses a GNSS spoofing detection method based on adaptive RAIM and a latent space diffusion model, addressing the vulnerability of GNSS receivers to spoofing attacks due to the inherent fragility of GNSS signals, and the low accuracy and stability of existing detection methods when facing diverse and covert spoofing attacks. This invention constructs an adaptive pseudorange residual feature extractor based on a weighted adaptive network and weighted least squares RAIM. The pseudorange observation weight matrix of the weighted least squares RAIM is adaptively modified through the weighted adaptive network to enhance the sensitivity of the pseudorange residuals to geometric consistency violations caused by spoofing, thereby extracting multi-scale trend features of the residuals. Simultaneously, a GNSS spoofing detection model based on a latent space diffusion model is constructed. The latent space diffusion model is trained using unspoofed GNSS samples, and spoofing detection is performed using a state evaluation function.
[0016] This invention can effectively improve the receiver's detection performance against GNSS spoofing attacks. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is an architecture diagram of a GNSS spoofing detection method based on adaptive RAIM and latent space diffusion model in an embodiment of the present invention.
[0019] Figure 2 Weighted adaptive network architecture diagram in this embodiment of the invention.
[0020] Figure 3 This invention presents an architecture diagram for GNSS spoofing detection based on a latent space diffusion model. Detailed Implementation
[0021] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.
[0022] This invention discloses a GNSS spoofing detection method based on adaptive RAIM and latent space diffusion model. The execution architecture of this detection method is as follows: Figure 1 As shown. In this embodiment, the method mainly includes the following steps.
[0023] S1. Construct a pseudorange residual feature extractor, which includes a weighted adaptive network and a weighted least squares RAIM. The weighted adaptive network is used to generate pseudorange observation weights that are adapted to the constellation structure based on the spatial geometric relationship between the satellite and the receiver. The weighted least squares RAIM is used to perform weighted least squares RAIM calculation on the pseudorange observation weight matrix to obtain pseudorange residual features. In this embodiment, the weight adaptive network includes a first graph convolutional neural network and a second graph convolutional neural network. For example... Figure 2 As shown, the first convolutional neural network processes the input adjacency matrix and feature matrix sequentially through adjacency normalization, feature aggregation, feature mapping, and ReLU activation, outputting the inter-layer feature matrix. The second convolutional neural network processes the adjacency matrix and inter-layer feature matrix after adjacency normalization sequentially through feature aggregation and weight mapping, outputting the pseudorange observation weights for each satellite.
[0024] Specifically, the weighted adaptive network is built upon a two-layer graph convolutional neural network. Its adjacency matrix is constructed from the adjacency matrix of a fully connected undirected graph based on all satellites captured within the current epoch. Each satellite serves as a node in the fully connected undirected graph, where any two satellites are connected by an edge, and the weights of each edge are set to equal constants. The feature matrix of the weighted adaptive network is constructed based on the satellite elevation vector, the satellite-receiver three-dimensional line-of-sight vector, and a unit constant bias vector. The bias vector serves as a baseline feature independent of satellite geometry, used to enhance the stability of the training process. It is assumed that a total of [number missing] satellites were captured within the current epoch. If there are 10 satellites, then the feature matrix of the weighted adaptive network can be expressed as: ,in For the first The feature vector of a satellite, which can be expressed as: ,in , , The first The three-dimensional components of the satellite-receiver line-of-sight vector of each satellite. Here, θ represents the satellite elevation angle, and 1 represents a constant bias. To avoid parameter degradation such as negative weights, the output of the weighted adaptive network is reparameterized as follows:
[0025] in, For the output of the weighted adaptive network, This is the reparameterized result of the weight adaptive network output. These are the prior weights, typically set to 1. The loss function for a weighted adaptive network is defined as follows:
[0026] in, The pseudorange observation weight vector is the output of the weighted adaptive network. The pseudorange observation weights output by the weighted adaptive network are exponentially reparameterized. Let be the 2-norm of the RAIM pseudorange residuals. The geometric matrix is obtained by linearizing the GNSS pseudorange observation equation. Used to transform a weight vector into a diagonal matrix. and is the penalty coefficient, where The range of values is within Between and 1, The range of values is within and between.
[0027] Weighted least squares RAIM uses pseudorange observation weights learned by a weighted adaptive network to perform one RAIM calculation, obtaining the adaptive pseudorange residuals for each satellite.
[0028] The pseudorange residuals are then further processed to extract multi-scale trend features. First, the pseudorange residual features are Z-score standardized. Then, their single-satellite scale trend features are extracted, including the mean, infinity norm, and 2-norm of the single-satellite pseudorange residuals. Next, their multi-satellite scale trend features are extracted, including the infinity norm, 2-norm, mean standard deviation, and maximum standard deviation of the global pseudorange residuals. The multi-satellite pseudorange residual set, single-satellite scale trend features, and multi-satellite scale features within the time window are defined as follows:
[0029]
[0030]
[0031] in This is the set of pseudorange residuals from multiple satellites within a time window. The time window length, The number of satellites acquired by the receiver. express Time of the first The pseudorange residual of each satellite. For the first The single-scale trend feature vector of a satellite has the following components: the first component is the mean within the time window, the second component is the infinite norm within the time window, and the third component is the 2-norm within the time window. Multi-satellite scale trend characteristics, for The standard deviation of all satellite residuals at any given time. The first term is the global infinity norm, the second term is the global 2-norm, the third term is the global mean standard deviation, and the fourth term is the global maximum standard deviation. Finally, geometric consistency features based on the RAIM residual sum are extracted, defined as follows:
[0032]
[0033] in Represents the RAIM residuals and the values relative to the chi-square threshold. The geometric consistency index is obtained by normalizing and taking the logarithm. represent The residual vector composed of the pseudorange residuals corresponding to each satellite at time step 1. Here is the RAIM pseudorange observation weight matrix, and the chi-square threshold is... Chi-square distribution quantiles, among which For pseudorange observation vector dimension, For the receiver state vector dimension, For the degrees of freedom of the chi-square distribution, The false alarm rate is set in advance. These are geometrically consistent eigenvectors. The first element represents The mean within the time window, the second element represents The maximum value within the time window, the third element represents The time slope, where Time index within the time window The mean, Within the time window The mean, the fourth element represents The 2-norm of within the time window.
[0034] S2. Obtain unspoofed GNSS pseudorange measurement data, satellite coordinates, and the initial position coordinates of the receiver. Train the weighted adaptive network in the pseudorange residual feature extractor based on the GNSS pseudorange measurement, satellite coordinates, and the initial position coordinates of the receiver.
[0035] S3. Calculate the RAIM pseudorange residual features based on the pseudorange observation weights output by the trained weighted adaptive network, as well as the GNSS pseudorange measurement, satellite coordinates, and the receiver's initial position coordinates.
[0036] S4. Standardize the pseudorange residual features and then obtain the multi-scale trend features used to train the latent space diffusion model.
[0037] S5. Construct a latent space diffusion model, which includes a latent space encoder, a forward diffusion process, and a noise prediction network. The latent space encoder maps high-dimensional multi-scale trend features to latent space features suitable for modeling the diffusion model, thereby improving the stability of the noise prediction network training process. The forward diffusion process progressively injects Gaussian noise into the latent space features to generate noisy latent space features. The noise prediction network performs noise prediction on the noisy latent space features.
[0038] The loss function of the latent space diffusion model is defined as follows:
[0039] in For the original sample, Its statistical distribution, The first step in the forward diffusion process The standard Gaussian noise applied in each diffusion step The unit covariance matrix, This represents the total number of diffusion steps. For the first Samples from each diffusion step Indicates the noise prediction network for The noise prediction results This represents the noise prediction error. The expected value of the noise prediction error across all diffusion steps is used as the loss.
[0040] The latent space diffusion model learns the statistical distribution of pseudorange residuals based on multi-scale trend features. It should be noted that during the training phase, the latent space diffusion model only uses unspoofed GNSS data for learning.
[0041] S6. Acquire the GNSS pseudorange measurement data to be detected, satellite coordinates, and the initial position coordinates of the receiver, and train the weighted adaptive network in the pseudorange residual feature extractor. Calculate the RAIM pseudorange residual features based on the pseudorange observation weights output by the trained weighted adaptive network, as well as the GNSS pseudorange measurements, satellite coordinates, and the initial position coordinates of the receiver. Standardize the pseudorange residual features, and then obtain the multi-scale trend features to be detected. Based on the trained latent space diffusion model, perform noise prediction on the multi-scale trend features to be detected, and obtain the corresponding noise prediction results. Perform noise consistency verification on the noise prediction results using a state evaluation function and an adaptive threshold, thereby obtaining the GNSS spoofing detection result.
[0042] In the deception detection phase, noise prediction results at different diffusion time steps are obtained through a latent space diffusion model, and deception detection is performed using a state evaluation function. The state evaluation function is defined as follows:
[0043] in, The latent space feature vector is the output of the latent space encoder, and the number of diffusion steps in the forward diffusion process is... , and These represent the odd-numbered diffusion steps in the forward diffusion process. and even-numbered diffusion steps The corresponding noisy latent space feature vectors and Representing noise prediction networks right and The noise prediction results are used. Based on the comparison between the state function value and the adaptive threshold, samples with state function values lower than the preset threshold are identified as deceptive samples.
[0044] Since deception is a low-probability event, it is assumed that the GNSS receiver will not be deceived for a period of time after it has been operating normally, using the mean of the state function within this confidence interval. and standard deviation The difference As a threshold, This is a scaling factor. Preferably, the length of the confidence interval is between 10 and 200 seconds after the GNSS receiver has been operating normally. Samples with state function values below a threshold are marked as spoofed samples.
[0045] The following specific application examples will further illustrate the solution and effects of the present invention.
[0046] This invention provides a GNSS spoofing detection method based on adaptive RAIM and latent space diffusion model, the algorithm architecture of which is as follows: Figure 1 As shown, this method consists of an adaptive pseudorange residual feature extractor and a latent space diffusion model. The specific implementation steps of this method will be described in detail below.
[0047] Step 1: Dataset Preparation This implementation first constructs a dataset for model training and model testing by acquiring GNSS data, which includes pseudorange measurement data, satellite coordinate data, and receiver initial coordinate data. Unspoofed GNSS data is used as the training dataset, and potentially spoofed GNSS data is used as the test dataset.
[0048] Step 2: Construction and Implementation of Adaptive Pseudorange Residual Feature Extractor This implementation uses a weighted adaptive network to learn the RAIM weights. The weighted adaptive network architecture is as follows: Figure 2 As shown. The weighted adaptive network is constructed based on a two-layer graph convolutional neural network. Its adjacency matrix is a fully connected undirected graph constructed based on all satellites captured within the current epoch, where each satellite is a graph node, and any two satellites are connected by an edge. The weights of each edge are set to equal weight constants, here set to 1. The feature matrix of the weighted adaptive network is constructed based on the satellite elevation vector, the satellite-receiver three-dimensional line-of-sight vector, and a unit constant bias vector. The bias vector serves as a baseline feature independent of satellite geometry, used to enhance the stability of the training process. The first graph convolutional layer of the weighted adaptive network takes the adjacency matrix and the feature matrix as input, and after feature aggregation, feature mapping, and ReLU activation, outputs an inter-layer feature matrix. The second graph convolutional layer takes the adjacency matrix and the inter-layer feature matrix as input, and its output is used to determine the pseudorange observation weights of each satellite in RAIM. The weight matrix of the first layer graph convolution is 5×32, and the weight matrix of the second layer graph convolution is 32×1. In this implementation, the adjacency matrix is set to the adjacency matrix of the fully connected undirected graph. The weighted adaptive network is trained according to the aforementioned loss function and then backpropagated. The training rate is 0.001, the geometric penalty term in the loss function is set to 0.1, and the numerical penalty term is set to 0.01. After training, a weighted least squares RAIM is performed with the current weight matrix to obtain the adaptive pseudorange residual features, and the multi-scale trend features are further extracted as follows: First, the adaptive pseudorange residual features are Z-score standardized. Then, their single-satellite scale trend features are extracted, including the mean, infinity norm, and 2-norm of the single-satellite pseudorange residuals. Next, their multi-satellite scale trend features are extracted, including the infinity norm, 2-norm, mean standard deviation, and maximum standard deviation of the global pseudorange residuals. The multi-satellite pseudorange residual set, single-satellite scale trend features, and multi-satellite scale features within the time window are defined as follows:
[0049]
[0050]
[0051] in This is the set of pseudorange residuals from multiple satellites within a time window. The time window length, The number of satellites acquired by the receiver. express Time of the first The pseudorange residual of each satellite. For the first The single-scale trend feature vector of a satellite has the following components: the first component is the mean within the time window, the second component is the infinite norm within the time window, and the third component is the 2-norm within the time window. Multi-satellite scale trend characteristics, for The standard deviation of all satellite residuals at any given time. The first term is the global infinity norm, the second term is the global 2-norm, the third term is the global mean standard deviation, and the fourth term is the global maximum standard deviation. Finally, geometric consistency features based on the RAIM residual sum are extracted, defined as follows:
[0052]
[0053] in Represents the RAIM residuals and the values relative to the chi-square threshold. The geometric consistency index is obtained by normalizing and taking the logarithm. represent The residual vector composed of the pseudorange residuals corresponding to each satellite at time step 1. Here is the RAIM pseudorange observation weight matrix, and the chi-square threshold is... Chi-square distribution quantiles, among which For pseudorange observation vector dimension, For the receiver state vector dimension, For the degrees of freedom of the chi-square distribution, The false alarm rate is set in advance. These are geometrically consistent eigenvectors. The first element represents The mean within the time window, the second element represents The maximum value within the time window, the third element represents The time slope, where Time index within the time window The mean, Within the time window The mean, the fourth element represents The 2-norm of within the time window.
[0054] The single-satellite-scale trend features, multi-satellite-scale trend features, and geometric consistency features are concatenated to construct a multi-scale trend feature vector. This multi-scale trend feature vector serves as the input sample for the latent space diffusion model.
[0055] Step 3: Construction and Implementation of the Latent Space Diffusion Model The architecture of the latent space diffusion model is as follows: Figure 3 As shown. The latent space diffusion model uses the multi-scale trend feature vector from step 2 as input samples. Considering that the multi-scale trend feature vector has a high dimension and is not suitable for direct use in diffusion model training, this implementation uses a two-layer multilayer perceptron as the latent space encoder to map the multi-scale trend features into latent space features suitable for diffusion modeling. The first hidden layer of the latent space encoder has a dimension of 64, and the second hidden layer has a dimension of 8. A ReLU activation function is used for non-linear activation between the first and second layers. The latent space feature vector output by the latent space encoder is embedded and fused with the normalized time step vector to obtain a feature vector with a dimension of 9. During the forward diffusion process, Gaussian noise is gradually injected into the latent space feature vector over 100 diffusion steps, with the noise intensity ranging from... linearly increase to The noise prediction network is a 3-layer multilayer perceptron with hidden layer dimensions of 64, 64, and 8. The network is trained using the noisy latent space features at each diffusion step as input, and employs the following loss function:
[0056] in For the original sample, Its statistical distribution, The first step in the forward diffusion process The standard Gaussian noise applied in each diffusion step The unit covariance matrix, This represents the total number of diffusion steps. For the first Samples from each diffusion step Indicates the noise prediction network for The noise prediction results This represents the noise prediction error. The expected value of the noise prediction error across all diffusion steps is used as the loss.
[0057] During the training phase, multi-scale trend feature vectors corresponding to unspoofed GNSS data are used as training samples to construct a latent space diffusion model. During the testing phase, multi-scale trend feature vectors corresponding to the GNSS data to be detected are used as input samples. Noise consistency prediction is performed through the latent space diffusion model, and spoofing detection is performed in conjunction with the state estimation function.
[0058] Step 4: Experimental Verification of the Spoofing Detection Algorithm Using unspoofed GNSS data as training data, pseudorange residual features are extracted using the feature extractor described in step 2. Then, multi-scale trend feature vectors of the residuals are extracted and used as training samples for the latent space diffusion model. The time window length of the feature extractor is set to 100, and the sampling step size is 3. The batch size for the latent space diffusion model training phase is set to 128, the learning rate is set to 0.001, and the number of training epochs is set to 500. The latent space diffusion model is trained based on these samples and parameters, and the resulting model is used for deception detection.
[0059] Using the GNSS data to be detected as test data, pseudorange residual features are extracted using the feature extractor described in step 2. Then, multi-scale trend feature vectors of the residuals are extracted, and these multi-scale trend feature vectors are used as test samples for the latent space diffusion model. The time window length of the feature extractor is set to 100, and the sampling step size is 3. The confidence interval is the first 100 seconds after the GNSS receiver has been operating normally. The trained latent space diffusion model is used to infer the test samples, and deception detection is performed using the aforementioned state evaluation function and adaptive threshold. At this time, the adaptive threshold is calculated based on the state function values within the confidence interval, and its scaling factor is... Set it to 1.
[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A GNSS spoofing detection method based on adaptive RAIM and latent space diffusion model, characterized in that, Includes the following steps: A pseudorange residual feature extractor is constructed, which includes a weighted adaptive network and a weighted least squares RAIM. The weighted adaptive network is used to generate a pseudorange observation weight matrix that is adapted to the constellation structure based on the spatial geometric relationship between the satellite and the receiver. The weighted least squares RAIM is used to calculate pseudorange residual features based on the pseudorange observation weight matrix. Unaffected GNSS pseudorange measurement data, satellite coordinates, and the receiver's initial position coordinates are obtained. Based on these data, a weighted adaptive network in the pseudorange residual feature extractor is trained. The pseudorange observation weights output by the trained network, along with the GNSS pseudorange measurements, satellite coordinates, and the receiver's initial position coordinates, are used to calculate RAIM pseudorange residual features. These features are then standardized, and multi-scale trend features are obtained for training the latent space diffusion model. A latent space diffusion model is constructed, comprising a latent space encoder, a forward diffusion process, and a noise prediction network. The latent space encoder maps high-dimensional multi-scale trend features to latent space features suitable for modeling the diffusion model, thereby improving the stability of the noise prediction network training process. The forward diffusion process progressively injects Gaussian noise into the latent space features to generate noisy latent space features. The noise prediction network performs noise prediction on the noisy latent space features. The latent space diffusion model is trained using the multi-scale trend features as the original input to learn the statistical distribution of the multi-scale trend features under normal GNSS conditions. The process involves acquiring GNSS pseudorange measurement data, satellite coordinates, and the receiver's initial position coordinates, and training a weighted adaptive network in the pseudorange residual feature extractor. Based on the pseudorange observation weights output by the trained weighted adaptive network, and the GNSS pseudorange measurements, satellite coordinates, and receiver's initial position coordinates, RAIM pseudorange residual features are calculated. These pseudorange residual features are then standardized, and subsequently, multi-scale trend features of the data to be detected are obtained. Based on the trained latent space diffusion model, noise prediction is performed on the multi-scale trend features of the data to be detected, and the corresponding noise prediction results are obtained. The noise prediction results are tested for noise consistency using a state evaluation function and an adaptive threshold, thereby obtaining the GNSS spoofing detection results.
2. The GNSS spoofing detection method based on adaptive RAIM and latent space diffusion model according to claim 1, characterized in that, The weight adaptive network includes a first graph convolutional neural network and a second graph convolutional neural network; The first convolutional neural network is used to perform adjacency relationship normalization, feature aggregation, feature mapping and ReLU activation function processing on the adjacency matrix and feature matrix in sequence, and output the inter-layer feature matrix. The adjacency matrix is an adjacency matrix of a fully connected undirected graph constructed based on the satellites in the current epoch. The feature matrix is constructed based on the satellite elevation vector, the satellite-receiver three-dimensional line-of-sight vector, and the unit constant bias vector. The second convolutional neural network is used to perform feature aggregation and weight mapping on the adjacency matrix and the inter-layer feature matrix after adjacency relationship normalization, and output the pseudorange observation weights of each satellite.
3. The GNSS spoofing detection method based on adaptive RAIM and latent space diffusion model according to claim 2, characterized in that, The fully connected undirected graph is constructed as follows: each satellite is treated as a graph node, and an edge is connected between any two satellites, with the weight of each edge set to an equal weight constant.
4. The GNSS spoofing detection method based on adaptive RAIM and latent space diffusion model according to claim 1, characterized in that, Statistical processing is performed on the standardized pseudorange residual features to obtain multi-scale trend features, including: The mean, infinite norm, and 2-norm of the pseudorange residuals of individual satellites are statistically analyzed as the single-satellite-scale trend characteristics of the residuals. The infinite norm, 2-norm, mean standard deviation, and maximum standard deviation of the global pseudorange residuals are used as multi-satellite scale trend characteristics of the residuals. The logarithmic mean, logarithmic maximum, logarithmic slope, and logarithmic 2-norm within the time window are used as geometric consistency features based on the sum of RAIM residuals. The single-satellite-scale trend features, multi-satellite-scale trend features, and geometric consistency features are concatenated to construct a multi-scale trend feature vector.
5. The GNSS spoofing detection method based on adaptive RAIM and latent space diffusion model according to claim 1, characterized in that, The weighted adaptive network is trained using the following loss function: in, The pseudorange observation weight vector is the output of the weighted adaptive network. The pseudorange observation weights output by the weighted adaptive network are exponentially reparameterized. Let be the 2-norm of the RAIM pseudorange residuals. The geometric matrix is obtained by linearizing the GNSS pseudorange observation equation. Used to transform a weight vector into a diagonal matrix. and This is the penalty coefficient.
6. The GNSS spoofing detection method based on adaptive RAIM and latent space diffusion model according to claim 1, characterized in that, The state evaluation function is defined as follows: in, The latent space feature vector is the output of the latent space encoder, and the number of diffusion steps in the forward diffusion process is... , and These represent the odd-numbered diffusion steps in the forward diffusion process. and even-numbered diffusion steps The corresponding noisy latent space feature vectors and Representing noise prediction networks right and The noise prediction results.