Vehicle-mounted GPS slow deception detection method and system based on LSTM and residual space-time accumulation

Abstract

This invention provides a method for detecting slow-speed GPS spoofing in vehicles based on LSTM and residual spatiotemporal accumulation, belonging to the field of information data security for autonomous vehicles. The method includes: constructing a Long Short-Term Memory (LSTM) vehicle dynamics model using IMU data, CAN bus data, and GPS observation data from the KITTI public dataset to calculate the vehicle's predicted displacement independent of the GPS signal. First, the vehicle's perceived displacement is calculated using the Havesing formula combined with GPS latitude and longitude data. Then, the Euclidean distance between the LSTM predicted displacement and the perceived displacement calculated from GPS latitude and longitude is calculated, and a sliding window accumulation mechanism is used to amplify the small deviations caused by slow-speed spoofing. An adaptive threshold based on background noise statistical characteristics is used for anomaly detection. A zero-speed correction strategy is introduced to suppress sensor drift noise when the vehicle is stationary. This invention effectively ensures the security and reliability of positioning information data transmission for autonomous vehicles, preventing traffic accidents.

Description

Technical Field

[0001] This invention belongs to the field of information data security technology for autonomous vehicles, and relates to a method and system for detecting slow-speed spoofing of vehicle GPS based on LSTM and residual spatiotemporal accumulation. Background Technology

[0002] In recent years, with the rapid development of autonomous driving technology and intelligent transportation systems, the security requirements of in-vehicle high-precision positioning systems have become increasingly urgent. Therefore, the Global Navigation Satellite System (GPS), a key technology capable of providing absolute position information around the clock, has been widely used. However, due to the publicly available structure of civilian GPS signals and the weak signal strength reaching the ground, it is highly vulnerable to various malicious interferences and attacks. Among these, spoofing attacks allow attackers to generate false signals highly synchronized with the real signal, inducing the receiver to lock onto incorrect position parameters. In particular, slow-traction spoofing involves attackers altering the signal phase extremely slowly, causing the calculated position to gradually deviate from the true trajectory by a tiny margin. Therefore, this attack method is highly concealed and difficult for drivers or traditional signal monitoring mechanisms to detect in a short time, thus posing a serious safety hazard to autonomous vehicles traveling at high speeds. In a multi-source sensor fusion defense architecture, the key lies in how to effectively distinguish between the inherent measurement noise of sensors and the small systematic deviations caused by slow-traction spoofing.

[0003] Multi-sensor data fusion technology is a suitable solution to enhance the robustness of in-vehicle navigation. Utilizing autonomous sensors such as inertial measurement units (IMUs) to verify the consistency of GPS data is a mainstream approach. Due to the high accuracy of IMUs over short periods and their complete immunity to external radio signal interference, incorporating inertial navigation calculations into spoofing detection offers significant advantages. IMU-assisted fusion detection systems can improve vehicle positioning accuracy and provide an independent position reference. However, accurate construction of the dynamic model is crucial to achieving these goals. Based on traditional control theory, researchers have proposed an extended Kalman filter (EKF) residual detection algorithm to identify anomalies using the statistical characteristics of innovation sequences. Based on machine learning theory, a method using support vector machines (SVMs) to classify signal features has been studied. However, these works primarily address large-scale, abrupt spoofing and do not adequately consider the issue of minimal initial deviations in slow spoofing. With breakthroughs in deep learning for time-series processing, how to utilize neural networks to extract nonlinear features of vehicle motion and maintain detection sensitivity under complex road conditions is an important research direction in the field of navigation safety. To improve the fitting ability of dynamic trajectories, a vehicle trajectory prediction model based on a GRU network was studied to assist anomaly detection. A multimodal feature fusion strategy based on self-supervised learning was explored to enhance the system's generalization ability. However, while traditional instantaneous residual detection algorithms can respond to large deviations, they do not effectively utilize the cumulative effect of errors over time. Since the deviation introduced by slow-speed deception is often smaller than the instantaneous noise amplitude of the vehicle sensor, traditional algorithms are prone to false alarms or missed detections in high-noise environments, such as bumpy roads. To balance the relationship between detection sensitivity and false alarm rate, statistical methods based on cumulative sum (CUSUM) were studied in some anomaly detection domains. However, the above methods are not specifically optimized for the zero-speed drift characteristics of vehicle inertial navigation systems in a stationary state and cannot be directly applied to the complex driving scenarios considered in this invention. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method and system for detecting slow-speed GPS spoofing based on LSTM and spatiotemporal accumulation of residuals, to improve the perception capability of small positional deviations, while using a zero-speed correction strategy to suppress sensor drift noise when the vehicle is stationary. This invention considers the security model of location information data for autonomous vehicles in complex dynamic environments, describing the spoofing detection problem as a multi-sensor trajectory consistency discrimination problem, while also considering the exponential growth characteristics of slow-speed spoofing and the statistical distribution of environmental background noise. An LSTM vehicle dynamics model is constructed based on the KITTI public dataset to calculate the predicted displacement independent of satellite signals, and a sliding window is introduced to accumulate and sum the instantaneous residuals, amplifying small systematic deviations. Based on the statistical characteristics of background noise, an adaptive threshold algorithm is designed to achieve anomaly detection.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for detecting slow-speed spoofing of vehicle GPS based on LSTM and residual spatiotemporal accumulation, the method specifically includes the following steps: S1. Collect vehicle inertial measurement unit (IMU) data, CAN bus data, and GPS observation data; S2. Select important features from the original data, perform standard normalization processing, and construct the input model sequence; S3. Convert latitude and longitude coordinates into real-time vehicle displacement using the Havessin formula; S4. Construct an LSTM model to estimate the predicted displacement for the next second using historical data frames; S5. Injecting exponential displacement deviation to simulate slow, covert deception attack signals; S6. Adopt a zero-speed correction strategy to eliminate false displacement noise; S7. Calculate the displacement residual for each frame and sum it using a sliding window. S8. Compare the residual sum with the adaptive threshold to determine whether to trigger an alarm.

[0006] Furthermore, in step S2, the step of screening important features in the original data and performing standard normalization processing to construct the input model sequence specifically includes: S21. Select the four-dimensional feature forward velocity. yaw rate , Sine component of heading angle With cosine component Construct the four-dimensional single-frame feature vector at time t , used to characterize the instantaneous motion state of the vehicle at time t; S22. Using the Z-Score normalization method, the selected feature data is mapped to a standard normal distribution:

[0007] In the formula and These are the mean and standard deviation of the training set samples, respectively. It is the raw sensor feature data Dimensionless numerical values ​​after Z-core transformation; S23. Construct the input sequence using a sliding window with a time length of L. The sequence of independent samples read in a single iteration of the neural network is the batch size Batch_Size, that is, the tensor shape of the input model is (Batch_Size,L,4).

[0008] Furthermore, in step S3, the latitude and longitude coordinates are converted into the vehicle's real-time perceived displacement using the Havesing formula, as follows:

[0009] In the formula, The distance between two points on the Earth's surface (in meters); The Earth's radius (6378 km); and The latitude (radians) of the two points; and The longitude (in radians) of the two points.

[0010] Furthermore, step S4 specifically includes: S41. The model structure includes: input layer (temporal data of length L), first LSTM layer (64 neurons), dropout layer (dropout rate of 0.2), second LSTM layer (32 neurons), and fully connected layer (activation function is Linear). S42. Since displacement prediction is a regression problem, the mean squared error (MSE) is used as the loss function for model training:

[0011] In the formula, N is the sample size. This represents the actual displacement at time i before the attack.

[0012] Furthermore, in step S5, the injection of exponential displacement deviation to simulate a slow, covert deception attack signal includes: assuming If the attack begins at time t, then the spoofed GPS position is... Defined as the real location and cumulative drift sum:

[0013] In the formula, the drift amount It follows the following exponential growth law: .

[0014] Furthermore, in step S6, the displacement modulus predicted by the monitoring model is | When this value is less than the preset static threshold, the vehicle is determined to be stationary, the predicted displacement is forcibly set to zero, and the accumulation of residuals is stopped.

[0015] Furthermore, step S7 specifically includes: S71. Calculate the displacement predicted by the model at the current time t. Displacement sensed by GPS receiver Define instantaneous residual :

[0016] Under normal driving conditions It mainly consists of sensor noise and model fitting error, and follows a low-amplitude random distribution; however, under a spoofing attack, This will include systemic biases introduced by the attack; S72. To extract the attack signal from the noise, a time window of length W is introduced. At any time t, the cumulative sum of all historical instantaneous residuals within the window is calculated. : .

[0017] Furthermore, in step S8, the system maintains a "security buffer" in real time, and counts the accumulated residual S over a period of time before the attack occurs (or during the system initialization phase). t mean and standard deviation The alarm threshold is defined as:

[0018] In the formula, K is the safety factor. According to Chebyshev's inequality or the 3-Sigma criterion, this threshold can cover most noise fluctuations generated during normal driving; when the cumulative error S calculated in real time... t > This triggers a system alarm.

[0019] The present invention also provides a vehicle-mounted GPS slow spoofing detection system based on LSTM and residual spatiotemporal accumulation.

[0020] The beneficial effects of this invention are as follows: This invention addresses the covert, slow-speed traction spoofing threat faced by GPS systems in autonomous vehicles by proposing an intrusion detection method based on an LSTM dynamics model and residual spatiotemporal accumulation. By constructing a vehicle trajectory estimation model independent of satellite signals using IMU and CAN bus data, and combining a sliding window integration mechanism and a zero-speed correction strategy, this method effectively overcomes the shortcomings of traditional instantaneous detection methods, such as insensitivity to small positional deviations and interference from stationary drift. This invention achieves rapid response to attacks while maintaining a low false alarm rate, providing reliable intrusion detection technology support for improving the positioning security of autonomous driving systems.

[0021] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is an overall flowchart of the intrusion detection method based on the LSTM dynamics model and residual spatiotemporal accumulation of the present invention; Figure 2 This is a schematic diagram illustrating the generation of the training dataset for this invention; Figure 3 This is a schematic diagram of the device-side LSTM model training of the present invention; Figure 4 This is a diagram showing the state of a car before and after a simulated attack, as presented in this invention. Figure 5 This is an example diagram illustrating the cumulative displacement error of the sliding window before and after the deception attack of this invention. Detailed Implementation

[0023] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0024] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0025] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0026] See Figures 1-5 This invention provides a vehicle-mounted GPS slow-speed spoofing detection method based on LSTM and residual spatiotemporal accumulation. It utilizes inertial measurement unit (IMU) data and GPS latitude and longitude data from the KITTI public dataset to construct a long short-term memory (LSTM) vehicle dynamics model, calculating the vehicle's predicted displacement independent of the GPS signal. By calculating the Euclidean distance residual between the LSTM predicted displacement and the perceived displacement calculated from GPS latitude and longitude, a sliding window accumulation mechanism is used to amplify the small deviations generated by slow-speed covert spoofing. An adaptive threshold based on background noise statistical characteristics is then used for anomaly detection.

[0027] Figure 1 This is an overall flowchart of the intrusion detection method based on the LSTM dynamics model and residual spatiotemporal accumulation of the present invention. The steps of this method are as follows: S1. Collect vehicle inertial measurement unit (IMU) data, CAN bus data, and GPS observation data; S2. Select important features from the original data, perform standard normalization processing, and construct the input model sequence; S3. Convert latitude and longitude coordinates into real-time vehicle displacement using the Havessin formula; S4. Construct an LSTM model to estimate the predicted displacement for the next second using historical data frames; S5. Injecting exponential displacement deviation to simulate slow, covert deception attack signals; S6. Adopt a zero-speed correction strategy to eliminate false displacement noise; S7. Calculate the displacement residual for each frame and sum it using a sliding window. S8. Compare the residual sum with the adaptive threshold to determine whether to trigger an alarm.

[0028] Example First, in step S1, this invention selects a data acquisition platform equipped with a high-precision OXTS RT3003 integrated navigation system (GPS / IMU), capable of providing centimeter-level positioning accuracy up to 100Hz, and high-precision inertial measurement datasets, specifically multiple residential and city scene data segments from the KITTI dataset collected on the date "2011_09_26". These scenes contain a rich variety of vehicle start-stop, acceleration, deceleration, right-angle turns, and lane-changing behaviors, which, compared to the simple highway scene, can more comprehensively test the LSTM model's ability to fit complex vehicle dynamics.

[0029] like Figure 2 The diagram shown illustrates the feature selection and input vector construction of this invention. Specific implementation S2 includes the following steps: To enable the model to independently calculate trajectories even in the absence of GPS signals, the input features must originate solely from in-vehicle sensors (such as IMU and wheel speedometers). Through feature engineering selection, this invention selects to construct the four-dimensional single-frame feature vector at time t. This is used to characterize the instantaneous motion state of the vehicle at time t: Forward velocity The longitudinal speed, calculated by the vehicle's CAN bus or inertial measurement unit, reflects how fast the vehicle is moving.

[0030] Yaw angular velocity The angular velocity around the Z-axis, collected by a gyroscope, reflects the dynamic characteristics of vehicle steering.

[0031] Sine component of heading angle With cosine component To address the issue that relying solely on velocity and angular velocity cannot determine the absolute direction of a vehicle, this invention introduces a heading angle feature. This feature is decomposed into sine and cosine components and input into the network, aiming to allow the model to learn the velocity vector in a local planar coordinate system.

[0032] Because the input features (such as velocity in m / s and angle in rad) and the output label (displacement in m) differ significantly in units and numerical range, directly inputting them into the network can lead to slow or even non-convergence of gradient descent. This invention employs the Z-Score normalization method to map all feature data to a standard normal distribution with a mean of 0 and a variance of 1.

[0033] In the formula and These are the mean and standard deviation of the training set samples, respectively. It is the raw sensor feature data Dimensionless numerical values ​​after Z-core transformation.

[0034] The model uses a sliding window of length L to construct the input sequence. The sequence of independent samples read in a single iteration of the neural network is the batch size Batch_Size, that is, the tensor shape of the input model is (Batch_Size,L,4). In the sequence generation stage, a sliding window technique is used to construct samples. The time step is set to 20, which means that the displacement increment of the current frame relative to the previous frame is predicted using the sensor data sequence of the past 20 frames (about 2 seconds). In order to enhance the generalization ability of the model and prevent overfitting to specific routes, the training set is randomly shuffled before being input into the model, while the test set maintains a strict time series order to simulate real online detection scenarios.

[0035] In the original KITTI data, the GPS output is longitude, latitude, and altitude in a geodetic coordinate system. Since latitude and longitude are non-linear coordinates with vastly different numerical ranges, they are not directly applicable to neural network training. Therefore, this invention first converts the geodetic coordinates to a local northeastern metric (ENU) plane coordinate system with the data acquisition origin as the origin. In step S3, the Havesing formula is used to calculate the actual driving distance based on the latitude and longitude coordinates, which includes the following steps:

[0036] In the formula, The distance between two points on the Earth's surface (in meters); The Earth's radius (6378 km); and The latitude (radians) of the two points; and The longitude (in radians) of the two points.

[0037] like Figure 3 The diagram shown is a design diagram of the model network structure of the present invention. Specific implementation S4 includes the following steps: S41. The model structure includes: input layer (temporal data of length L), first LSTM layer (64 neurons), dropout layer (dropout rate of 0.2), second LSTM layer (32 neurons), and fully connected layer (activation function is Linear). S42. Since displacement prediction is a regression problem, the mean squared error (MSE) is used as the loss function for model training:

[0038] In the formula, N is the sample size. This represents the actual displacement at time i before the attack.

[0039] like Figure 4As shown, in this attack mode, the attacker does not instantly change the vehicle's position, but rather takes over the receiver tracking loop and gradually and maliciously modifies the positioning results with extremely small offset rates. To simulate this highly covert attack, an exponentially growing bias model was designed. In step S5, it is assumed that... If the attack begins at time t, then the spoofed GPS position is... Defined as the real location and cumulative drift sum:

[0040] In the formula, the drift amount It follows the following exponential growth law:

[0041] In the formula, A is the basic amplitude parameter (set to 0.05 in this invention), which controls the degree of concealment in the early stage of the attack. k is the growth rate parameter (set to 0.03 in this invention), which controls the speed at which the deception deviation diverges.

[0042] In step S6, the displacement modulus predicted by the monitoring model is | When the value is less than the preset static threshold (set to 0.05m), the vehicle is determined to be stationary, the predicted displacement is forcibly set to zero, and the accumulation of residuals is stopped.

[0043] like Figure 5 As shown, step S7 calculates the cumulative displacement error after the attack occurs, including the following steps: S71. Calculate the displacement predicted by the model at the current time t. Displacement sensed by GPS receiver Define instantaneous residual :

[0044] Under normal driving conditions It mainly consists of sensor noise and model fitting error, which follows a certain low-amplitude random distribution; however, under a deception attack, This will include systemic biases introduced by the attack; S72. To extract the attack signal from the noise, a time window of length W is introduced. At any time t, the cumulative sum of all historical instantaneous residuals within the window is calculated. :

[0045] The physical significance of this mechanism lies in the fact that random noise exhibits the characteristics of mutual cancellation or mean regression during the time integration process, thus the cumulative sum under normal conditions... It will remain at a relatively stable low level; while slow deception introduces a directional DC bias, which, under integration, will amplify even small deviations through continuous accumulation, leading to... It exhibits a significant monotonic growth trend. This greatly improves the sensitivity of the detection system in low signal-to-noise ratio environments.

[0046] In step S8, the system maintains a "security buffer" in real time, and counts the accumulated residual S over a period of time before the attack occurs (or during the system initialization phase). t mean and standard deviation The alarm threshold (Threshold) is defined as follows:

[0047] In the formula, K is the safety factor (taken as K=3). According to Chebyshev's inequality or the 3-Sigma criterion, this threshold can cover most noise fluctuations generated during normal driving. When the cumulative error S calculated in real time... t > This triggers a system alarm.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting slow-speed spoofing of vehicle-mounted GPS based on LSTM and residual spatiotemporal accumulation, characterized in that, The method specifically includes the following steps: S1. Collect vehicle inertial measurement unit (IMU) data, CAN bus data, and GPS observation data; S2. Select important features from the original data, perform standard normalization processing, and construct the input model sequence; S3. Convert latitude and longitude coordinates into real-time vehicle displacement using the Havessin formula; S4. Construct an LSTM model to estimate the predicted displacement for the next second using historical data frames; S5. Injecting exponential displacement deviation to simulate slow, covert deception attack signals; S6. Adopt a zero-speed correction strategy to eliminate false displacement noise; S7. Calculate the displacement residual for each frame and sum it using a sliding window. S8. Compare the residual sum with the adaptive threshold to determine whether to trigger an alarm.

2. The vehicle-mounted GPS slow spoofing detection method based on LSTM and residual spatiotemporal accumulation according to claim 1, characterized in that, In step S2, the step of screening important features in the original data and performing standard normalization to construct the input model sequence specifically includes: S21. Select the four-dimensional feature forward velocity. yaw rate , Sine component of heading angle With cosine component Construct the four-dimensional single-frame feature vector at time t , used to characterize the instantaneous motion state of the vehicle at time t; S22. Using the Z-Score normalization method, the selected feature data is mapped to a standard normal distribution: In the formula and These are the mean and standard deviation of the training set samples, respectively. It is the raw sensor feature data Dimensionless numerical values ​​after Z-core transformation; S23. Construct the input sequence using a sliding window with a time length of L. The sequence of independent samples read in a single iteration of the neural network is the batch size Batch_Size, that is, the tensor shape of the input model is (Batch_Size,L,4).

3. The vehicle-mounted GPS slow spoofing detection method based on LSTM and residual spatiotemporal accumulation according to claim 2, characterized in that, In step S3, the latitude and longitude coordinates are converted into the vehicle's real-time perceived displacement using the Havesing formula, as follows: In the formula, The distance between two points on the Earth's surface; The radius of the Earth; and The latitude of the two points; and The longitudes of the two points are [longitude].

4. The vehicle-mounted GPS slow spoofing detection method based on LSTM and residual spatiotemporal accumulation according to claim 3, characterized in that, Step S4 specifically includes: S41. The model structure includes: input layer, first LSTM layer, dropout layer, second LSTM layer, and fully connected layer. S42. Since displacement prediction is a regression problem, the mean squared error (MSE) is used as the loss function for model training: In the formula, N is the sample size. Let i be the actual displacement before the attack. This represents the displacement predicted by the model.

5. The vehicle-mounted GPS slow spoofing detection method based on LSTM and residual spatiotemporal accumulation according to claim 4, characterized in that, In step S5, the injection of exponential displacement deviation to simulate a slow, covert deception attack signal includes: assuming If the attack begins at time t, then the spoofed GPS position is... Defined as the real location and cumulative drift sum: In the formula, the drift amount It follows the following exponential growth law: A is the basic amplitude parameter, and k is the growth rate parameter.

6. The method for slow-speed spoofing detection of vehicle-mounted GPS based on LSTM and residual spatiotemporal accumulation according to claim 5, characterized in that, In step S6, the displacement modulus predicted by the monitoring model is | When this value is less than the preset static threshold, the vehicle is determined to be stationary, the predicted displacement is forcibly set to zero, and the accumulation of residuals is stopped.

7. The vehicle-mounted GPS slow spoofing detection method based on LSTM and residual spatiotemporal accumulation according to claim 6, characterized in that, Step S7 specifically includes: S71. Calculate the displacement predicted by the model at the current time t. Displacement sensed by GPS receiver Define instantaneous residual : Under normal driving conditions It mainly consists of sensor noise and model fitting error, and follows a low-amplitude random distribution; however, under a spoofing attack, This will include systemic biases introduced by the attack; S72. To extract the attack signal from the noise, a time window of length W is introduced. At any time t, the cumulative sum of all historical instantaneous residuals within the window is calculated. : 。 8. The method for slow-speed spoofing detection of vehicle-mounted GPS based on LSTM and residual spatiotemporal accumulation according to claim 7, characterized in that, In step S8, the system maintains a "security buffer" in real time, and counts the accumulated residual S over a period of time before the attack occurs. t mean and standard deviation The alarm threshold is defined as: In the formula, K is the safety factor. According to Chebyshev's inequality or the 3-Sigma criterion, this threshold can cover most noise fluctuations generated during normal driving; when the cumulative error S calculated in real time... t > This triggers a system alarm.

9. A vehicle-mounted GPS slow spoofing detection system based on LSTM and residual spatiotemporal accumulation, characterized in that, The system employs the method as described in any one of claims 1 to 8.

Images