A multi-feature satellite navigation signal deception jamming detection method based on an autoencoder

Abstract

The application discloses a kind of multi-feature satellite navigation signal deception interference detection methods based on auto-encoder, belong to satellite navigation and signal processing technical field.This method is by extracting four key characteristic values such as carrier-to-noise ratio, in-phase component, quadrature component, doppler shift, carries out Z-score standardization processing to received satellite navigation signal, eliminates dimension difference;Normal signal mode is trained and predicted in auto-encoder model, and whether signal is abnormal is judged by comparing reconstruction error with set threshold value.The application does not need additional hardware support, is suitable for resource limited equipment, has efficient, accurate detection capability in unmanned aerial vehicle navigation, vehicle navigation and military communication etc.Scenarios, with high flexibility, low cost, wide applicability advantages, especially in detecting complex deception signal, the application has significant advantages.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of satellite navigation and signal processing, and particularly relates to a multi-feature satellite navigation signal anti-deception interference detection method based on an autoencoder. BACKGROUND

[0002] As a core technology of modern communication, transportation, emergency management and military strategy, the global navigation satellite system (GNSS) has been deeply integrated into people's daily life and key industries. The positioning, navigation and timing services provided by the system have been widely used in transportation, finance, military and other fields. However, as the dependence on GNSS increases, the signal security problem has gradually become prominent, especially when encountering deceptive jamming, GNSS signals are easily affected. This kind of interference confuses the receiver by generating false signals or repeating real signals, making it produce false positioning, speed and time judgment, leading to the deviation of user or system behavior from normal, and even causing traffic accidents, financial system chaos, and even military decision-making errors. Therefore, improving the anti-deception interference ability of GNSS system has become an urgent problem to be solved in the industry.

[0003] The existing GNSS anti-deception detection method mainly relies on the matching check of Doppler positioning and pseudo-range positioning or digital signature encryption technology. For example, by calculating the Euclidean distance between the Doppler positioning and pseudo-range positioning results and combining the generalized likelihood ratio test, it can be judged whether there is deceptive signal. This kind of method is suitable for detecting typical pseudo-range deception attacks, but its effect is limited when dealing with other complex types of attacks, and the setting of accurate parameters is required, which cannot be widely applied to cost-sensitive civilian equipment. In addition, some methods use public key encryption to sign navigation data and combine statistical hypothesis testing to verify the reasonableness of signal arrival time. Although this method can enhance signal security, the requirements for key management, hardware performance and computing resources are too high, and it is difficult to promote, especially not suitable for low-power and lightweight devices.

[0004] Current anti-interference technologies mostly use signal processing or encryption technology, which can alleviate signal interference to some extent, but their flexibility and applicability are limited, especially when facing diversified and increasingly complex interference environment, their performance is slightly insufficient. Therefore, it is particularly necessary to develop an efficient and flexible GNSS anti-deception interference detection method for application on low-cost and resource-constrained devices to improve the anti-interference ability of navigation system in various environments and ensure the safe operation of key systems. SUMMARY

[0005] To address the limitations of existing spoofing signal detection methods, such as high complexity, strong parameter dependence, and high resource requirements, this invention provides a highly efficient anti-spoofing interference detection method for satellite navigation signals. This invention employs a multi-feature fusion detection approach, utilizing an autoencoder model to jointly analyze the main features of satellite navigation signals, thereby improving the flexibility and applicability of spoofing detection.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] A method for detecting anti-spoofing interference of multi-feature satellite navigation signals based on autoencoders includes the following steps:

[0008] Step S1: Extract key features from the received satellite navigation signal: carrier-to-noise ratio, in-phase component, quadrature component, and Doppler shift;

[0009] The main feature values, such as carrier-to-noise ratio, in-phase component, quadrature component, and Doppler shift, are extracted from the received satellite navigation signal. These feature values ​​contain the phase, frequency, and intensity information of the navigation signal, accurately reflecting the signal's changes during reception. Compared to single feature analysis, this invention improves the comprehensiveness and accuracy of detection through multi-feature fusion.

[0010] Step S2: Standardize the extracted main feature values ​​to eliminate dimensional differences;

[0011] The extracted feature values ​​are Z-score standardized to eliminate dimensional differences, ensuring that the feature values ​​have a mean of 0 and a standard deviation of 1. Standardization reduces the bias caused by different data scales among the features, ensuring data consistency during model training.

[0012] Step S3: Train and predict the standardized main feature values ​​based on the autoencoder model;

[0013] The standardized feature values ​​are input into the autoencoder model for training, and the encoder and decoder structures of the autoencoder are optimized. The encoder consists of two layers of neurons, with 3 and 2 neurons respectively, both using the hyperbolic tangent (tanh) activation function to compress the input signal into a low-dimensional space. The decoder then restores the low-dimensional encoding to reconstruct an output that approximates the original features.

[0014] Step S4: Calculate the reconstruction error of the autoencoder model, set a threshold for the reconstruction error, and determine whether the signal is abnormal based on the comparison result between the reconstruction error and the threshold. When the reconstruction error exceeds the threshold, it is determined that the received signal may be subject to deception interference.

[0015] The reconstruction error of the input signal is calculated and compared with a preset threshold. When the reconstruction error is higher than the threshold, it is determined that the signal may be abnormal. The reconstruction error calculation of the present application uses the mean square error formula, which ensures the sensitivity of the model when detecting signal deviation. In addition, the present application sets error thresholds in different scenarios to adapt to various application scenarios (such as unmanned aerial vehicles, vehicle navigation systems, etc.).

[0016] Preferably, the satellite navigation signal data in step S1 is based on the TEXBAT dataset for feature extraction, wherein the TEXBAT dataset contains a plurality of real and simulated interference scenarios for training the autoencoder model and optimizing the threshold setting.

[0017] Preferably, the standardization of the feature data in step S2 uses the Z-score standardization method, which includes the following steps:

[0018] The average value μ of the feature data is calculated as follows:

[0019]

[0020] where f i represents the i-th original feature value, n is the total number of feature data sets, μ is the mean value of the feature data, which reflects the central tendency of the feature data and reduces the bias;

[0021] The standard deviation σ of the feature data is calculated as follows:

[0022]

[0023] where σ is the standard deviation of the feature data, which measures the dispersion of the feature data relative to the mean value μ, ensuring the scale consistency between different feature values;

[0024] Each feature value is standardized to convert the original feature value f i to the standardized value f i ′, which is calculated as follows:

[0025]

[0026] where f i ′ is the i-th standardized feature value;

[0027] After the Z-score standardization, all feature data will conform to the standard normal distribution with a mean of 0 and a standard deviation of 1.

[0028] Preferably, the autoencoder model in step S3 compresses and reconstructs the input data through an encoder and a decoder, specifically including:

[0029] The encoder part compresses the input four feature values into a low-dimensional space, reducing the data dimension layer by layer through a double-layer encoder;

[0030] The first layer encoder is set with 3 neurons, and the second layer encoder is set with 2 neurons, and both layers of the encoder adopt the hyperbolic tangent tanh activation function;

[0031] The decoder part restores the low-dimensional data of the encoder to the original high-dimensional space, and reconstructs the output data consistent with the input data as much as possible.

[0032] Preferably, the encoder comprises two layers:

[0033] The first layer encoder receives an input vector x = [x1, x2, x3, x4], where x ∈ R 4 represents an input vector containing four feature values;

[0034] The first layer encoder calculates the hidden layer output h1 ∈ R3 through a weight matrix W1 ∈ R 3×4 and a bias vector b1 ∈ R 3 .

[0035] The first layer encoder adopts the hyperbolic tangent activation function tanh, and the calculation formula is:

[0036] h1 = tanh(W1x + b1)

[0037] For each hidden neuron i = 1, 2, 3, the calculation process is as follows:

[0038]

[0039] where, is the weight of the weight matrix W1 from the input feature x j to the hidden neuron h 1i , b 1i is the bias term, which allows the model to still produce a non-zero output when the input is zero.

[0040] Preferably, the encoder comprises two layers:

[0041] The first layer encoder receives an input vector x = [x1, x2, x3, x4], where x ∈ R 4 represents an input vector containing four feature values;

[0042] The first layer encoder calculates the hidden layer output h1 ∈ R3 through a weight matrix W1 ∈ R 3×4 and a bias vector b1 ∈ R 3 .

[0043] The first layer encoder adopts the hyperbolic tangent activation function tanh, and the calculation formula is:

[0044] h1 = tanh(W1x + b1)

[0045] For each hidden neuron i = 1,2,3, the calculation process is as follows:

[0046]

[0047] where, is the weight in the weight matrix W1 from the input feature x j to the hidden neuron h 1i , b 1i is the bias term, allowing the model to still produce a non-zero output when the input is zero.

[0048] Preferably, the decoder part contains two layers:

[0049] The first layer decoder receives the 2-dimensional encoded representation h2 = [h 21 , h 22 ] T from the encoder.

[0050] The output vector h3 ∈ R 3×2 is calculated by the weight matrix W3 ∈ R 3 and the bias vector b3 ∈ R 3 ;

[0051] The activation function uses the sigmoid function, and the calculation formula is as follows:

[0052] h3 = sigmoid(W3h2 + b3)

[0053] For each hidden neuron i = 1,2,3, the specific calculation process is as follows:

[0054]

[0055] where, is the element in the weight matrix W3, representing the weight from the kth input node to the ith hidden neuron, b 3i is the corresponding bias term.

[0056] Preferably, the second layer of the decoder receives the output h3 of the first layer decoder and restores it to the 4-dimensional space:

[0057] The second layer decoder calculates the final reconstructed output vector by the weight matrix W4 ∈ R 4×3 and the bias vector b4 ∈ R 4 , so that it is close to the original input vector x;

[0058] The activation function uses the sigmoid function, and the calculation formula is as follows:

[0059]

[0060] For each output node j = 1, 2, 3, 4, the specific calculation process is as follows:

[0061]

[0062] wherein, is the reconstructed output value, is an element in the weight matrix W4, representing the weight from the hidden neuron h 3i to the output node , b 4j is the corresponding bias term.

[0063] Preferably, the step of calculating the reconstruction error and setting the threshold in the step S4 comprises:

[0064] S41: the input multi-dimensional feature signal is a feature vector y = [y1, y2, …, yn], wherein: y represents the original feature vector; n represents the dimension of the feature vector, i.e. the number of signal components; y n represents the nth component of the original signal; n

[0065] S42: the input signal is reconstructed using the trained autoencoder model to obtain a reconstructed signal vector wherein: represents the feature vector reconstructed by the autoencoder model; represents the ith component in the reconstructed signal vector, i.e. the ith signal component reconstructed by the autoencoder;

[0066] S43: the reconstruction error is expressed by the mean square error, and the calculation formula is as follows:

[0067]

[0068] wherein: MSE represents the mean square error, used to measure the difference between the reconstructed signal and the original signal y; n represents the dimension of the feature vector; and y i represent the ith component of the reconstructed signal and the original signal, respectively;

[0069] S44: based on the statistical characteristics of the data, the threshold T is set to 37, in the non-spoofing scenario, the value of MSE is low, usually below 20, and the maximum does not exceed 37; while in the spoofing scenario, the value of MSE increases significantly, with a maximum value of more than 1200, wherein: T represents the threshold of the mean square error, used to judge whether the signal is disturbed or not;

[0070] ​S45: comparing the calculated reconstruction error MSE with a set threshold T;

[0071] If MSE≤T, it is judged that the signal is normal.

[0072] If MSE>T, an anomaly is detected, and it is judged that the signal may be subject to spoofing interference.

[0073] Preferably, the method is suitable for interference detection in UAV navigation, vehicle navigation system and military communication, and specifically comprises:

[0074] In UAV navigation, multi-dimensional data of the UAV when normally operating are input into the autoencoder as a training data set for training, and the abnormality of the input data is judged based on the reconstruction error, and if the calculated reconstruction error exceeds the set threshold, it is judged that the UAV may have suffered a GPS spoofing attack.

[0075] In vehicle navigation system and military communication, the vehicle navigation system and military communication rely on the positioning, navigation and timing services provided by the global navigation satellite system (GNSS), and the main characteristic values used by both are general signal parameters.

[0076] Compared with the prior art, the present application has the beneficial effects that: the multi-parameter detection method based on the autoencoder of the present application does not require additional hardware support and is suitable for wide application. By extracting multiple key features such as carrier-to-noise ratio, in-phase component, quadrature component and Doppler shift from the received satellite navigation signal and performing Z-score standardization processing on these features, the same scale is obtained in the model, ensuring the detection stability and applicability of the model. Compared with the traditional detection method relying on a single characteristic value, the present application jointly analyzes multiple characteristic values through the autoencoder model, making up for the limitations of single parameter detection, thereby greatly improving the accuracy of detection. The autoencoder model adopts a two-layer encoder and two-layer decoder structure, which can effectively reconstruct the feature mode of the input data under the condition of no spoofing signal, and judge the signal anomaly through the reconstruction error. The adopted reconstruction error threshold is simple and has low calculation cost, and by setting the threshold to 37, when the reconstruction error exceeds this value, it is judged that the signal may be subject to spoofing interference. The detection method of the present application is suitable for various scenarios such as UAV navigation, vehicle navigation system and military communication, and can significantly improve the GNSS signal security in these fields and reduce potential safety hazards caused by spoofing interference. BRIEF DESCRIPTION OF DRAWINGS

[0077] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0078] Wherein:

[0079] Figure 1 The figure is a flow chart of the method of the embodiment of the present application.

[0080] Figure 2 The figure is a flow chart of the implementation process of the embodiment of the present application in the actual detection process.

[0081] Figure 3 The figure is a data flow chart of the feature value standardization process in the embodiment of the present application.

[0082] Figure 4 The figure is a structure diagram of the auto-encoder model in the embodiment of the present application.

[0083] Figure 5 The figure is a change curve diagram of the training loss in the training process of the auto-encoder in the embodiment of the present application.

[0084] Figure 6 The figure is a change curve diagram of the verification loss in the training process of the auto-encoder in the embodiment of the present application. DETAILED DESCRIPTION

[0085] In order to make the purpose, technical solutions and advantages of the embodiments of the present application more clear, the technical solutions of the embodiments of the present application will be described clearly and completely in combination with the drawings of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the described embodiments of the present application, all other embodiments obtained by those skilled in the art belong to the scope of protection of the present application.

[0086] Embodiment 1

[0087] As shown in Figure 1 , Figure 2 , the embodiment of the present application provides a multi-feature satellite navigation signal anti-deception interference detection method based on an auto-encoder, which comprises:

[0088] Step S1: extracting main feature values from the received satellite navigation signal: carrier-to-noise ratio, in-phase component, quadrature component, Doppler frequency shift;

[0089] Specifically, the satellite navigation signal data in step S1 is based on the TEXBAT dataset for feature extraction, wherein the TEXBAT dataset contains a plurality of real and simulated interference scenarios for training the autoencoder model and optimizing the threshold setting.

[0090] In the present application, four key feature values are first extracted from the received satellite navigation signal: carrier-to-noise ratio (C / N0), in-phase component (Prompt_i), quadrature component (Prompt_q), and Doppler shift (Carrier_Doppler_Hz).

[0091] Carrier-to-noise ratio (C / N0): This parameter represents the signal-to-noise ratio of the signal, which is a key indicator for GNSS receivers to evaluate signal quality, and can measure signal strength and noise interference level. Lower C / N0 may indicate that the signal is disturbed.

[0092] In-phase component (Prompt_i) and quadrature component (Prompt_q): These two parameters represent the amplitude information and phase information of the signal, respectively, for judging the integrity of the signal. Their changes can reflect the possible phase or amplitude anomalies of the signal, which are suitable for detecting deceptive interference.

[0093] Doppler shift (Carrier_Doppler_Hz): Reflects the frequency change caused by relative motion, widely used in positioning and speed measurement. Abnormal frequency shift may indicate that the signal is disturbed.

[0094] By monitoring the changes of these feature values, the method of the present application provides basic data for detecting complex deceptive interference.

[0095] Further, as shown in Figure 3 Step S2: Standardize the extracted main feature values to eliminate dimensional differences;

[0096] In step S2, the standardization of feature data uses the Z-score standardization method, which includes the following steps:

[0097] Calculate the mean value μ of the feature data, the formula is as follows:

[0098]

[0099] Where f i represents the i-th original feature value, n is the total number of feature data sets, μ is the mean value of the feature data, which reflects the central tendency of the feature data and reduces the bias;

[0100] Calculate the standard deviation σ of the feature data, the formula is as follows:

[0101]

[0102] Where σ is the standard deviation of the feature data, which is used to measure the dispersion of the feature data relative to the mean μ, and to ensure the scale consistency among different feature values;

[0103] Standardize each eigenvalue, and then standardize the original eigenvalue f. i Convert to the standardized value f i The formula is as follows:

[0104]

[0105] Among them, f i ′ represents the i-th eigenvalue after standardization;

[0106] After Z-score standardization, all feature data will conform to a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0107] To eliminate dimensional differences in feature values, the feature values ​​extracted in step S1 are Z-score standardized. Standardization ensures that all feature values ​​conform to a standard normal distribution with a mean of 0 and a standard deviation of 1. This process ensures scale consistency among different feature values, which helps the autoencoder model accurately learn the normal patterns of the signal and improves the adaptability of detection in complex environments.

[0108] Furthermore, such as Figure 4 As shown, step S3: train and predict the standardized main feature values ​​based on the autoencoder model;

[0109] Specifically, the autoencoder model in step S3 compresses and reconstructs the input data through an encoder and a decoder, including:

[0110] The encoder part compresses the four input feature values ​​into a low-dimensional space, reducing the data dimensionality layer by layer through a two-layer encoder;

[0111] The first encoder layer has 3 neurons, the second encoder layer has 2 neurons, and both encoder layers use the hyperbolic tangent tanh activation function;

[0112] The decoder part restores the low-dimensional data from the encoder back to the original high-dimensional space, reconstructing output data that is as consistent as possible with the input data.

[0113] The encoder consists of two layers:

[0114] The first layer encoder receives the input vector x = [x1, x2, x3, x4], where x ∈ R. 4 This represents an input vector containing four eigenvalues;

[0115] The first encoder uses a weight matrix W1∈R 3×4 and bias vector b1∈R 3 Calculate the hidden layer output h1∈R3;

[0116] The first-layer encoder uses the hyperbolic tangent activation function tanh, calculated as follows:

[0117] h1 = tanh(W1x + b1)

[0118] For each hidden neuron i = 1, 2, 3, the calculation process is as follows:

[0119]

[0120] in, It is the weight matrix W1 derived from the input feature x j To the hidden neuron h 1i The weight, b 1i It is a bias term that allows the model to produce a non-zero output even when the input is zero.

[0121] Specifically, the second layer of the encoder receives the output h1 from the first layer and further compresses the data into a 2D space:

[0122] The second encoder uses the weight matrix W2∈R 2×3 and bias vector b2∈R 2 Calculate the output vector h2∈R 2 ;

[0123] The activation function is also the hyperbolic tangent tanh, and the calculation formula is as follows:

[0124] h2 = tanh(W2h1 + b2)

[0125] Specifically, for each hidden neuron k=1,2, the calculation process is as follows:

[0126]

[0127] in, These are elements in the weight matrix W2, representing the number of hidden neurons h from the first layer. 1i To the second layer of hidden neurons h 2k The weight, b 2k This is the corresponding bias term.

[0128] The decoder section consists of two layers:

[0129] The first-layer decoder receives the 2D encoded representation h2 = [h] from the encoder. 21 ,h 22 ] T ;

[0130] The output vector h3 is calculated by the weight matrix W3 and the bias vector b3. 3×2 3 The output vector h3 is calculated by the weight matrix W3 and the bias vector b3. 3 ;

[0131] The activation function adopts a sigmoid function, and the calculation formula is as follows:

[0132] h3 = sigmoid(W3h2 + b3)

[0133] For each hidden neuron i = 1, 2, 3, the specific calculation process is as follows:

[0134]

[0135] wherein, is an element in the weight matrix W3, representing the weight from the kth input node to the ith hidden neuron, b 3i is the corresponding bias term.

[0136] Specifically, the second layer of the decoder receives the output h3 of the first layer of the decoder and restores it to a 4-dimensional space:

[0137] The second layer of the decoder calculates the final reconstruction output vector h4 by the weight matrix W4 and the bias vector b4. 4×3 4 so as to approach the original input vector x.

[0138] The activation function adopts a sigmoid function, and the calculation formula is as follows:

[0139]

[0140] For each output node j = 1, 2, 3, 4, the specific calculation process is as follows:

[0141]

[0142] wherein, is the reconstructed output value, is an element in the weight matrix W4, representing the weight from the hidden neuron h 3i to the output node , b 4j is the corresponding bias term.

[0143] The application adopts an autoencoder model to train the normalized feature values, and establishes a normal mode of the navigation signal.

[0144] ​​​Encoder: The input layer consists of four neurons, corresponding to the four normalized feature values. The encoder contains two layers, the first layer compresses the features into an intermediate representation of 3 neurons, and the second layer further compresses it to 2 neurons, with the activation function being hyperbolic tangent (tanh). In this way, the model can effectively reduce the data dimension and extract the main features of the signal.

[0145] Decoder: The encoded data is gradually restored to the original 4-dimensional space, and the decoder outputs reconstructed data as close as possible to the input by trying to reconstruct the input data.

[0146] Training and prediction: The autoencoder model is trained on normal data to learn the reconstruction pattern of normal signals. When detecting fraudulent signals, the test data is input into the model, and the reconstructed signal data is output.

[0147] As shown in Figure 5 , the change curve of the training loss during the training process of the autoencoder model is shown. The training loss decreases rapidly within the first 20 training cycles (epochs), indicating that the model quickly learns the basic features of the data in the initial stage. Subsequently, the training loss gradually stabilizes, and eventually stabilizes around 50 training cycles (epochs), indicating that the learning process of the model has achieved good reconstruction results.

[0148] Further, step S4: calculate the reconstruction error of the autoencoder model, set a threshold for the reconstruction error, and determine whether the signal is abnormal according to the comparison result of the reconstruction error and the threshold, wherein when the reconstruction error exceeds the threshold, it is determined that the received signal may be subject to fraudulent interference.

[0149] Specifically, the steps of calculating the reconstruction error and setting the threshold in step S4 include:

[0150] S41: The input multi-dimensional feature signal is a feature vector y = [y1, y2,..., yn], where: y represents the original feature vector; n represents the dimension of the feature vector, i.e. the number of signal components; y n represents the nth component of the original signal; n

[0151] S42: Use the trained autoencoder model to reconstruct the input signal to obtain the reconstructed signal vector where: represents the feature vector reconstructed by the autoencoder model; represents the ith component in the reconstructed signal vector, i.e. the ith signal component reconstructed by the autoencoder;

[0152] S43: The reconstruction error is represented by the mean square error, and the calculation formula is as follows:

[0153] ​

[0154] Wherein: MSE represents the mean square error, used to measure the difference between the reconstructed signal and the original signal y; n represents the feature

[0155] dimension of the vector; and y i respectively represent the i-th component of the reconstructed signal and the original signal;

[0156] S44: Based on the statistical characteristics of the data, the threshold T is set to 37, and in the non-fraud scene, the value of MSE is low, usually below 20, and the maximum does not exceed 37; while in the fraud scene, the value of MSE is significantly increased, and the maximum value can reach more than 1200, wherein: T represents the threshold of the mean square error, used to judge whether the signal is disturbed;

[0157] S45: The calculated reconstruction error MSE is compared with the set threshold T:

[0158] If MSE≤T, it is judged that the signal is normal;

[0159] If MSE>T, an anomaly is detected, and it is judged that the signal may be disturbed by fraud.

[0160] As Figure 6 shown in the figure, the change curve of the validation loss in the training process of the autoencoder is shown, and the validation loss curve also presents a similar change trend as the training loss, which rapidly decreases at the beginning, and then tends to be stable, and maintains a close level with the training loss, which indicates that the autoencoder model performs well on the validation set, and there is no obvious overfitting phenomenon.

[0161] Further, the method is applicable to interference detection in unmanned aerial vehicle navigation, vehicle navigation system and military communication, specifically including:

[0162] In the unmanned aerial vehicle navigation, the multi-dimensional data of the normal operation of the unmanned aerial vehicle is input into the autoencoder as the training data set for training, and the abnormal situation of the input data is judged based on the reconstruction error, and if the calculated reconstruction error exceeds the set threshold, it is judged that the unmanned aerial vehicle may have suffered GPS spoofing attack;

[0163] In the vehicle navigation system and military communication, since the vehicle navigation system and military communication rely on the positioning, navigation and timing services provided by the global navigation satellite system GNSS, and the main characteristic values used by both are general signal parameters.

[0164] The advantages of the embodiment are:

[0165] The application is suitable for multiple scenes such as unmanned aerial vehicle navigation, vehicle navigation system and military communication, and can still work stably on resource-limited devices, thereby improving the flexibility of GNSS spoofing detection.

[0166] The low-dimensional encoding and high-dimensional reconstruction characteristics of the autoencoder model make the model have a higher matching degree for normal signals, so that when a spoofing signal enters, the model will produce a significant reconstruction error, thereby greatly improving the detection accuracy of the spoofing signal.

[0167] Compared with encryption technology, the application does not rely on key management and high-performance hardware devices, and can be implemented on low-power devices, thereby saving the cost and resource consumption in the application scene.

[0168] The application can resist multiple types of interference signals through the multi-feature fusion mode, and can still stably detect spoofing signals in complex environments, thereby significantly improving the anti-interference ability of GNSS signals.

[0169] By adjusting the parameters of the feature extraction module and the threshold setting of the autoencoder model, the method can adapt to the interference detection requirements in different scenes, and has wide application potential in military communication and low-power navigation devices.

[0170] Embodiment 2

[0171] In this embodiment, the satellite navigation signal anti-spoofing interference detection method of the application is tested in multiple scenes based on the TEXBAT dataset. The TEXBAT dataset contains multiple simulated and real interference scenes, including scenes 1, 2, 4 and 8, which represent different types of spoofing interference and are suitable for evaluating the applicability and accuracy of the model in complex environments.

[0172] Experimental steps Data preparation: select the satellite navigation signal data of scenes 1, 2, 4 and 8 from the TEXBAT dataset. The spoofing signals of these scenes contain different characteristic value changes to simulate spoofing interference in real environments.

[0173] Feature extraction: for the signals in each scene, four main characteristic values are extracted: carrier-to-noise ratio (C / N0), in-phase component (Prompt_i), quadrature component (Prompt_q) and Doppler frequency shift (Carrier_Doppler_Hz). These features reflect the changes in signal strength, phase and frequency, and are used to detect whether the signal is interfered.

[0174] Data standardization: the extracted characteristic values are standardized by Z-score to eliminate the dimensional difference and ensure that each characteristic value has the same scale, thereby improving the detection accuracy of the model.

[0175] Autoencoder model training: The standardized data is input into the autoencoder model for training. The model contains two layers of encoder and two layers of decoder. The encoder compresses the input data into a low-dimensional space, while the decoder restores it. The reconstruction error between the input and output is calculated to determine the abnormality of the signal.

[0176] Reconstruction error calculation and threshold setting: After the model reconstructs the input signal, the reconstruction error (mean square error, MSE) is calculated, and the threshold T is set to 37. If the MSE exceeds the threshold, it is determined that the signal may be subject to fraudulent interference.

[0177] Experimental results

[0178] In scenarios 1, 2, 4 and 8, the detection results of the model are shown in the following table:

[0179] Scenario Total number of channels Number of alarm triggers Alarm rate Remark 1 8 August 8 100％ Detection accuracy 2 8 August 8 100％ Detection accuracy 3 8 June 8 75％ Partial missed detection 4 8 August 8 100％ Detection accuracy 8 8 August 8 100％ Detection accuracy

[0180] In scenarios 1, 2, 4 and 8, the method of the present application can detect 8 channels of fraudulent signals with 100% accuracy, showing high sensitivity and accuracy of the detection model. In scenario 3, the model detects 6 channels of fraudulent signals, and some channels fail to trigger an alarm, with a detection rate of 75%, slightly lower than other scenarios. Overall, the model shows high detection accuracy in most scenarios.

[0181] Analysis of experimental results: The 100% detection rate in scenarios 1, 2, 4 and 8 proves the effectiveness of the autoencoder model in multi-feature joint analysis, which can accurately judge various types of fraudulent signals.

[0182] Adaptability and robustness: The model can maintain high detection accuracy in multiple scenarios, suitable for complex navigation signal interference environments, especially for unmanned aerial vehicles, vehicle navigation systems and military communications.

[0183] This embodiment verifies the applicability and efficiency of the method of the present application in different scenarios, providing strong data support for practical applications.

[0184] To sum up, the application significantly improves the detection ability of complex interference by jointly analyzing the multi-parameter characteristic values through the auto-encoder model, and has high creativity. The method combines the four key features of carrier-to-noise ratio, in-phase component, quadrature component and Doppler shift together for the first time in spoofing detection, overcomes the defect that the single parameter detection of the traditional method is easily limited, and establishes the normal feature mode of the signal through the low-dimensional coding and high-dimensional reconstruction process of the auto-encoder model. In particular, the application uses the Z-score standardization method to make the dimensions of different features consistent, effectively improving the adaptability and detection accuracy of the model in various environments. In addition, the reconstruction error threshold of the application is set to 37, which not only makes the detection judgment more rapid and simple, but also ensures the reliability of the detection result by statistically setting the signal reconstruction error in different scenes. Without additional hardware support, the detection process is suitable for deployment in low-cost devices with limited resources, and can effectively adapt to the actual application requirements in the fields of unmanned aerial vehicle navigation, vehicle navigation system and military communication. Compared with the existing encryption or complex hardware method, the application has the significant advantages of low cost, convenient deployment and sensitive detection, has high application and promotion potential, fully meets the high security requirements of various navigation systems, and shows unique technical innovation and good authorization prospects.

[0185] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" and the like 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 present application. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or features of the embodiments or examples described in the specification without contradiction.

[0186] Any process or method descriptions in flow charts or described elsewhere herein can be understood as representing code modules, segments, or portions of code that include one or more executable instructions for implementing specific logic functions or other processes. And the scope of preferred embodiments of the present application includes additional implementation in which the functions described in the illustrated or discussed order are not necessarily performed in that order. For example, functions described sequentially can be performed at least partially concurrently or in reverse order.

[0187] The above description is merely one specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for spoofing detection of multi-features GNSS signals based on autoencoder, characterized in that, The method comprises the following steps: Step S1: extracting main characteristic values from the received satellite navigation signal: carrier-to-noise ratio, in-phase component, quadrature component, Doppler shift; Step S2: standardizing the extracted main characteristic values to eliminate dimensional differences; Step S3: training and predicting the standardized main characteristic values based on an autoencoder model; Step S4: calculating the reconstruction error of the autoencoder model, setting a threshold value for the reconstruction error, and determining whether the signal is abnormal according to the comparison result of the reconstruction error and the threshold value, wherein when the reconstruction error exceeds the threshold value, it is determined that the received signal may be subject to spoofing interference.

2. The self-encoder based multi-feature GNSS spoofing detection method according to claim 1, characterized in that, The satellite navigation signal data in step S1 is based on the TEXBAT dataset for feature extraction, wherein the TEXBAT dataset contains multiple real and simulated interference scenarios for training the autoencoder model and optimizing the threshold setting.

3. The self-encoder based multi-feature GNSS spoofing detection method according to claim 1, characterized in that, The standardization of feature data in step S2 adopts the Z-score standardization method, which includes the following steps: Calculate the average value μ of the feature data, as follows: wherein f i represents the ith original feature value, n is the total number of the feature dataset, μ is the mean of the feature data, used to reflect the central tendency of the feature data, and reduce the bias; Calculate the standard deviation σ of the feature data, as follows: Where σ is the standard deviation of the feature data, used to measure the dispersion of the feature data relative to the mean μ, ensuring the scale consistency between different characteristic values; The original feature value f i is converted into a normalized value f i ′, according to the following formula: wherein f i is the standardized i-th feature value; After Z-score standardization, all feature data will conform to the standard normal distribution, with a mean of 0 and a standard deviation of 1.

4. The self-encoder based multi-feature GNSS anti-spoofing jamming detection method according to claim 1, characterized in that, The autoencoder model in step S3 compresses and reconstructs the input data through an encoder and a decoder, specifically including: The encoder part compresses the input four characteristic values to a low-dimensional space, reducing the data dimension layer by layer through a double-layer encoder; The first layer encoder sets 3 neurons, the second layer encoder sets 2 neurons, and both layers of the encoder use the hyperbolic tangent tanh activation function; The decoder part restores the low-dimensional data of the encoder to the original high-dimensional space, reconstructing output data as consistent as possible with the input data.

5. The self-encoder based multi-feature GNSS spoofing detection method according to claim 4, characterized in that, The encoder contains two layers: The first layer encoder receives an input vector x = [x1, x2, x3, x4], where x e R 4 represents an input vector containing four feature values; the first layer encoder computes a hidden layer output h1 e R3 through a weight matrix W1 e R 3×4 and a bias vector b1 e R 3 and a bias vector b1 e R The first layer encoder uses the hyperbolic tangent activation function tanh, with the calculation formula as follows: h1=tanh(W1x+b1) For each hidden neuron i=1,2,3, the calculation process is as follows: where, is the weight of the input feature x j to the hidden neuron h 1i in the weight matrix W1, b 1i is the bias term, allowing the model to still produce a non-zero output if the input is zero.

6. The self-encoder based multi-feature GNSS anti-spoofing jamming detection method according to claim 5, characterized in that, The second layer of the encoder receives the output h1 of the first layer and further compresses the data to a 2-dimensional space: The second layer encoder computes an output vector h2 e R 2×3 and a bias vector b2 e R 2 from an input vector x2 e R 2 ; The activation function also uses the hyperbolic tangent tanh, with the calculation formula as follows: h2=tanh(W2h1+b2) Specifically, for each hidden neuron k=1,2, the calculation process is as follows: wherein, is an element in the weight matrix W2representing the weight from the first layer hidden neuron h 1i to the second layer hidden neuron h 2k b 2k is the corresponding bias term.

7. The self-encoder based multi-feature GNSS anti-spoofing jamming detection method according to claim 4, characterized in that, The decoder part contains two layers: The first layer decoder receives the 2-dimensional encoded representation h2= [h 21 ,h 22 ] T from the encoder. The output vector h3 e R 3×2 is calculated by the weight matrix W3 e R 3 and the bias vector b3 e R 3 ; The activation function uses the sigmoid function, with the calculation formula as follows: h3=sigmoid(W3h2+b3) For each hidden neuron i=1,2,3, the specific calculation process is as follows: wherein, are elements in the weight matrix W3representing the weight from the kth input node to the ith hidden neuron, b 3i is the corresponding bias term.

8. The self-encoder based multi-feature GNSS anti-spoofing jamming detection method according to claim 7, characterized in that, The second layer of the decoder receives the output h3 of the first layer of the decoder and restores it to a 4-dimensional space: The second layer decoder computes the final reconstructed output vector 4×3 y4 = W4x + b4 4 y4 = W4x + b4 such that it approximates the original input vector x; The activation function uses the sigmoid function, with the calculation formula as follows: For each output node j=1,2,3,4, the specific calculation process is as follows: wherein, is the reconstructed output value, is an element in the weight matrix W4, representing the weight from the hidden neuron h 3i to the output node b 4j is the corresponding bias term.

9. The self-encoder based multi-feature GNSS anti-spoofing jamming detection method according to claim 1, characterized in that, The steps of calculating the reconstruction error and setting the threshold value in step S4 include: S41: the input multi-dimensional feature signal is a feature vector y = [y1, y2,..., yn]T, wherein: y represents an original feature vector; n represents the dimension of the feature vector, i.e., the number of components of the signal; yn represents the nth component of the original signal. n ], wherein: y represents an original feature vector; n represents the dimension of the feature vector, i.e., the number of components of the signal; yn represents the nth component of the original signal. n ], wherein: y represents an original feature vector; n represents the dimension of the feature vector, i.e., the number of components of the signal; yn represents the nth component of the original signal. S42: reconstruct the input signal using the trained autoencoder model to obtain a reconstructed signal vector wherein: represents the feature vector reconstructed by the autoencoder model; represents the i-th component in the reconstructed signal vector, i.e., the i-th signal component reconstructed by the autoencoder S43: The reconstruction error is expressed by mean square error, and the calculation formula is as follows: where MSE denotes the mean square error, which measures the difference between the reconstructed signal and the original signal y; n denotes the dimension of the feature vector; and y i denote the i-th component of the reconstructed signal and the original signal, respectively. S44: Based on the statistical characteristics of the data, the threshold T is set to 37, in the non-deception scene, the value of MSE is low, below 20, and the maximum does not exceed 37; while in the deception scene, the value of MSE increases significantly, the maximum value can reach more than 1200, wherein: T represents the threshold of mean square error, which is used to judge whether the signal is disturbed or not; S45: The calculated reconstruction error MSE is compared with the set threshold T: If MSE≤T, it is judged that the signal is normal; If MSE>T, an anomaly is detected, and it is judged that the signal may be subject to deception interference.

10. The self-encoder based multi-feature GNSS spoofing detection method according to claim 1, characterized in that, The method is suitable for interference detection in unmanned aerial vehicle navigation, vehicle navigation system and military communication, specifically including: in unmanned aerial vehicle navigation, the multi-dimensional data of the unmanned aerial vehicle in normal operation is input into the self-encoder as a training data set for training, and the abnormality of the input data is judged based on the reconstruction error, if the calculated reconstruction error exceeds the set threshold, it is judged that the unmanned aerial vehicle may be subject to GPS deception attack; In vehicle navigation system and military communication, vehicle navigation system and military communication rely on global navigation satellite system (GNSS) to provide positioning, navigation and timing services, and the main characteristic values used by both are general signal parameters.

Images