Multi-task spiking neural network assisted satellite positioning deception defense method and system

Abstract

The invention provides a multi-task pulse neural network assisted satellite positioning deception defense method and system, and belongs to the technical field of satellite navigation positioning. And processing satellite intermediate frequency signal data acquired in real time by using a satellite positioning receiver embedded with a pre-trained multi-task pulse neural network model to obtain a carrier credible operation state. According to the method, a unified optimization architecture of multi-task collaborative modeling is developed by adopting a pulse neural network model, so that joint optimization of deception detection, attack situation identification and deception signal parameter estimation is realized; an adaptive state solver is further constructed to restore the real operation state of the satellite positioning receiver in combination with specific output of the model, satellite measurement after real / deception elimination and auxiliary sensor measurement, the application potential of a neural network in a satellite positioning deception attack'detection-suppression 'integrated path is expanded, a deception defense closed loop is realized, and the method is suitable for a deception defense system. And the safety credibility of the satellite navigation system position service is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of satellite navigation and positioning technology, and in particular to a multi-task pulse neural network-assisted satellite positioning deception defense method and system. Background Art

[0002] Global Navigation Satellite System (GNSS) spoofing attacks, which broadcast erroneous, forged signals to manipulate nearby receivers into calculating false positioning information, pose a serious cybersecurity threat to safety-related applications that rely on satellite positioning for autonomous driving and train safety control. To address this threat, effective active defenses are urgently needed, including the construction of dynamic virtual protection barriers to effectively identify, isolate, and mitigate spoofing effects, thereby ensuring the security and reliability of GNSS positioning services. When spoofed satellite positioning signals are mixed with genuine signals and enter a victim receiver, the receiver simultaneously captures and tracks both signals. In the signal tracking loop, the mixed signal is correlated and phase-detected with a locally generated carrier signal. The spoofed signal significantly alters the autocorrelation function, causing the zero crossing of the phase detection curve to shift, making it impossible for the locally generated signal to accurately synchronize with the pseudo-code phase and carrier phase of the genuine signal. The resulting phase mismatch propagates errors throughout the positioning process, ultimately generating spoofed, false positioning information. Spoofing signals with strong active intrusion intentions are more likely to overwrite real signals and be captured by general receivers that do not have anti-spoofing functions. Therefore, it is necessary to introduce active defense mechanisms into the receiver's signal processing logic to improve its anti-spoofing capabilities and ensure the quality of GNSS positioning services.

[0003] GNSS anti-spoofing and spoof generation form a dynamic game of mutual constraints, with one constraining the other and driving the evolution of countermeasures. As spoof generation technology continues to evolve, traditional anti-spoofing methods are gradually exposing their limitations in adaptability and generalization. Simultaneously, anti-spoofing technology itself is also evolving, evolving from detection mechanisms focused on identification and early warning to defense mechanisms that actively suppress and even eliminate the effects of spoofing, reflecting a paradigm shift from passive response to active countermeasures. A single defense mechanism is often unable to fully address diverse and intelligent spoofing threats. In recent years, new data-driven approaches have been widely applied in fields such as wireless communications and navigation and positioning. Data-driven approaches, particularly neural network models, offer powerful expressive capabilities, flexible modeling strategies, and excellent scalability. As a robust nonlinear filter, they use processing units (neurons) to automatically extract features and model complex nonlinear mappings. Neural networks not only accurately estimate data through training and learning from existing data but also predict and extend these results using new data, demonstrating strong adaptability. With the introduction of new data-driven methods, the coverage and countermeasure capabilities of GNSS anti-spoofing technology have been significantly expanded, giving rise to a large number of adversarial protection measures targeting different attack modes.

[0004] GNSS spoofing detection, as the first step in identifying spoofing, has a relatively clear technical path. Detection mechanisms based on neural network methods have been widely adopted, including support vector machines, perceptrons, autoencoders, convolutional neural networks, generative adversarial networks, long short-term memory neural networks, recurrent neural networks, deep reinforcement learning, graph neural networks, and transformers. These data-driven models leverage key features extracted during signal processing and, through a classification / reconstruction / prediction approach, learn the mapping between these features and spoofing attack behaviors, thereby enabling efficient identification and discrimination of spoofing attacks. While neural network models demonstrate excellent recognition accuracy in spoofing detection tasks, research on interference suppression and performance recovery against spoofing attacks is still in its infancy. Some studies have introduced neural network models into the GNSS intermediate frequency signal processing stage to reconstruct the true signal or related function sequences, thereby improving the availability and accuracy of the original signal. Furthermore, in the satellite measurement processing and positioning solution stages, some studies have attempted to use neural networks to predict and compensate for spoofing-induced errors in key indicators (such as pseudorange, carrier phase, receiver clock error, and position increment), thereby directly recovering the true positioning solution. However, existing neural network-based GNSS anti-spoofing technologies still face numerous limitations, including high-performance requirements for spoofing suppression, split modeling for spoofing detection and suppression, and high energy consumption. Spiking Neural Networks (SNNs), a third-generation artificial intelligence approach, are inspired by the brain's powerful and energy-efficient information processing capabilities. They use binary pulses to transmit information. Their low power consumption, inherent noise suppression capabilities, and enhanced dynamic robustness have garnered attention in areas such as image classification, target detection and tracking, and speech recognition. SNN models offer new insights and technical approaches for building GNSS anti-spoofing systems suitable for low-resource, high-real-time satellite positioning environments. Summary of the Invention

[0005] The purpose of the present invention is to provide a multi-task pulse neural network assisted satellite positioning deception defense method and system to solve at least one technical problem existing in the above-mentioned background technology.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a multi-task pulse neural network assisted satellite positioning deception defense method, comprising:

[0008] A satellite positioning receiver with a built-in deception countermeasure module is used to extract and process binary pulse deception features from the collected satellite intermediate frequency signal data, which are used for offline independent training of a multi-task pulse neural network model, and store and export the optimal model parameters as a target format that can be called by the satellite positioning receiver; the binary pulse processing deception features are used to integrate continuous real-valued features into an adversarial pulse feature image; a satellite positioning receiver embedded with a pre-trained multi-task pulse neural network model is used to process the satellite intermediate frequency signal data acquired in real time to obtain a trusted operating state of the carrier; the multi-task pulse neural network model includes a multi-task shared backbone network, a neck network and a head network; the multi-task shared backbone network is used to extract relevant semantic information from the deception features; the neck network is used to interactively fuse and enhance the semantic information extracted by the shared backbone network; the head network includes two classification heads and one regression head; the two classification heads are used for deception detection and attack situation awareness tasks respectively, and the regression head is used for deception estimation tasks.

[0009] As a further limitation of the first aspect of the present invention, the multi-task shared backbone network includes a feature downsampling module based on a pulse neuron-convolution layer-pooling layer structure, a pulse residual module based on a void convolution-depth separable convolution, and a pulse attention module; the downsampling module is used to compress the size of the adversarial pulse feature image and expand the number of channels; the pulse residual module is used to extract features from the adversarial pulse feature image; the pulse attention module is used to enhance the model's ability to capture key features and suppress irrelevant information, thereby improving the accuracy of deception recognition.

[0010] As a further limitation of the first aspect of the present invention, in the pulse residual module, its input and output pulse feature maps have the same dimension, its void convolution is used for multi-scale feature extraction and expanded receptive field, so that the model can capture a wider range of contextual information, the depthwise separable convolution is used to reduce the number of parameters and computational complexity of the model, keep the model lightweight, and the residual structure is used to alleviate the gradient vanishing problem that may exist in deep networks; several downsampling modules and pulse residual modules are alternately connected in series to realize high-level semantic information extraction; a pulse attention module is connected in series after the last pulse residual module, aiming to enhance the model's ability to capture key features and suppress irrelevant information, thereby improving the accuracy of deception recognition.

[0011] As a further limitation of the first aspect of the present invention, the neck network includes a multi-level feature fusion module and a pulse residual module; wherein the multi-level feature fusion module fuses and interacts information from three different scales: high-level, low-level and current layers; high-level features are processed using an upsampling-pulse neuron-1*1 convolution layer-pooling layer structure and scale-aligned with the current layer features; low-level features are processed using a downsampling module to reduce the scale of the feature map and expand the channel dimension to align the scale of the current layer features; the current feature map is processed by the pulse neuron and used as a 0 / 1 gating mechanism to automatically match and fuse high-level semantic features and low-level semantic features, and the processed three-level features are element-wise summed to obtain a fused adversarial pulse feature image; the pulse residual module is used to enhance the fused feature map and capture key information in the adversarial pulse feature image.

[0012] As a further limitation of the first aspect of the present invention, the two classification heads only take the highest-level features as input, and adopt the average pooling-spiking neuron-linear layer structure to predict the deception confidence and attack situation label respectively. The last layer of the classification head is the spiking neuron layer, and the average firing rate of the spiking neurons is used as the final output; for the deception detection task, the last spiking neuron layer contains only one neuron, and the average firing rate in the time step dimension is output as the deception confidence; for the deception attack situation perception task, the last spiking neuron layer contains several neurons corresponding to different deception attack situations in turn, and the average firing rate of each neuron is directly output as the attack situation score, and the final situation prediction label is the attack situation corresponding to the neuron with the largest spiking firing rate.

[0013] As a further limitation of the first aspect of the present invention, the regression head uses fused features of different levels as input to enhance the ability to recognize sub-signals of different intensities; the regression task is used to complete the estimation of the deceptive signal parameters, intending to decompose the autocorrelation function to solve the parameters of several sub-signals.

[0014] In the second aspect, the present invention provides a multi-task pulse neural network assisted satellite positioning deception defense system, including: an offline training module, which uses a satellite positioning receiver with a built-in deception countermeasure module to extract and process binary pulse deception features from the collected satellite intermediate frequency signal data, and is used for offline independent training of the multi-task pulse neural network model, storing and exporting the optimal model parameters as a target format callable by the satellite positioning receiver; the binary pulse processing deception features are used to integrate continuous real-valued features into an adversarial pulse feature image; a real-time processing module, which uses a satellite positioning receiver embedded with a pre-trained multi-task pulse neural network model to process the satellite intermediate frequency signal data acquired in real time to obtain a trusted operating state of the carrier; the multi-task pulse neural network model includes a multi-task shared backbone network, a neck network and a head network; the multi-task shared backbone network is used to extract relevant semantic information from the deception features; the neck network is used to interactively fuse and enhance the semantic information extracted by the shared backbone network; the head network includes two classification heads and one regression head; the two classification heads are used for deception detection and attack situation awareness tasks respectively, and the regression head is used for deception estimation tasks.

[0015] In a third aspect, the present invention provides a non-transitory computer-readable storage medium, which is used to store computer instructions. When the computer instructions are executed by a processor, the multi-task pulse neural network assisted satellite positioning deception defense method as described in the first aspect is implemented.

[0016] In a fourth aspect, the present invention provides a computer device comprising a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the multi-task pulse neural network assisted satellite positioning deception defense method as described in the first aspect.

[0017] In a fifth aspect, the present invention provides an electronic device comprising: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory so that the electronic device executes instructions for implementing the multi-task pulse neural network assisted satellite positioning deception defense method as described in the first aspect.

[0018] The beneficial effects of the present invention are as follows: a unified optimization architecture for multi-task collaborative modeling is developed using a pulse neural network model to achieve joint optimization of deception detection, attack situation identification, and deception signal parameter estimation. An adaptive state solver is further constructed to combine model-specific outputs, satellite measurements after real / deception elimination, and auxiliary sensor measurements to restore the true operating state of the satellite positioning receiver. This expands the application potential of neural networks in the integrated "detection-suppression" path of GNSS deception attacks, realizes a closed loop for deception defense, and enhances the security and credibility of GNSS location services.

[0019] Additional advantages of the present invention will be more clearly given in the following description or learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] Figure 1 This is a flow chart of the multi-task pulse neural network assisted satellite positioning deception defense method described in an embodiment of the present invention.

[0022] Figure 2 Schematic diagram of possible in-phase branch output vectors and their corresponding binary sparse matrices according to an embodiment of the present invention.

[0023] Figure 3 Schematic diagram of a possible implementation of a multi-task spiking neural network model for countering spoofing attacks according to an embodiment of the present invention.

[0024] Figure 4 Schematic diagram of the deployment and application of a possible deception adversarial network model described in an embodiment of the present invention on a satellite positioning receiver.

[0025] Figure 5 A schematic diagram of the deception confidence output results of a possible deception adversarial network model described in an embodiment of the present invention.

[0026] Figure 6 This is a schematic diagram of possible satellite positioning errors after deception elimination according to an embodiment of the present invention. DETAILED DESCRIPTION

[0027] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention and are not to be construed as limiting the present invention.

[0028] Example 1

[0029] In this embodiment 1, a multi-task pulse neural network assisted satellite positioning deception defense system is first provided, including: an offline training module, which uses a satellite positioning receiver with a built-in deception countermeasure module to extract and process binary pulse deception features from collected satellite intermediate frequency signal data, and is used for offline independent training of a multi-task pulse neural network model, and stores and exports the optimal model parameters as a target format callable by the satellite positioning receiver; the binary pulse processing deception features are used to integrate continuous real-valued features into an adversarial pulse feature image; a real-time processing module, which uses a satellite positioning receiver embedded with a pre-trained multi-task pulse neural network model to process the satellite intermediate frequency signal data acquired in real time to obtain a trusted operating state of the carrier; the multi-task pulse neural network model includes a multi-task shared backbone network, a neck network and a head network; the multi-task shared backbone network is used to extract relevant semantic information from the deception features; the neck network is used to interactively fuse and enhance the semantic information extracted by the shared backbone network; the head network includes two classification heads and one regression head; the two classification heads are used for deception detection and attack situation awareness tasks respectively, and the regression head is used for deception estimation tasks.

[0030] In this embodiment, the above-mentioned system is used to implement a multi-task pulse neural network assisted satellite positioning deception defense method, including: using a satellite positioning receiver with a built-in deception countermeasure module to extract and process binary pulse deception features from the collected satellite intermediate frequency signal data, for offline independent training of the multi-task pulse neural network model, storing and exporting the optimal model parameters as a target format callable by the satellite positioning receiver; the binary pulse processing deception features are used to integrate continuous real-valued features into an adversarial pulse feature image; using a satellite positioning receiver embedded with a pre-trained multi-task pulse neural network model to process the satellite intermediate frequency signal data acquired in real time to obtain a trusted operating state of the carrier; the multi-task pulse neural network model includes a multi-task shared backbone network, a neck network and a head network; the multi-task shared backbone network is used to extract relevant semantic information from the deception features; the neck network is used to interactively fuse and enhance the semantic information extracted by the shared backbone network; the head network includes two classification heads and one regression head; the two classification heads are used for deception detection and attack situation awareness tasks respectively, and the regression head is used for deception estimation tasks.

[0031] The multi-task shared backbone network includes a feature downsampling module based on a spike neuron-convolutional layer-pooling layer structure, a spike residual module based on a dilated convolution-depthwise separable convolution, and a spike attention module. The downsampling module is used to compress the size of the adversarial spike feature image and expand the number of channels. The spike residual module is used to extract features from the adversarial spike feature image. The spike attention module is used to enhance the model's ability to capture key features and suppress irrelevant information, thereby improving deception recognition accuracy. In the spike residual module, its input and output spike feature maps have the same dimensions. Its dilated convolution is used for multi-scale feature extraction and expanded receptive field, enabling the model to capture a wider range of contextual information. The depthwise separable convolution is used to reduce the number of model parameters and computational complexity, keeping the model lightweight. The residual structure is used to alleviate the gradient vanishing problem that may exist in deep networks. Several downsampling modules and spike residual modules are alternately connected in series to achieve high-level semantic information extraction. A spike attention module is connected in series after the last spike residual module to enhance the model's ability to capture key features and suppress irrelevant information, thereby improving deception recognition accuracy.

[0032] Specifically, the shared backbone network consists of three main components: a feature downsampling module (Spike Neuron-Convolution-Batch Normalization (SCB)) based on a spiking neuron-convolution layer-pooling layer structure, a Dilated Separable Residual Block (DSResblock) based on a dilated convolution-depthwise separable convolution, and a spiking attention module. The downsampling module is used to compress the size of the adversarial spike feature image and expand the number of channels. The spiking residual block is used for feature extraction from the adversarial spike feature image. Its input and output spike feature maps have the same dimensions, and its dilated convolution is used for multi-scale feature extraction and expanded receptive field, enabling the model to capture a wider range of contextual information. The depthwise separable convolution is intended to reduce the number of model parameters and computational complexity, keeping the model lightweight. The residual structure is designed to alleviate the vanishing gradient problem that may exist in deep networks. Several downsampling modules and spiking residual modules are alternately connected in series to extract high-level semantic information. A spiking attention module is connected in series after the last spiking residual block to enhance the model's ability to capture key features and suppress irrelevant information, thereby improving deception recognition accuracy. Low-level semantic features exhibit pulse fluctuations and unclear semantic information. Information interaction at this stage will introduce information redundancy and internal conflicts. However, high-level semantic information, after being processed at multiple levels, exhibits rich semantic information, which facilitates the mining of information contributions. Therefore, the pulse attention module only acts on the highest-level semantic features.

[0033] The neck network includes a multi-level feature fusion module and a pulse residual module; among them, the multi-level feature fusion module fuses and interacts information from three different scales of high layer, low layer and current layer; high-level features are processed using an upsampling-pulse neuron-1*1 convolution layer-pooling layer structure to align the scale with the current layer features; low-level features are processed using a downsampling module to reduce the feature map scale and expand the channel dimension to align the scale of the current layer features; the current feature map is processed by the pulse neuron as a 0 / 1 gating mechanism to automatically match and fuse high-level semantic features and low-level semantic features, and the processed three-level features are summed element-by-element to obtain the fused adversarial pulse feature image; the pulse residual module is used to enhance the fused feature map and capture the key information in the adversarial pulse feature image.

[0034] Specifically, the neck network, located between the backbone and head networks, primarily fuses and enhances features at different scales from the shared backbone network to obtain rich semantic information and feature representations, thereby improving the model's ability to recognize signals of varying intensities. It comprises a multi-level feature fusion module (MFM) and a spike residual module. The MFM integrates and interacts information from three different scales: high-level, low-level, and current layers. High-level features are processed using an upsampling-spiking neuron-1x1 convolutional layer-pooling layer structure to align scales with the current layer features. Low-level features are processed using a downsampling module to reduce the feature map scale and expand the channel dimension to align with the scale of the current layer features. The current feature map is processed by spike neurons and used as a 0 / 1 gating mechanism to automatically match and fuse high-level and low-level semantic features. The processed three-level features are element-wise summed to produce the fused adversarial spike feature image. The spike residual module then enhances the fused feature map to capture key information within the adversarial spike feature image.

[0035] The two classification heads take only the highest-level features as input and employ an average pooling-spiking neuron-linear layer structure to predict the deception confidence and attack situation label, respectively. The final layer of the classification head is a spiking neuron layer, whose average firing rate is used as the final output. For the deception detection task, the final spiking neuron layer contains only one neuron, and the average firing rate in the time step dimension is used as the output for the deception confidence. For the deception attack situation awareness task, the final spiking neuron layer contains four neurons corresponding to four attack situations. The average firing rate of each neuron is directly output as the attack situation score, and the final situation prediction label is the attack situation corresponding to the neuron with the largest spiking firing rate. The regression head takes fused features of different levels as input to enhance the ability to recognize sub-signals of varying strengths. The regression task is used to estimate the parameters of the deception signal, aiming to decompose the autocorrelation function to solve for the parameters of several sub-signals.

[0036] Specifically, the two classification heads only use the highest-level features as input, and use the average pooling-spiking neuron-linear layer structure to predict the deception confidence and attack situation label respectively. The last layer of the classification head is the spiking neuron layer, and the average firing rate of the spiking neuron is used as the final output. For the deception detection task, the last spiking neuron layer contains only one neuron, and the average firing rate in the time step dimension is output as the deception confidence; for the deception attack situation perception task, the last spiking neuron layer contains several neurons corresponding to different deception attack situations in turn, and the average firing rate of each neuron is directly output as the attack situation score. The final situation prediction label is the attack situation corresponding to the neuron with the largest spiking firing rate. Finally, for each sample, the classification head used for deception detection outputs the deception confidence Its value is distributed between 0 and 1; the classification head used for deception attack situation awareness outputs a situation score vector O SA , whose elements are distributed between 0 and 1, and the situation corresponding to the highest score is the final predicted situation category of the deceptive adversarial network model

[0037] A regression head uses fused features at different levels as input, aiming to enhance the ability to recognize sub-signals of varying strengths. Compared to classification, regression is used to estimate the parameters of spoofing signals, aiming to decompose the autocorrelation function to determine the parameters of several sub-signals. This task presents a greater challenge, therefore requiring additional operations to improve regression accuracy.

[0038] Example 2

[0039] The application scenario of the multi-task spiking neural network-assisted satellite positioning deception defense method provided in Example 2 is as follows: a deception adversarial network model based on a multi-task spiking neural network model is constructed to simultaneously achieve deception detection, deception estimation, and deception attack situation awareness, and is embedded in a satellite positioning receiver. During the offline training phase of the network model, the network model embedded in the satellite positioning receiver is set to remain silent. The receiver is used to collect and extract deception features of the model input, perform binary pulse processing on them, and convert them into adversarial pulse feature images. The deception adversarial network model is independently trained using the constructed deception training set and validation set, and the model parameters are saved to the corresponding modules of the satellite positioning receiver. During online detection, the deception adversarial module embedded in the satellite positioning receiver onboard the carrier is activated. The satellite positioning receiver receives satellite intermediate frequency signals in real time and sequentially completes signal capture, multi-correlator calculation, deception feature binary pulse processing, deception adversarial network model interaction, deception signal rejection, and true signal re-tracking, outputting satellite measurements after deception components have been eliminated. Deception adaptive positioning solution is performed based on the output results of the deception adversarial network model and satellite multi-type measurements / multi-sensor measurements to restore the carrier's true operating state.

[0040] like Figure 1 As shown, the multi-task pulse neural network assisted satellite positioning deception defense method provided in this embodiment includes:

[0041] S1. Build a satellite positioning receiver embedded with a deception adversarial module, initialize its relevant parameters, extract and binary pulse process deception features from the satellite positioning intermediate frequency signal collected by the receiver, and establish a training set and validation set for offline training of the deception adversarial network model.

[0042] According to the deception adversarial target and the input requirements of the adversarial network model, a satellite positioning receiver capable of embedding a deception adversarial module is designed. The parameters of the satellite positioning receiver, such as signal capture, tracking, and multi-correlator structure, are initialized. The satellite positioning intermediate frequency signal is collected and processed. The input features of the deception adversarial network model are extracted, and binary pulse processing is performed to convert the image into an adversarial pulse feature image to meet the standard input format of the model. The training set and validation set of the deception adversarial network are constructed.

[0043] S1.1. Construct a satellite positioning receiver embedded with a deception countermeasure module, initialize the satellite positioning receiver's parameters such as signal capture, tracking, and multi-correlator structure, and use the multi-correlator structure to extract deception features.

[0044] The deception countermeasure module is embedded in the signal tracking loop of the satellite positioning receiver. A multi-correlator structure is set up in each satellite signal tracking loop to extract the autocorrelation function (ACF). The code delay interval of the multi-correlator is Δd τchip, covering ±T of the prompt correlator τ Within the chip range, a total of N τ =2T τ / Δd τ +1 correlator. The code delay interval for the standard advanced (E), prompt (P), and lagging (L) correlators used for phase detection is 0.5 chips. The autocorrelation function of the visible satellite signal can comprehensively characterize the signal's health and parameter properties. Therefore, the autocorrelation function is used as input to a deception adversarial model developed based on a multi-task spiking neural network. Each satellite channel shares the same network structure and parameters.

[0045] Under a spoofing attack, the intermediate frequency signal of the satellite receiver consists of real satellite signals, spoofing signals, and noise. The signal of a single satellite under spoofing attack can be modeled as

[0046]

[0047] Among them, s au (t) is the real satellite signal, s sp (t) is the deceptive signal, w(t) is the additive white Gaussian noise, P is the signal power, C is the satellite pseudo-code sequence, D is the satellite modulated data code information, τ is the propagation delay of the satellite signal, f IF is the intermediate frequency, f d is the Doppler shift, is the initial phase of the carrier. The code delay of the spoofing signal may be ahead of or behind the real signal, and the navigation data of the spoofing signal is consistent with the real signal to achieve the purpose of interfering with the target receiver. The spoofing signal has a carrier frequency close to the real signal to avoid the receiver from losing lock, that is,

[0048] For this satellite, the signal generated locally by the satellite positioning receiver is

[0049]

[0050] Among them, τ g , f g , are the code delay of the local C / A code, the Doppler shift of the carrier, and the initial phase, respectively. Assuming that the frequency has been compensated and the noise is ignored, the satellite signal is mixed with the locally generated carrier signal, and the high-frequency components are filtered out to obtain the correlation results of the in-phase I branch and the orthogonal Q branch, which can be written as

[0051]

[0052] Where Δτ s is the code delay between the actual deception position and the local instant code in the deception environment. Under the action of the phase detector, this value is stable near 0. Δτau and Δτ sp are the code delays of the real signal, the deceptive signal and the actual deceptive position, (Δτ au -Δτ s ) and (Δτ sp -Δτ s ) are the code delay distances between the real signal, the deceptive signal and the local real-time code, is the carrier phase error, R(x) is the standard normalized autocorrelation function, which can be modeled as

[0053]

[0054] In the time [t1, t1+T], the correlation results are coherently integrated and the result is obtained.

[0055]

[0056] The coherent integration of the received signal is the superposition of the coherent integration of the real signal and the spoofed signal, and then the code phase delay corresponding to the multi-correlator is set to δτ i , then the coherent integral value of the i-th in-phase branch and the orthogonal branch is

[0057]

[0058] Finally, at a certain epoch, the one-dimensional vector of the autocorrelation function of the received signal is expressed as

[0059]

[0060] S1.2. Process the autocorrelation function into an adversarial pulse feature image to meet the standard input format of the deceptive adversarial network model, and construct a training dataset and a verification dataset for offline training of the deceptive adversarial network model.

[0061] Binary pulse processing is to convert the autocorrelation function output by the multi-correlator from a continuous real value into a binary sparse matrix to meet the discrete binary input of the pulse neural network, enabling the image processing capability of the neural network to complete specific tasks. Take the following example to describe the steps of obtaining the adversarial pulse feature image.

[0062] (1) The autocorrelation function (ACF) is normalized based on the predefined average amplitude of the autocorrelation function of the interference-free signal, and its amplitude is constrained within [-3, 3]. Compared with normalization, this autocorrelation function processing method can retain the amplitude information and reduce the loss of deception-related information;

[0063] (2) Use linear interpolation method to convert the in-phase branch vector Expanding to longer vectors (W≥N τ ), W is the length of the adversarial pulse feature image;

[0064] (3) Expand the range of the in-phase branch vector after expansion with a scaling factor of H / 2, where H represents the width of the adversarial pulse feature image, and perform integer rounding and offset processing to obtain the subscript index vector

[0065]

[0066] (4) Construct a sparse matrix of all 0s Search the sparse matrix column by column according to the subscript index vector, setting the corresponding positions to 1.

[0067] According to the above steps, the in-phase branch vector is processed into a binary sparse matrix The same steps are used to perform binary pulse processing on the orthogonal branches to obtain the corresponding sparse matrix By concatenating the two binary sparse matrices in the channel dimension, we can get the standard input of the deceptive adversarial network model, that is, the adversarial feature image with a pulse form.

[0068] All extracted samples are labeled with the following: a binary label of 0 / 1 for deception detection, indicating the presence / absence of a deception attack; and four-category labels for deception attack situational awareness, namely, deception latency, attack period, induction period, and capture period, represented by 0, 1, 2, and 3, respectively. The deception latency indicates the presence of a deception attack, but due to low signal strength, no change in the correlation function occurs. The deception attack period indicates that the signal power of the deception attack begins to increase, causing a change in the amplitude of the correlation function, but the deception attack does not significantly distort the symmetry of the correlation function. The deception induction period indicates that there is a code delay between the deception signal and the true signal, destroying the symmetry of the correlation function. The deception capture period indicates that the correlation function has two distinct separate peaks, representing the true signal peak and the deception signal peak, respectively, and that the true signal peak does not affect the phase detection processing of the deception signal. Deception estimation aims to decompose the input correlation function and is designed to be trained without labels.

[0069] The labeled deception feature map set is divided into a training set and a validation set for offline training of the deception adversarial network model.

[0070] Figure 2 A possible in-phase branch output vector and its corresponding binary sparse matrix are given.

[0071] S2. Build a multi-task spiking neural network model capable of performing deception detection, deception estimation, and deception attack situation recognition. Use the dataset to independently train the model offline, store and export the optimal model parameters into a target format that can be called by the satellite positioning receiver.

[0072] S2.1. Independent of the satellite positioning receiver, a multi-task deception adversarial network model capable of performing deception detection, deception estimation, and deception attack situation recognition is built using a pulse neural network architecture. The model includes a shared backbone network for extracting relevant semantic information from the input adversarial pulse feature image, a neck network for interactively fusing and enhancing the semantic information extracted by the shared backbone network, and a head network for three subtasks (two classification heads for deception detection and deception attack situation awareness tasks, respectively, and a regression head for deception estimation).

[0073] Pulse neural networks use information encoding and processing mechanisms similar to those of the biological brain to achieve information extraction. Unlike existing artificial neural networks, the binary pulse information transmission characteristics avoid the high energy consumption required for floating-point calculations and have natural noise robustness. They have become a key technical path for edge and mobile devices to break through energy consumption and inference time. With the development of neuromorphic computing hardware, the large-scale deployment of pulse neural networks on edge nodes and mobile devices has been accelerated. Therefore, building a deception adversarial network model based on pulse neural networks is suitable for resource-constrained scenarios of carrier mobile devices and meets their high inference speed requirements for network models. In the deception adversarial task, the input of the model is the accumulation of the constructed deception feature map within time step T. The output can be modeled as

[0074]

[0075] in, is the deception confidence, which indicates whether the visible satellite is under deception attack at the current epoch, and is used for deception detection tasks; It is a deception attack situation prediction category, used for deception attack situation awareness tasks; They represent the amplitude, code delay and carrier phase difference of the autocorrelation function subcomponents respectively, and Num is the predefined number of subcomponents used for the deception estimation task.

[0076] like Figure 3 As shown, the multi-task spiking neural network model for countering deception attacks provided by this embodiment includes a shared backbone network, a neck network and a head network.

[0077] The shared backbone network consists of three main components: a feature downsampling module (Spike Neuron-Convolution-Batch Normalization (SCB)) based on a spiking neuron-convolution layer-pooling layer structure, a dilated separable residual block (DSResblock) based on a dilated convolution-depthwise separable convolution layer, and a spiking attention module. The downsampling module is used to compress the size of the adversarial spike feature image and expand the number of channels. The spiking residual block is used for feature extraction from the adversarial spike feature image. Its input and output spike feature maps have the same dimensions, and its dilated convolution is used for multi-scale feature extraction and expanded receptive field, enabling the model to capture a wider range of contextual information. The depthwise separable convolution is intended to reduce the number of model parameters and computational complexity, maintaining the model's lightweightness. The residual structure is designed to alleviate the vanishing gradient problem that may exist in deep networks. Several downsampling modules and spiking residual modules are alternately connected in series to extract high-level semantic information. A spiking attention module is connected in series after the last spiking residual block to enhance the model's ability to capture key features and suppress irrelevant information, thereby improving deception recognition accuracy. Low-level semantic features exhibit pulse fluctuations and unclear semantic information. Information interaction at this stage will introduce information redundancy and internal conflicts. However, high-level semantic information, after being processed at multiple levels, exhibits rich semantic information, which facilitates the mining of information contributions. Therefore, the pulse attention module only acts on the highest-level semantic features.

[0078] The neck network, located between the backbone and head networks, primarily fuses and enhances features at different scales from the shared backbone network to obtain rich semantic information and feature representations, thereby improving the model's ability to recognize signals of varying intensities. It comprises a multi-level feature fusion module (MFM) and a spike residual module. The MFM integrates information from three different scales: high-level, low-level, and current layers. High-level features are processed using an upsampling-spiking neuron-1x1 convolutional layer-pooling layer structure to align their scales with those of the current layer. Low-level features are processed using a downsampling module to reduce the feature map scale and expand the channel dimension to align with the scale of the current layer. The current feature map is processed by spike neurons and used as a 0 / 1 gating mechanism to automatically match and fuse high-level and low-level semantic features. The processed three-level features are element-wise summed to produce the fused adversarial spike feature image. The spike residual module then enhances the fused feature map to capture key information within the adversarial spike feature image.

[0079] The two classification heads only use the highest-level features as input, and use the average pooling-spiking neuron-linear layer structure to predict the deception confidence and attack situation label respectively. The last layer of the classification head is the spiking neuron layer, and the average firing rate of the spiking neuron is used as the final output. For the deception detection task, the last spiking neuron layer contains only one neuron, and its average firing rate in the time step dimension is output as the deception confidence; for the deception attack situation perception task, the last spiking neuron layer contains four neurons corresponding to four attack situations in turn, and the average firing rate of each neuron is directly output as the attack situation score. The final situation prediction label is the attack situation corresponding to the neuron with the largest spiking firing rate. Finally, for each sample, the classification head for deception detection outputs the deception confidence Its value is distributed between 0 and 1; the classification head used for deception attack situation awareness outputs a situation score vector Its elements are distributed between 0 and 1, and the situation corresponding to the highest score is the final predicted situation category of the deceptive adversarial network model.

[0080] A regression head uses fused features at different levels as input, aiming to enhance the ability to recognize sub-signals of different strengths. Compared to classification tasks, regression tasks are used to estimate the parameters of spoofing signals, aiming to decompose the autocorrelation function to solve the parameters of several sub-signals. This task is more challenging, so additional operations are needed to improve regression accuracy. From Equation (6), it can be seen that the autocorrelation function is a superposition of the real signal and the spoofing signal. Under the influence of observation noise, the correlation function will be composed of several sub-signals. The regression head design method is based on grid partitioning.

[0081] The autocorrelation function space [-T τ ,T τ ] code chips with finer granularity Δd l (Δd l ≤Δd τ ) chips are divided into (2T τ / Δd l +1) predefined grid points, assuming that there are sub-signals near the predefined grid points, the regression head of the deception estimation predicts the code delay off-grid error Δτ∈[-Δd l / 2, Δd l / 2] Code chip, signal amplitude A and carrier phase difference For the i-th sub-signal, its code delay can be calculated as (i-1)Δd l -T τ +Δτ iThe deception estimation method based on grid division can avoid a sub-signal from being decomposed multiple times, but its disadvantage is that it cannot effectively estimate the code delay difference between the real signal and the deception signal less than Δd l signal.

[0082] The regression head uses a long-short-term memory (LSTM) spike processing layer-linear layer structure. Long-short-term memory (LSTM) neural networks, developed based on artificial neurons, demonstrate significant performance advantages in time series forecasting tasks, particularly when processing sequences with strong temporal correlations, significant trends / periodicities, or significant long-term temporal dependencies. The LSTM spike processing layer is a LSTM neural network layer built on spiking neurons, an implementation of a recurrent spiking neural network. It includes three control gates and two spiking neurons (σ1 and σ2).

[0083]

[0084] Among them, ⊙ is the Hadamard product, w and b are weights and biases, and the forget gate f t Decide how much to keep from the current cell state m t Information, input gate i t Controls how much new information should be written to the current cell state, the output gate o t Determines how much information about the cell state is output to the hidden state h t The output of the LSTM spike processing layer at the last time step is input to the linear layer, which outputs a signal parameter matrix Each row represents the code delay, amplitude, and carrier phase difference of a sub-signal, respectively.

[0085] The classification head used for deception detection and deception attack situational awareness combines high-level semantic information from multiple time steps to make final decisions, consistent with the stable nature of deception attacks. To prevent the victim receiver from losing lock, deception signals often have signal parameters that change slowly over time, which gives deception attacks a temporal drift property. Therefore, when determining whether a deception attack exists and what attack situation it is in, combining multiple samples from the most recent epoch to jointly determine the current system state can enhance classification accuracy. The deception estimation task aims to determine the deception signal parameters at the current moment. Therefore, a long short-term memory spike processing layer is used to reflect the information from the most recent epoch on the semantic information at the current moment. This further focuses on the semantic information at the current epoch to regress the signal parameters of interest. This effectively balances the impact of the semantic information from the most recent epoch on the current epoch, thereby enhancing the accuracy of parameter regression prediction.

[0086] S2.2. Use the training set to perform offline training on the deception adversarial network model. The validation set is used to verify the performance of the model. After the model performance is improved, the model parameters are saved. After the training is completed, the saved model is exported to the target format for deployment on satellite positioning receivers.

[0087] The loss function is used to quantify the gap between the model prediction output and the true label. For the deception detection task, the binary cross-entropy loss (BCE) is used. SD It is used to measure the difference between the deception confidence output by the model and the true deception label. For the deception attack situation awareness task, the Hinge loss (Multi Margin Loss) of multi-category classification is used. SA This loss encourages the model to maximize the difference between the scores of the correct class and the scores of the incorrect class, hoping that the correct class score is higher than the scores of other classes by at least a margin. To determine an appropriate margin value, a dynamic adjustment strategy is adopted, where the margin value is set to increase linearly with the number of model training epochs. In the early stages of training, a small margin allows the model to quickly learn preliminary distinctions. As training epochs increase, the margin value is gradually increased to refine the classification boundaries.

[0088] For the deception estimation task, first, the signal parameter matrix output by the regression head is Perform scale scaling to satisfy the actual constraint range of the parameters and obtain a parameter matrix with actual physical meaning. Then, the correlation function sequence of the received signal is reconstructed according to the parameter matrix It can be written as

[0089]

[0090] The mean squared error (MSE) is used to measure the difference between the reconstructed correlation function and the true input. In addition, in order to automatically determine the truly effective sub-signals (real signals and deceptive signals), a sparse regularization term is added to the loss function to suppress the amplitude of meaningless sub-signals. The loss function of the regression task can be written as

[0091]

[0092] Among them, λ is the penalty factor.

[0093] Finally, the overall loss function of the multi-task deception adversarial network model can be modeled as

[0094] Among them, the coefficient κ is used to adjust the weight coefficient of the task.

[0095] Set model training parameters such as the epoch, learning rate, and optimizer. Use the training set to perform multiple batches of forward propagation, loss calculation, and proxy-based gradient backpropagation. After each round of training, use the validation set to infer the model and calculate performance metrics. If performance improves, save the current model parameters. Model performance metrics include classification accuracy and relevant function fitting error. After model training, export the best-performing model parameters on the validation set to the data format required by the satellite positioning receiver for deployment. For example, if the neural network model is built and trained using PyTorch and the satellite positioning receiver is implemented using Matlab, export the trained model parameters to a target format that can be used by Matlab.

[0096] like Figure 4 As shown in the figure, the deployment and application of the deception adversarial network model on the satellite positioning receiver.

[0097] S3. Deploy the deception adversarial network model on the satellite positioning receiver and activate the adversarial module. Through the relevant interface, it interacts with the pre-stored network model in real time. Based on the model prediction output, it is decided whether the specific visible satellite is under deception attack to achieve diversion control.

[0098] The satellite positioning receiver program configures the relevant environment and interfaces for invoking the spiking neural network model, specifies the path to the derived model parameters, and activates the receiver's deception adversarial network module. Using offline satellite intermediate frequency signals, the receiver is tested for its ability to properly invoke the deception adversarial network model. During real-time application, the receiver sequentially performs signal acquisition, signal tracking, and correlation function calculations. The correlation functions are then processed using binary pulses and stacked in time. The sparse pulse signatures of the correlation functions within a specific window are then passed to the pre-existing deception adversarial network model via the relevant program interface. The model's real-time outputs, including deception confidence, deception attack situation labels, and sub-signal parameter matrices, are then passed to the receiver via the program interface. The receiver then performs subsequent processing based on the model's outputs.

[0099] First, according to the predicted deception confidence and the preset threshold T h The decision of whether the tracked satellite is under spoofing attack and the realization of diversion control can be modeled as

[0100]

[0101] If it is determined that the tracked satellite has not been subjected to a spoofing attack, the satellite is identified as a real satellite, and the satellite positioning receiver performs related operations normally, including satellite navigation message decoding and non-spoofing measurement calculations; if it is determined that a spoofing attack has occurred, the satellite signal is a mixture of real signals and spoofing signals, and the satellite positioning receiver needs to perform subsequent related operations to eliminate the spoofing components in the satellite signal and mitigate the impact of the spoofing attack on the estimation of the carrier's operating status, such as the position.

[0102] S4. Under a spoofing attack, reconstruct and remove the spoofing components in the received signal based on the spoofing estimation results, calibrate the phase detector of the satellite positioning receiver to maintain stable tracking of the real signal, and calculate the satellite measurement after spoofing correction.

[0103] Calculate the correlation function of each sub-signal in the received signal based on the deception estimation result for

[0104]

[0105] In the received signal, the real signal and the deception signal often have a significant amplitude advantage compared to the noise signal. Therefore, by traversing the combination of all sub-signals, two sub-signals with the smallest fitting error can be easily obtained, where one sub-signal corresponds to the real signal and the other sub-signal corresponds to the deception signal. In order to determine the identity of the two sub-signals, it is necessary to further combine the characteristics of the deception signal and the auxiliary information to make a judgment. A successful deception often shows a stable change in signal parameters to avoid attracting the attention of the target receiver and losing lock. The deception signal has a larger pseudorange offset than the real signal. Based on this, the signal parameters corresponding to the deception component can be determined. Then the deceptive component is extracted from the correlation function of the received signal, and the correlation function of the restored real signal is obtained as

[0106]

[0107] Based on the recovered true signal correlation function, the standard EPL correlator value is extracted to calibrate the phase detector, thereby mitigating the impact of spoofing attacks on the signal tracking loop. Subsequently, the satellite positioning receiver decodes the signal navigation potential and calculates the satellite measurement after the spoofing is eliminated.

[0108] S5. Evaluate the current satellite positioning quality based on the deception attack situational awareness results, build an adaptive state estimator, and combine the true / deception-eliminated satellite measurements and multi-sensor measurements to estimate the true state of the satellite positioning receiver before being deceived.

[0109] If the deception adversarial network model predicts that the system is currently in the deception latency period, the deception attack is very weak and will not affect the calculation of satellite measurements. However, if the prediction results indicate that the system is currently in the deception attack period, induction period, or acquisition period, the deception attack will introduce varying degrees of bias in the satellite measurements. After removing the deception component based on the deception estimation, the satellite measurement bias is corrected to a certain extent. However, due to the introduction of the deception attack, the satellite measurements still have varying degrees of noise floor and residual bias that are not effectively corrected. Therefore, it is necessary to further design a state estimator with adaptive characteristics for satellite measurement errors based on the deception attack situation. In addition, multi-sensor measurements can be further combined with adaptive state estimators to enhance positioning accuracy.

[0110] An adaptive state estimator based on multi-sensor fusion is built on the factor graph optimization framework. The core of the algorithm is to automatically match the optimal satellite measurement covariance matrix (R matrix) according to the deception attack situation to improve the robustness and accuracy of the state estimator. The specific steps are as follows:

[0111] 1. Use the constructed training set and validation set to establish the observation satellite measurement samples ρ under different deception attack situations obs And the reference trajectory set, under each attack situation and each visible satellite, solve and store the optimal observation error variance. The solution methods include the method based on mathematical statistics and the regression prediction method based on the neural network model.

[0112] The method for solving the observation covariance matrix based on mathematical statistics: For each visible satellite, assume that at a certain epoch, the measurement of the satellite is ρ obs , solve the reference satellite measurement ρ using the reference trajectory and the satellite's three-dimensional spatial position and velocity ref , then the observation error of satellite measurement can be written as

[0113] e ρ =ρ obs -ρ ref (17)

[0114] The probability density function of the satellite's observation error vector is fitted with a priori distribution to obtain the variance of the observation error, which is the diagonal element of the observation covariance matrix R corresponding to the satellite.

[0115] 2. In real-time positioning solution, first, based on the deception attack situation category predicted by the deception adversarial network model, the corresponding stored observation error variance is automatically extracted, and the observation covariance matrix R of the state estimator is constructed and adjusted. It is further combined with auxiliary sensor measurements to jointly optimize and solve the carrier positioning information.

[0116] The observation covariance matrix adaptive adjustment strategy based on the deception attack situation can balance the calculation real-time and positioning accuracy, and dynamically adapt to the more complex carrier operating environment under the action of deception attacks.

[0117] like Figure 5 and Figure 6 The figure shows a schematic diagram of the deception confidence level and the satellite positioning error after deception elimination. This method is suitable for detecting and mitigating satellite positioning deception interference in different operating scenarios of satellite positioning receiver carriers, and has significant engineering application value.

[0118] Example 3

[0119] This embodiment 3 provides a non-transitory computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, they implement the multi-task spiking neural network-assisted satellite positioning deception defense method described above. The method includes: using a satellite positioning receiver with a built-in deception countermeasure module to extract and process binary pulse deception features from collected satellite intermediate frequency signal data, for offline independent training of a multi-task spiking neural network model, and storing and deriving optimal model parameters in a target format callable by the satellite positioning receiver; the binary pulse processing deception features are used to integrate continuous real-valued features into an adversarial pulse feature image; using a satellite positioning receiver embedded with a pre-trained multi-task spiking neural network model to process the satellite intermediate frequency signal data acquired in real time to obtain a trusted operating state of the carrier; the multi-task spiking neural network model includes a multi-task shared backbone network, a neck network, and a head network; the multi-task shared backbone network is used to extract relevant semantic information from the deception features; the neck network is used to interactively fuse and enhance the semantic information extracted by the shared backbone network; the head network includes two classification heads and one regression head; the two classification heads are used for deception detection and attack situation awareness tasks, respectively, and the regression head is used for deception estimation tasks.

[0120] Example 4

[0121] This embodiment 4 provides a computer device, including a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions executable by the processor, and the processor calls the program instructions to execute the multi-task pulse neural network assisted satellite positioning deception defense method as described above, the method comprising: using a satellite positioning receiver with a built-in deception countermeasure module to extract and binary pulse process deception features from collected satellite intermediate frequency signal data, for offline independent training of a multi-task pulse neural network model, and storing and exporting the optimal model parameters as a target format callable by the satellite positioning receiver; the binary pulse processing deception features are used to convert continuous real The value features are integrated into an adversarial pulse feature image; a satellite positioning receiver embedded with a pre-trained multi-task pulse neural network model is used to process the satellite intermediate frequency signal data acquired in real time to obtain the trusted operating state of the carrier; the multi-task pulse neural network model includes a multi-task shared backbone network, a neck network and a head network; the multi-task shared backbone network is used to extract relevant semantic information from the deception features; the neck network is used to interactively fuse and enhance the semantic information extracted by the shared backbone network; the head network includes two classification heads and one regression head; the two classification heads are used for deception detection and attack situation awareness tasks respectively, and the regression head is used for deception estimation tasks.

[0122] Example 5

[0123] This embodiment 5 provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to execute instructions for implementing the multi-task pulse neural network assisted satellite positioning deception defense method as described above, the method comprising: using a satellite positioning receiver with a built-in deception countermeasure module to extract and binary pulse process deception features from collected satellite intermediate frequency signal data for offline independent training of a multi-task pulse neural network model, and storing and deriving optimal model parameters as a target format that can be called by the satellite positioning receiver; the binary pulse processing The deception features are used to integrate continuous real-valued features into adversarial pulse feature images; a satellite positioning receiver embedded with a pre-trained multi-task pulse neural network model is used to process the satellite intermediate frequency signal data acquired in real time to obtain a trusted operating state of the carrier; the multi-task pulse neural network model includes a multi-task shared backbone network, a neck network and a head network; the multi-task shared backbone network is used to extract relevant semantic information from the deception features; the neck network is used to interactively fuse and enhance the semantic information extracted by the shared backbone network; the head network includes two classification heads and one regression head; the two classification heads are used for deception detection and attack situation awareness tasks respectively, and the regression head is used for deception estimation tasks.

[0124] In summary, the present invention provides a multi-task pulse neural network assisted satellite positioning deception defense method. Among them, a multi-task pulse neural network model is used to realize the cognition and confrontation of satellite positioning deception attacks, including a two-classification task of deception detection, a parameter regression task of deception estimation, and a four-classification task of deception attack situation awareness. It is embedded in the satellite positioning receiver to construct a closed-loop defense technology framework covering deception detection, deception estimation, deception attack situation awareness, deception elimination and real carrier operation status recovery. The multi-task pulse neural network model includes a multi-task shared backbone network for extracting relevant semantic information from the input deception features, a neck network for interactively fusing and enhancing the semantic information extracted by the shared backbone network, and a head network for three subtasks (two classification heads for deception detection and attack situation awareness tasks, and one regression head for deception estimation task). During the offline training phase, the deception countermeasure module is kept silent, and the satellite positioning receiver uses a multi-correlator structure to extract deception features for model input, and performs binary pulse processing on the deception features to construct a training set; offline training is performed on an independently constructed multi-task pulse neural network model, and the relevant interactive interface between the satellite positioning receiver and the pulse neural network model is configured, and the trained optimal model parameters are saved in the deception countermeasure module of the satellite positioning receiver; during online detection, the deception countermeasure module of the satellite positioning receiver is activated, and the satellite positioning receiver receives satellite intermediate frequency signals, signal capture, calculation model input, network model information interaction and signal re-tracking in real time. During the positioning solution process, the trusted operating state of the carrier is estimated based on the predicted output of the deception countermeasure module, satellite measurements after the deception components are eliminated, and auxiliary sensor measurements. The present invention designs a multi-task collaboratively optimized pulse neural network model for recognizing and locating deception components in satellite signals. Combined with the embedded deployment of the satellite positioning receiver, it realizes a full-process defense architecture from signal layer analysis to recovery of the carrier's true operating state, providing a systematic solution for achieving active and trusted GNSS positioning anti-deception in specific carrier target tasks.

[0125] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solutions disclosed in the present invention without the need for creative work should be included in the scope of protection of the present invention.

Claims

1. A multi-task pulse neural network assisted satellite positioning deception defense method, characterized in that: include: A satellite positioning receiver with a built-in deception countermeasure module is used to extract binary pulse processing deception features from collected satellite intermediate frequency signal data. The features are used to independently train a multi-task spiking neural network model offline, and the optimal model parameters are stored and exported in a target format that can be called by the satellite positioning receiver. The binary pulse processing deception features are used to integrate continuous real-valued features into an adversarial pulse feature image. A satellite positioning receiver embedded with a pre-trained multi-task spiking neural network model is used to process satellite intermediate frequency signal data acquired in real time to obtain a trusted operating state of the carrier; the multi-task spiking neural network model includes a multi-task shared backbone network, a neck network, and a head network; the multi-task shared backbone network is used to extract relevant semantic information from deception features; the neck network is used to interactively fuse and enhance the semantic information extracted by the shared backbone network; the head network includes two classification heads and one regression head; the two classification heads are used for deception detection and attack situation awareness tasks respectively, and the regression head is used for deception estimation tasks.

2. The multi-task pulse neural network assisted satellite positioning deception defense method according to claim 1 is characterized in that: The multi-task shared backbone network includes a feature downsampling module based on a spike neuron-convolution layer-pooling layer structure, a spike residual module based on a dilated convolution-depthwise separable convolution, and a spike attention module. The downsampling module is used to compress the size of the adversarial spike feature image and expand the number of channels. The spike residual module is used to extract features from the adversarial spike feature image. The pulse attention module is used to enhance the model's ability to capture key features and suppress irrelevant information, thereby improving the accuracy of deception recognition.

3. The multi-task pulse neural network assisted satellite positioning deception defense method according to claim 2 is characterized in that: In the pulse residual module, its input and output pulse feature maps have the same dimension, its void convolution is used for multi-scale feature extraction and expanded receptive field, enabling the model to capture a wider range of contextual information, depthwise separable convolution is used to reduce the number of model parameters and computational complexity, keeping the model lightweight, and the residual structure is used to alleviate the gradient vanishing problem that may exist in deep networks; several downsampling modules and pulse residual modules are alternately connected in series to achieve high-level semantic information extraction; a pulse attention module is connected in series after the last pulse residual module, aiming to enhance the model's ability to capture key features and suppress irrelevant information, thereby improving the accuracy of deception recognition.

4. The multi-task spiking neural network assisted satellite positioning deception defense method according to claim 1, characterized in that: The neck network includes a multi-level feature fusion module and a pulse residual module; among them, the multi-level feature fusion module fuses and interacts information from three different scales of high layer, low layer and current layer; high-level features are processed using an upsampling-pulse neuron-1*1 convolution layer-pooling layer structure to align the scale with the current layer features; low-level features are processed using a downsampling module to reduce the feature map scale and expand the channel dimension to align the scale of the current layer features; the current feature map is processed by the pulse neuron as a 0 / 1 gating mechanism to automatically match and fuse high-level semantic features and low-level semantic features, and the processed three-level features are summed element-by-element to obtain the fused adversarial pulse feature image; the pulse residual module is used to enhance the fused feature map and capture the key information in the adversarial pulse feature image.

5. The multi-task spiking neural network assisted satellite positioning deception defense method according to claim 1, characterized in that: The two classification heads only take the highest-level features as input and adopt the average pooling-spiking neuron-linear layer structure to predict the deception confidence and attack situation label respectively. The last layer of the classification head is the spiking neuron layer, and the average firing rate of the spiking neurons is used as the final output; for the deception detection task, the last spiking neuron layer contains only one neuron, and the average firing rate in the time step dimension is output as the deception confidence; for the deception attack situation perception task, the last spiking neuron layer contains several neurons corresponding to different deception attack situations in turn, and the average firing rate of each neuron is directly output as the attack situation score. The final situation prediction label is the attack situation corresponding to the neuron with the largest spiking firing rate.

6. The multi-task spiking neural network assisted satellite positioning deception defense method according to claim 5, characterized in that: The regression head takes fused features of different levels as input to enhance the ability to recognize sub-signals of different intensities; the regression task is used to complete the estimation of deceptive signal parameters, aiming to decompose the autocorrelation function and solve the parameters of several sub-signals.

7. A multi-task pulse neural network assisted satellite positioning deception defense system, characterized in that: include: An offline training module uses a satellite positioning receiver with a built-in deception countermeasure module to extract and process binary pulse deception features from collected satellite intermediate frequency signal data for offline independent training of a multi-task spiking neural network model, storing and exporting optimal model parameters in a target format callable by the satellite positioning receiver; the binary pulse processing deception features are used to integrate continuous real-valued features into an adversarial pulse feature image; The real-time processing module uses a satellite positioning receiver embedded with a pre-trained multi-task spiking neural network model to process the satellite intermediate frequency signal data acquired in real time to obtain the carrier's trusted operating state; the multi-task spiking neural network model includes a multi-task shared backbone network, a neck network, and a head network; the multi-task shared backbone network is used to extract relevant semantic information from deception features; the neck network is used to interactively fuse and enhance the semantic information extracted by the shared backbone network; the head network includes two classification heads and one regression head; the two classification heads are used for deception detection and attack situation awareness tasks respectively, and the regression head is used for deception estimation tasks.

8. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by the processor, the multi-task pulse neural network assisted satellite positioning deception defense method as described in any one of claims 1-6 is implemented.

9. A computer device, characterized in that: It includes a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the multi-task pulse neural network assisted satellite positioning deception defense method as described in any one of claims 1 to 6.

10. An electronic device, characterized in that: include: A processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to execute instructions for implementing the multi-task pulse neural network assisted satellite positioning deception defense method as described in any one of claims 1-6.

Images