A Multi-Modal GNSS Spoofing Detection Method Based on Dynamic Weighting

Abstract

The present invention relates to the technical field of GNSS spoofing detection, and provides a multi-modal GNSS spoofing detection method based on dynamic weighting, including obtaining a one-dimensional SQM time series and a two-dimensional correlation function and performing preprocessing; extracting features from the standardized two-dimensional correlation function through a Transformer channel to obtain global features; extracting features from the standardized two-dimensional correlation function through a CNN channel to obtain local features; obtaining a dynamic weight matrix for the Transformer channel and a dynamic weight matrix for the CNN channel through the one-dimensional SQM time series; multiplying and adding the global features and the local features with the corresponding weights to obtain fused features; performing global average pooling on the fused features, and mapping them to a spoofing probability through a fully connected network; detecting spoofing according to the spoofing probability and a probability threshold. The present invention significantly improves the detection accuracy in complex environments and enhances the security and reliability of the system.

Description

Technical Field

[0001] The present invention relates to the technical field of GNSS spoofing detection, and in particular to a multi-modal GNSS spoofing detection method based on dynamic weighting. Background Art

[0002] The signals of the Global Navigation Satellite System (GNSS) are vulnerable to various spoofing attacks, such as signal retransmission, replay, and forgery. These spoofing attacks will seriously interfere with the normal operation of GNSS receivers, resulting in errors or even failures in functions such as positioning, navigation, and timing, thus posing a threat to the safe and stable operation of related fields.

[0003] Currently, the methods for GNSS spoofing detection are mainly divided into two categories. The first category is traditional signal processing methods, which are based on correlation function analysis, such as using signal quality monitoring (SQM) indicators, carrier-to-noise ratio, etc. for detection. However, these methods rely on manually designed features. For complex and variable spoofing attack patterns, the difficulty of feature design is large and it is difficult to cover comprehensively, so the detection effect is limited.

[0004] The second category is deep learning methods, which mainly use models such as convolutional neural networks (CNNs) or Transformers for feature extraction and detection. However, the existing deep learning methods have some obvious defects. First, most of them are single-modal processing, only dealing with one-dimensional time series or two-dimensional correlation functions, ignoring the complementarity between multi-modal data, and unable to fully mine the information contained in different types of data, so that the detection model cannot comprehensively capture the features of spoofing signals, thus reducing the accuracy and reliability of detection. Second, in the feature fusion process, static weights are used, and the importance of features cannot be dynamically adjusted according to the real-time signal quality and the changes of spoofing scenarios, making it difficult to adapt to complex and variable spoofing scenarios, resulting in a decline in detection performance in some special cases. Third, traditional Transformer models have a large number of parameters, high computational complexity, and high requirements for computing resources, resulting in slow inference speed and difficulty in effective deployment on embedded devices with high real-time requirements. Summary of the Invention

[0005] The present invention aims to at least solve one of the technical problems existing in the related technologies. For this purpose, the present invention provides a multi-modal GNSS spoofing detection method based on dynamic weighting, which significantly improves the detection ability of dynamic spoofing signals through multi-modal input, spatio-temporal alignment preprocessing, lightweight dual-channel feature extraction, and dynamic weight generation, significantly improves the accuracy of detection in complex environments, enhances the security and reliability of the system, and provides a more efficient and secure solution for the application of GNSS.

[0006] The present invention provides a multi-modal GNSS spoofing detection method based on dynamic weighting, including:

[0007] S1: Obtain a one-dimensional SQM time series and a two-dimensional correlation function, and preprocess the one-dimensional SQM time series and the two-dimensional correlation function to obtain a standardized one-dimensional SQM time series and a standardized two-dimensional correlation function;

[0008] S2: Extract global features through the Transformer channel for the standardized two-dimensional correlation function, and extract local features through the CNN channel for the standardized two-dimensional correlation function;

[0009] S3: Extract the temporal characteristics of the one-dimensional SQM time series through LSTM to obtain the dynamic weight matrix of the Transformer channel; extract the channel features of the one-dimensional SQM time series through the squeeze-and-excitation network to obtain the dynamic weight matrix of the CNN channel;

[0010] S4: Multiply the global features by the dynamic weight matrix of the Transformer channel to obtain the Transformer channel features, multiply the local features by the dynamic weight matrix of the CNN channel to obtain the CNN channel features, and fuse the Transformer channel features and the CNN channel features to obtain the fused features;

[0011] S5: Perform global average pooling on the fused features and map them to the spoofing probability through a fully connected network;

[0012] S6: Detect spoofing according to the spoofing probability and the probability threshold.

[0013] Further, step S1 includes:

[0014] S11: Obtain a one-dimensional SQM time series by acquiring the real-time measurement values of the GNSS receiver, and the real-time measurement values include the carrier-to-noise ratio, code phase error, phase jitter, and multipath index;

[0015] S12: Obtain the code phase by searching for the phase with the strongest correlation between the local code and the received signal in the GNSS receiver, perform frequency offset adjustment on the received signal through Doppler frequency compensation to obtain the Doppler frequency, use the code phase as the horizontal axis, the Doppler frequency as the vertical axis, and the correlation value as the pixel value to construct a two-dimensional correlation function;

[0016] S13: Select the minimum time step of the one-dimensional SQM time series and the two-dimensional correlation function, and align the one-dimensional SQM time series and the two-dimensional correlation function in the time dimension through the minimum time step;

[0017] S14: Normalize the aligned one-dimensional SQM time series to obtain a standardized one-dimensional SQM time series, and normalize the aligned two-dimensional correlation function to obtain a standardized two-dimensional correlation function.

[0018] Furthermore, extract features from the standardized two-dimensional correlation function through the Transformer channel to obtain global features including:

[0019] S211: Extract features from the standardized two-dimensional correlation function through the multi-head self-attention mechanism to obtain multi-head attention features;

[0020] S212: Retain the order of the standardized two-dimensional correlation function through sinusoidal position encoding to obtain position encoding information;

[0021] S213: Add the multi-head attention features and the position encoding information to obtain global features.

[0022] Furthermore, extract features from the standardized two-dimensional correlation function through the CNN channel to obtain local features including:

[0023] S221: Construct a residual block, which includes a 3×3 convolutional layer, a batch normalization layer, and an activation function layer;

[0024] S222: Extract local features of the standardized two-dimensional correlation function through the residual block.

[0025] Furthermore, step S3 includes:

[0026] S31: Convert the standardized one-dimensional SQM time series into a high-dimensional feature vector through a one-dimensional convolutional layer;

[0027] S32: Capture the temporal dynamic features of the high-dimensional feature vector through LSTM to obtain a dynamic weight matrix, and adjust the dimension of the dynamic weight matrix through bilinear interpolation to obtain the dynamic weight matrix of the Transformer channel;

[0028] S33: Extract channel features of the high-dimensional feature vector through a squeeze-and-excitation network to obtain the CNN channel dynamic weight matrix, and the calculation expression is:

[0029]

[0030] Among them, is the CNN channel attention weight, is the non-linear activation function, is the second fully connected layer, is the operation sequence connection symbol, is the linear activation function, is the first fully connected layer, is the global average pooling function, is the high-dimensional feature vector.

[0031] Furthermore, in step S5,

[0032] perform global average pooling on the fused feature to obtain the fused feature vector,

[0033]

[0034] where, is the fused feature vector, is the value of the fused feature at the th time step, the th row, and the th column, H is the length, and W is the width;

[0035] map the fused feature vector to the spoofing probability through a fully connected network;

[0036]

[0037] where, is the spoofing probability, is the non-linear activation function, is the first weight matrix, is the second weight matrix, is the linear activation function, is the first bias term, is the second bias term, is the transpose of the matrix.

[0038] Furthermore, in step S6,

[0039] if the spoofing probability is greater than or equal to the probability threshold, then the current signal is a spoofing signal;

[0040] if the spoofing probability is less than the probability threshold, then the current signal is a genuine signal.

[0041] Furthermore, perform a moving average on the spoofing probabilities of consecutive time steps through a time series smoothing method to obtain the smoothed spoofing probability;

[0042] if multiple consecutive smoothed spoofing probabilities are greater than or equal to the probability threshold, then trigger an alarm, and the current signal is a spoofing signal; otherwise, maintain the current determination result.

[0043] Furthermore, the probability threshold selection conditions include the false alarm rate, the miss rate, and environmental parameters;

[0044] balance the false alarm rate and the miss rate through cross-validation or the ROC curve to obtain the optimal probability threshold for the false alarm rate and the miss rate,

[0045] Dynamically select the probability threshold through linear regression or gating mechanism according to the optimal probability threshold of false alarm rate and missed alarm rate and environmental parameters.

[0046] Furthermore, the false alarm rate is the probability that a true signal is misjudged as a spoofing signal, and the missed alarm rate is the probability that a spoofing signal is misjudged as a true signal.

[0047] One or more of the above technical solutions in the embodiments of the present invention have at least one of the following technical effects:

[0048] The present invention significantly improves the detection ability of dynamic spoofing signals through multi-modal input, spatio-temporal alignment preprocessing, lightweight dual-channel feature extraction, and dynamic weight generation, significantly improves the detection accuracy in complex environments, enhances the security and reliability of the system, and provides a more efficient and secure solution for the application of GNSS.

[0049] Additional aspects and advantages of the present invention will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0051] Figure 1 is a schematic flowchart of a multi-modal GNSS spoofing detection method based on dynamic weighting provided by the present invention.

[0052] Figure 2 is a comparison chart of the detection accuracy of the present invention and the existing method under different signal-to-noise ratios. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. The following embodiments are used to illustrate the present invention, but cannot be used to limit the scope of the present invention.

[0054] In the description of this specification, the description referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the embodiments of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0055] The following combines Figures 1 to 2 to describe a multi-modal GNSS spoofing detection method based on dynamic weighting of the present invention.

[0056] As Figure 1 shown, a multi-modal GNSS spoofing detection method based on dynamic weighting includes:

[0057] S1: Obtain a one-dimensional SQM time series and a two-dimensional correlation function, and preprocess the one-dimensional SQM time series and the two-dimensional correlation function to obtain a normalized one-dimensional SQM time series and a normalized two-dimensional correlation function;

[0058] Generate a one-dimensional SQM time series and a two-dimensional correlation function using a multi-correlator;

[0059] S11: Obtain a one-dimensional SQM time series by acquiring real-time measurements of a GNSS receiver, where the real-time measurements include carrier-to-noise ratio, code phase error, phase jitter, and multipath index; the time step of the one-dimensional SQM time series is , and the dimension is .

[0060] In some specific embodiments of the present invention, the one-dimensional SQM time series has 1000 time points and a sampling rate of 100 Hz, that is, a sample is output every 10 ms.

[0061] S12: Obtain the code phase by searching for the phase with the strongest correlation between the local code and the received signal in the GNSS receiver, perform frequency offset adjustment on the received signal through Doppler frequency compensation to obtain the Doppler frequency, use the code phase as the horizontal axis, the Doppler frequency as the vertical axis, and the correlation value as the pixel value to construct a two-dimensional correlation function;

[0062] The two-dimensional correlation function is the cross-correlation result between the local code and the received signal in the GNSS receiver, which reflects the matching degree of the signal in terms of code phase and Doppler frequency and is presented in the form of a two-dimensional heat map.

[0063] In some specific embodiments of the present invention, the local code is slid within a certain range (such as ±2 chips), and the correlation value between the local code and the received signal is calculated. By continuously changing the phase of the local code, the position with the strongest correlation with the received signal is found to obtain the code phase;

[0064] Due to the relative motion between the satellite and the receiver, the frequency of the received signal will shift. Therefore, it is necessary to adjust the frequency shift of the received signal to compensate for the influence of the Doppler effect. The frequency shift of the received signal is adjusted through Doppler frequency compensation to obtain the Doppler frequency.

[0065] Taking the code phase as the horizontal axis, the Doppler frequency as the vertical axis, and the correlation value as the pixel value, a matrix is formed, where is the length, is the width,

[0066] The time step of the two-dimensional correlation function is The spatial dimension is The number of channels is C, and the dimension is .

[0067] In some specific embodiments of the present invention, such as is 41×41, .

[0068] The correlation function of the real signal usually shows a single-peak symmetric distribution, and the main peak corresponds to the code phase and Doppler of the real satellite signal; while the spoofing signal will cause the correlation function to have multiple peaks, offsets, or distortions, etc. For example, the repeat-back spoofing will form a double-peak structure.

[0069] S13: Select the minimum time step of the one-dimensional SQM time series and the two-dimensional correlation function, and align the one-dimensional SQM time series and the two-dimensional correlation function in the time dimension through the minimum time step;

[0070] To ensure the synchronization of the one-dimensional SQM time series and the two-dimensional correlation function in the time dimension, select the minimum time step of the two,

[0071]

[0072] where is the minimum time step, is the time step of the one-dimensional SQM time series, is the time step of the two-dimensional correlation function, is the minimum value function.

[0073] In some specific embodiments of the present invention, if the time step of the one-dimensional SQM time series is 1000 and the time step of the two-dimensional correlation function is 800, then take This can ensure that in subsequent processing, the two types of data are time-corresponding, avoiding information mismatch problems caused by inconsistent time.

[0074] S14: Normalize the aligned one-dimensional SQM time series to obtain a standardized one-dimensional SQM time series, and normalize the aligned two-dimensional correlation function to obtain a standardized two-dimensional correlation function.

[0075] S141: Normalize the aligned one-dimensional SQM time series to obtain a standardized one-dimensional SQM time series. The calculation expression is:

[0076]

[0077] where is the standardized one-dimensional SQM time series at the th time step, is the one-dimensional SQM time series at the th time step, is the mean of the one-dimensional SQM time series, is the standard deviation of the one-dimensional SQM time series.

[0078] Through standardization processing, the influence of dimensions can be eliminated, making the data have a unified distribution and facilitating the learning of subsequent models.

[0079] S142: Normalize the aligned two-dimensional correlation function to obtain a standardized two-dimensional correlation function. The calculation expression is:

[0080]

[0081] where is the standardized two-dimensional correlation function at the th time step, the th row, and the th column, is the two-dimensional correlation function at the th time step, the th row, and the th column, is the minimum value of the two-dimensional correlation function, is the maximum value of the two-dimensional correlation function.

[0082] Normalization processing compresses the pixel values into the [0, 1] interval, avoiding the problem of numerical overflow, and also helps to improve the training efficiency and stability of the model.

[0083] S2: Extract global features by performing feature extraction on the normalized two-dimensional correlation function through the Transformer channel; extract local features by performing feature extraction on the normalized two-dimensional correlation function through the CNN channel;

[0084] Performing feature extraction on the normalized two-dimensional correlation function through the Transformer channel to obtain global features includes:

[0085] S211: Perform feature extraction on the normalized two-dimensional correlation function through the multi-head self-attention mechanism to obtain multi-head attention features; the calculation expression of the multi-head attention features is:

[0086]

[0087] Among them, The multi-head attention features, is the query matrix, is the key matrix, is the value matrix, is the first attention head, is the second attention head, is the th attention head, is the concatenation function, is the learnable weight matrix used to perform a linear transformation on the concatenated multi-head attention results and map them back to the original feature dimension, The calculation expression of

[0088]

[0089] Among them, is the query matrix of the mth attention head, is the key matrix of the mth attention head, is the value matrix of the mth attention head, is the dimension of the key, is the transpose of the matrix, is the normalization activation function.

[0090] In some specific embodiments of the present invention, , .

[0091] S212: Retain the order of the normalized two-dimensional correlation function through sine position encoding to obtain position encoding information;

[0092] The position encoding calculation expression is:

[0093]

[0094] Among them, is the The position encoding of the -dimensional feature at the th time step, is the position encoding of the -dimensional feature at the th time step, is the dimension index,

[0095] In some specific embodiments of the present invention, .

[0096] S213: Add the multi-head attention feature and the position encoding information to obtain the global feature.

[0097] Extract features from the normalized two-dimensional correlation function through the CNN channel to obtain local features including:

[0098] S221: Construct a residual block, the residual block includes a 3×3 convolutional layer, a batch normalization layer, and an activation function layer;

[0099]

[0100] Among them, is the residual block, is the normalized two-dimensional correlation function, ReLU (Rectified LinearUnit) is the linear activation function, is the 3×3 convolutional layer, is the operation sequence connection symbol, (BatchNormalization) is the batch normalization operation;

[0101] The use of batch normalization and linear activation functions helps to improve the training effect and stability of the model, is used to accelerate the model convergence speed and improve training stability, and the linear activation function is used to introduce linear operations, stride 1, padding 1, means to perform a 3×3 convolution operation first, then a batch normalization operation, and then another 3×3 convolution operation.

[0102] S222: Extract the local features of the normalized two-dimensional correlation function through the residual block.

[0103] S3: Extract the temporal characteristics of the standard one-dimensional SQM time series through LSTM to obtain the Transformer channel dynamic weight matrix; extract the channel features of the standard one-dimensional SQM time series through the squeeze-and-excitation network to obtain the CNN channel dynamic weight matrix;

[0104] S31: Convert the normalized one-dimensional SQM time series into a high-dimensional feature vector through a one-dimensional convolutional layer; the calculation expression is:

[0105]

[0106] where is the high-dimensional feature vector, is the one-dimensional convolutional layer, is the normalized one-dimensional SQM time series, The kernel size is 3 and the output dimension is 128.

[0107] S32: Capture the temporal dynamic features of the high-dimensional feature vector through LSTM to obtain a dynamic weight matrix, and adjust the dimension of the dynamic weight matrix through bilinear interpolation to obtain the dynamic weight matrix of the Transformer channel;

[0108] S33: Extract the channel features of the high-dimensional feature vector through Squeeze-and-Excitation Networks (SE-Net) to obtain the CNN channel dynamic weight matrix, and the calculation expression is:

[0109]

[0110] where is the CNN channel attention weight, is the non-linear activation function, is the second fully-connected layer, is the operation sequence connection symbol, is the linear activation function, is the first fully-connected layer, is the global average pooling function, is the high-dimensional feature vector,

[0111] for dimensionality reduction to 32, for restoring to the number of channels C.

[0112] SE-Net can automatically learn the importance of different channels, so as to allocate appropriate weights for the features of the CNN channels.

[0113] S4: Multiply the global features by the dynamic weight matrix of the Transformer channel to obtain the Transformer channel features, multiply the local features by the CNN channel dynamic weight matrix to obtain the CNN channel features, and fuse the Transformer channel features and the CNN channel features to obtain the fused features;

[0114] Multiply the global feature by the dynamic weight matrix of the Transformer channels to obtain the Transformer channel features. The calculation expression is:

[0115]

[0116] Where, is the Transformer channel feature, is the global feature, is the dynamic weight matrix of the Transformer channels;

[0117] Multiply the local feature by the dynamic weight matrix of the CNN channels to obtain the CNN channel features. The calculation expression is:

[0118]

[0119] Where, is the CNN channel feature, is the local feature, is the dynamic weight matrix of the CNN channels;

[0120] Fuse the Transformer channel features and the CNN channel features to obtain the fused features. The calculation expression is:

[0121]

[0122] Where, is the fused feature;

[0123] Through this weighted fusion method, dynamic adjustment can be made according to the importance of different channel features, enhancing the model's ability to perceive deception signals.

[0124] S5: Perform global average pooling on the fused features and map them to the deception probability through a fully connected network;

[0125] Perform global average pooling on the fused features to obtain the fused feature vector,

[0126]

[0127] Where, is the fused feature vector, is the value of the fused feature at the th time step, the th row, and the th column; H is the length, W is the width,

[0128] Map the fused feature vector to the deception probability through a fully connected network;

[0129]

[0130] Among them, is the deception probability, is a non-linear activation function, is the first weight matrix, is the second weight matrix, is a linear activation function, is the first bias term, is the second bias term, is the transpose of the matrix.

[0131] S6. Detect deception according to the deception probability and the probability threshold;

[0132] The conditions for selecting the probability threshold include the false positive rate (FPR), the false negative rate (FNR), and environmental parameters;

[0133] The false positive rate is the probability that a true signal is misjudged as a deception signal, and the false negative rate is the probability that a deception signal is misjudged as a true signal.

[0134] Balance the false positive rate and the false negative rate through cross-validation or the ROC curve (Receiver Operating Characteristic Curve) to obtain the optimal probability threshold for the false positive rate and the false negative rate;

[0135] According to the changes in the real-time signal environment, such as the noise level of the signal, the reliability requirements, etc., dynamically adjust the probability threshold. Specifically, in a high-noise environment, in order to reduce the false negative situation, appropriately lower the threshold; in a scenario with high reliability requirements, in order to reduce the false positive rate, increase the threshold. To achieve this dynamic adjustment, introduce environmental parameters (such as the carrier-to-noise ratio) as auxiliary inputs, and dynamically select the probability threshold through linear regression or a gating mechanism.

[0136] If the deception probability is greater than or equal to the probability threshold, the current signal is a deception signal;

[0137] If the deception probability is less than the probability threshold, the current signal is a true signal.

[0138] In order to suppress instantaneous noise interference and avoid single misjudgment, post-process and optimize the output probability. Perform a moving average on the deception probabilities of consecutive time steps through a time series smoothing method to obtain the smoothed deception probability;

[0139] With as the window size, average the probabilities of time step and the previous time steps to obtain the smoothed probability. The calculation expression is:

[0140]

[0141] For the smooth deception probability, is the deception probability at the

[0142] If the smoothed probabilities for multiple consecutive time steps (such as 3) exceed the threshold, an alarm is triggered and it is determined that there is a deception signal; otherwise, the current determination result is maintained.

[0143] If multiple consecutive smooth deception probabilities are greater than or equal to the probability threshold, an alarm is triggered, and the current signal is a deception signal; otherwise, the current determination result is maintained.

[0144] The present invention significantly improves the detection ability of dynamic deception signals through multi-modal input, spatio-temporal alignment preprocessing, lightweight dual-channel feature extraction, and dynamic weight generation, significantly improves the detection accuracy in complex environments, enhances the security and reliability of the system, and provides a more efficient and secure solution for the application of GNSS.

[0145] Embodiment:

[0146] Perform deception detection on in-vehicle navigation. Data collection: The GNSS receiver outputs 100 groups of SQM data per second, and the baseband processor generates a correlation function every 10 ms. The data collection duration is 24 hours, covering various scenarios such as urban roads, highways, and tunnels.

[0147] Real-time inference: The collected data is preprocessed and then input into a dual-channel network. According to the characteristics of the SQM sequence and the correlation function, the weights are dynamically adjusted for feature fusion and deception detection.

[0148] Decision output: When the detection probabilities for 3 consecutive frames exceed the set threshold, an alarm is triggered and the system switches to the inertial navigation system.

[0149] The deception detection accuracy of in-vehicle navigation in different scenarios is shown in Table 1.

[0150] Table 1 Deception detection accuracy of in-vehicle navigation in different scenarios

[0151]

[0152] Detection latency: In the urban road scenario, the average detection latency is 18 ms; in the highway scenario, the average detection latency is 15 ms; in the tunnel entrance scenario, the average detection latency is 20 ms.

[0153] The in-vehicle navigation is respectively subjected to deception detection using the method of the present invention, traditional CNN, and pure Transformer, and the performance comparison results are shown in Table 2.

[0154] Table 2 Performance Comparison among the Method of the Present Invention, Traditional CNN, and Pure Transformer

[0155] As can be seen from Table 2, compared with the pure Transformer model, the model size, inference latency, false alarm rate, and miss rate of the method of the present invention are significantly reduced.

[0156] In order to verify the anti-interference ability of the present invention in different noise environments, detection experiments were carried out at different signal-to-noise ratios. In the experiment, Gaussian white noise was artificially added to simulate different signal-to-noise ratio environments. The experimental data are as Figure 2 shown. The accuracy of the present invention is 95.3% at -20dB, while that of the pure Transformer model is only 82.1%. The present invention can still maintain a high detection accuracy in a low signal-to-noise ratio environment, showing a significant advantage compared with the pure Transformer model.

[0157] The present invention significantly improves the anti-noise ability through multi-modal fusion and dynamic weights, and is suitable for complex electromagnetic environments.

[0158] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-modal GNSS spoofing detection method based on dynamic weighting, characterized in that Including: S1: Obtain a one-dimensional SQM time series and a two-dimensional correlation function, and preprocess the one-dimensional SQM time series and the two-dimensional correlation function to obtain a normalized one-dimensional SQM time series and a normalized two-dimensional correlation function; S2: Extract features of the normalized two-dimensional correlation function through the Transformer channel to obtain global features; extract features of the normalized two-dimensional correlation function through the CNN channel to obtain local features; S3: Extract the time characteristics of the normalized one-dimensional SQM time series through LSTM to obtain the dynamic weight matrix of the Transformer channel; extract the channel features of the normalized one-dimensional SQM time series through the squeeze-and-excitation network to obtain the dynamic weight matrix of the CNN channel; S4: Multiply the global features by the dynamic weight matrix of the Transformer channel to obtain the Transformer channel features, multiply the local features by the dynamic weight matrix of the CNN channel to obtain the CNN channel features, and fuse the Transformer channel features and the CNN channel features to obtain the fused features; S5: Perform global average pooling on the fused features and map them to the spoofing probability through a fully connected network; S6: Detect spoofing based on the spoofing probability and the probability threshold.

2. The multi-modal GNSS spoofing detection method based on dynamic weighting according to claim 1, wherein Step S1 includes: S11: Obtain a one-dimensional SQM time series by acquiring the real-time measurement values of the GNSS receiver, where the real-time measurement values include the carrier-to-noise ratio, code phase error, phase jitter, and multipath index; S12: Obtain the code phase by searching for the phase with the strongest correlation between the local code and the received signal in the GNSS receiver, perform frequency offset adjustment on the received signal through Doppler frequency compensation to obtain the Doppler frequency, use the code phase as the horizontal axis, the Doppler frequency as the vertical axis, and the correlation value as the pixel value to construct a two-dimensional correlation function; S13: Select the minimum time step of the one-dimensional SQM time series and the two-dimensional correlation function, and align the one-dimensional SQM time series and the two-dimensional correlation function in the time dimension through the minimum time step; S14: Perform normalization processing on the aligned one-dimensional SQM time series to obtain a normalized one-dimensional SQM time series, and perform normalization processing on the aligned two-dimensional correlation function to obtain a normalized two-dimensional correlation function.

3. A multi-modal GNSS spoofing detection method based on dynamic weighting according to claim 1, characterized in that, Extracting features of the normalized two-dimensional correlation function through the Transformer channel to obtain global features includes: S211: Extract features of the normalized two-dimensional correlation function through the multi-head self-attention mechanism to obtain multi-head attention features; S212: Retain the order of the normalized two-dimensional correlation function through sinusoidal position encoding to obtain position encoding information; S213: Add the multi-head attention features and the position encoding information to obtain global features.

4. A multi-modal GNSS spoofing detection method based on dynamic weighting according to claim 1, characterized in that, Extracting features of the normalized two-dimensional correlation function through the CNN channel to obtain local features includes: S221: Construct a residual block, where the residual block includes a 3×3 convolutional layer, a batch normalization layer, and an activation function layer; S222: Extract the local features of the normalized two-dimensional correlation function through the residual block.

5. A multi-modal GNSS spoofing detection method based on dynamic weighting according to claim 1, characterized in that Step S3 includes: S31: Convert the normalized one-dimensional SQM time series into a high-dimensional feature vector through a one-dimensional convolutional layer; S32: Capture the temporal dynamic features of the high-dimensional feature vector through LSTM to obtain a dynamic weight matrix, and adjust the dimension of the dynamic weight matrix through bilinear interpolation to obtain the Transformer channel dynamic weight matrix; S33: Extract the channel features of the high-dimensional feature vector through a squeeze-and-excitation network to obtain the CNN channel dynamic weight matrix, and the calculation expression is: Among them, is the CNN channel attention weight, is the non-linear activation function, is the second fully connected layer, is the operation sequence connection symbol, is the linear activation function, is the first fully connected layer, is the global average pooling function, is the high-dimensional feature vector.

6. A multi-modal GNSS spoofing detection method based on dynamic weighting according to claim 1, characterized in that In step S5, Perform global average pooling on the fused features to obtain a fused feature vector, Among them, is the fused feature vector, is the value of the fused feature at the th time step, the th row, and the th column; H is the length and W is the width. Map the fused feature vector into a spoofing probability through a fully connected network; wherein, is the deception probability, is the non-linear activation function, is the first weight matrix, is the second weight matrix, is the linear activation function, is the first bias term, is the second bias term, is the transpose of the matrix.

7. A multi-modal GNSS spoofing detection method based on dynamic weighting according to claim 1, characterized in that In step S6, If the spoofing probability is greater than or equal to the probability threshold, the current signal is a spoofing signal; If the spoofing probability is less than the probability threshold, the current signal is a genuine signal.

8. A multi-modal GNSS spoofing detection method based on dynamic weighting according to claim 7, characterized in that Perform a moving average on the spoofing probabilities of consecutive time steps through a time series smoothing method to obtain a smoothed spoofing probability; If multiple consecutive smoothed spoofing probabilities are greater than or equal to the probability threshold, trigger an alarm, and the current signal is a spoofing signal; otherwise, maintain the current judgment result.

9. A multi-modal GNSS spoofing detection method based on dynamic weighting according to claim 7, characterized in that The probability threshold selection conditions include false alarm rate, miss rate, and environmental parameters; Balance the false alarm rate and the miss rate through cross-validation or ROC curve to obtain the optimal probability threshold for the false alarm rate and the miss rate, Dynamically select the probability threshold through linear regression or gating mechanism according to the optimal probability threshold of the false alarm rate and the miss rate and the environmental parameters.

10. A multi-modal GNSS spoofing detection method based on dynamic weighting according to claim 9, where the false alarm rate is the probability that a genuine signal is misjudged as a spoofing signal, and the miss rate is the probability that a spoofing signal is misjudged as a genuine signal.

Images