An adaptive vehicle sensor attack detection method based on multi-modal fusion

Abstract

The application discloses a kind of self-adapting vehicle sensor attack detection methods based on multi-modal fusion.The application constructs cross-modal verification mechanism by multi-sensor space-time alignment and feature consistency check, so as to more accurately locate attack timestamp under complex interference and information incomplete condition and reduce the misjudgment rate.Integrating learning idea is used, and different detection models are complementary in different attack types, to avoid the performance degradation of single model in specific working condition or specific attack mode, to avoid missed detection or false alarm, and to improve the cross-scene generalization ability.In the threshold determination stage, the statistical distribution of normalized abnormal score is modeled by Beta distribution, and the threshold is adaptively adjusted based on online updating mechanism, so as to effectively reduce the false alarm rate while ensuring the detection sensitivity, and improve the availability of long-term online deployment.

Description

Technical Field

[0001] This invention belongs to the field of connected vehicle fleet communication security technology, and particularly relates to a method for detecting vehicle sensor attacks based on multimodal fusion and adaptive technology. Background Technology

[0002] With the widespread application of IoT technology, connected autonomous driving platooning has become an important technical means to improve road traffic safety. In this technical system, the integrity and accuracy of information exchange between vehicles within the platoon are key prerequisites for ensuring coordinated platoon control.

[0003] However, the open communication environment makes the convoy vulnerable to network attacks. Attacks targeting observation data can be divided into two categories: 1) Status observation delay attacks: Attackers extend the delay time for vehicles to acquire status information by launching DoS attacks (delay or packet loss) or MITM attacks (tampering with delay information). 2) Status observation modification attacks: Attackers directly modify the status observation values ​​acquired by the vehicles through FDI attacks (tampering with sensor data), MITM attacks (tampering with communication information), or replay information.

[0004] Although there are detection schemes based on statistical rules, traditional machine learning, and deep learning, they still have the following limitations when facing the above-mentioned highly concealed and dynamically changing network attacks: 1) The feature representation is simple and it is difficult to capture cross-domain coupling anomalies; 2) The model has poor dynamic adaptability and is difficult to cope with the shift in operating conditions.

[0005] Therefore, how to provide a multimodal detection method that can accurately identify the above-mentioned network attacks in real time and has online adaptive capabilities is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive vehicle sensor attack detection method based on multimodal fusion, which aims to solve the problems of single feature dimension, insufficient mining of time series correlation features, and poor dynamic adaptability of the model.

[0007] In a first aspect, embodiments of this application provide an adaptive vehicle sensor attack detection method based on multimodal fusion, comprising the following steps:

[0008] S1. Acquire the raw observation data of the connected autonomous driving platoon vehicles and the vehicle control inputs from the previous moment. Then, based on the data-driven state estimation model, model the vehicle state evolution process to obtain state estimates.

[0009] S2. Construct a time-sliding window and input the original observation data, state estimates, and control input sequences within the window into a causal dilated convolutional network (TCN) to extract local temporal dependency features. Simultaneously, calculate the residual features between the original observation data and the state estimates within the window, and construct consistency verification features based on the consistency of vehicle kinematics and the consistency relationship between visual perception state.

[0010] S3. Calculate the statistics (including mean, variance, maximum and minimum) of the original observation data and local time-dependent features within the sliding window. Then, concatenate and fuse the local time-dependent features, residual features, consistency check features, and statistics to generate a multimodal feature vector.

[0011] S4. Training the anomaly detection model

[0012] Frequency domain augmented variational autoencoder (FVAE), deep support vector data description (Deep SVDD), and a detection network based on prototype representation and memory update mechanism (MemoryBank) were trained separately. After training, anomaly evaluation of multimodal feature vectors was performed by each model, and initial anomaly scores were generated.

[0013] S5. Based on the trained anomaly detection model, construct a multi-channel integrated detection framework, calculate the anomaly score of the multimodal feature vector of each channel when subjected to different types of network attacks, and finally perform dynamic weighted fusion to obtain a comprehensive anomaly score.

[0014] S6. Based on the obtained comprehensive anomaly score, construct an adaptive threshold adjustment mechanism based on Beta distribution, determine the dynamic threshold, and perform anomaly judgment.

[0015] In one possible implementation, the data-driven state estimation model employs a prediction-update two-stage architecture, specifically including:

[0016] First, construct the vehicle state vector. With observation vector The state vector is represented as:

[0017]

[0018] in, For the vehicle's global coordinate information, For heading angle, For longitudinal velocity, Let be the yaw rate. It is expressed using sin and cos trigonometric functions to avoid abrupt changes in angle.

[0019] Observation vector This represents the raw observation data collected by the vehicle's onboard sensors at time t. When the vehicle is attacked at time t, the observation vector may exhibit anomalies such as sensor zeroing, increased measurement noise, or drift.

[0020] Then, a predictor generates a priori state estimate based on historical multi-sensor sequences. The predictor uses a Long Short-Term Memory (LSTM) network to fuse features from the historical multi-sensor sequences and uses a masking mechanism to suppress abnormal or malfunctioning sensor features. Subsequently, the predictor outputs the vehicle's motion increment in the body coordinate system and transforms this motion increment to the global coordinate system based on the vehicle's current heading angle to obtain the vehicle's prior state in the global coordinate system.

[0021] Secondly, the prior state is corrected by an updater to obtain the posterior state estimate of the vehicle. The specific operation of the updater is as follows: First, a feature vector including the state prediction residual and the observation residual is constructed, where the state prediction residual is the difference between the prior state and the vehicle state vector at the current time, and the observation residual is the difference between the current observation value and the prior state. Then, a Kalman filter Kalman gain matrix is ​​dynamically generated from the above feature vector using a gated recurrent unit (GRU), and the posterior state estimate is calculated based on the Kalman filter update formula.

[0022] Finally, the accuracy of vehicle state estimation is improved by minimizing the loss function. This loss function consists of the weighted mean square error between the posterior state and the observation vector, the weighted mean square error between the prior state and the observation vector, and a smoothing constraint term for the heading angle change. After the model is trained, a vehicle state estimation model is obtained, and the state estimates of vehicles in the platoon are calculated based on this model.

[0023] In one possible implementation, the construction of the consistency verification feature specifically involves the following steps:

[0024] First, time consistency features are constructed from three dimensions, specifically composed of "stepping ratio features, multi-vehicle synchronization gap deviation features, and sensor hysteresis features":

[0025] (1) Step ratio characteristics: Define the step ratio Where i represents the sampling index in the sliding window, The difference in timestamps between adjacent sensors. This represents the time difference between adjacent systems. The mean offset, minimum offset, standard deviation, and stagnation rate of the step ratio within the sliding window are extracted and concatenated to form a four-dimensional timestamp step ratio feature.

[0026] (2) Multi-vehicle synchronization gap deviation characteristics: The synchronization gap between the timestamps of sensors of two adjacent vehicles is defined as... .in This represents the sensor timestamp at the i-th sampling time while following the vehicle. This represents the sensor timestamp of the preceding vehicle at the i-th sampling moment. Under normal communication conditions, this synchronization gap should remain relatively stable, but will exhibit abnormal jumps when a vehicle suffers a delay or replay attack. The aforementioned synchronization gap between two vehicles is used to characterize the time synchronization relationship between any two vehicles. For a convoy system containing multiple vehicles, the synchronization gap between each vehicle and the preceding vehicle is calculated separately, thus forming a multi-vehicle synchronization gap sequence. Then, within a sliding time window, the current synchronization gap is compared with the median of the historical synchronization gap sequence within the window, and its deviation is calculated as the multi-vehicle synchronization gap deviation feature. This deviation is close to 0 under normal circumstances, but increases significantly when an attack occurs.

[0027] (3) Sensor hysteresis characteristics: defining the hysteresis quantity This reflects the difference between the system time and the sensor timestamp, where... This represents the timestamp of the corresponding sensor. Under normal conditions, due to continuous sensor updates, the hysteresis is approximately constant. When sensor data transmission is obstructed, system time continues to advance while the sensor timestamp remains stagnant, causing the hysteresis to increase linearly with time. The hysteresis at consecutive moments is represented as... The hysteresis time series is constructed, and the mean, standard deviation, and hysteresis difference between the beginning and end of the window are extracted from the hysteresis window series. These are then spliced ​​together to obtain the sensor hysteresis characteristics.

[0028] Then, based on image data collected by the vehicle-mounted camera following the vehicle, and combining temporal state vectors and bounding box geometric features, a multimodal fusion regression model is used to infer the state of the preceding vehicle. This model consists of three parts: an image branch that crops and normalizes the preceding vehicle image and extracts visual features using ResNet18; a state sequence branch that extracts the vehicle motion state sequence of the past T frames and encodes it into temporal state features using LSTM; and a bounding box sequence branch that extracts the 6-dimensional bounding box geometric feature sequence of the past T frames and encodes it into temporal geometric features using LSTM. The output features of the three branches are concatenated and regressed using a multilayer perceptron to obtain a 3-dimensional visual inference value. The visual inference value is compared with the observation vector, and the residual is calculated. Statistical features of the residual sequence are extracted within a sliding time window, including the window mean and maximum deviation, as visual-kinematic cross-consistency features.

[0029] Secondly, based on the position coordinates of the preceding and following vehicles obtained from the Global Navigation Satellite System (GNSS), the GNSS-estimated distance is converted to metric distance using an equidistant rectangular approximation. Then, the GNSS-estimated distance is compared with the radar-measured distance, and the absolute difference between the two is calculated. This difference is used as the GNSS-radar consistency residual characteristic.

[0030] Finally, the average absolute value of the speed of the two vehicles within the sliding window is calculated as the speed reference, and the average value of the accelerator pedal opening is used as the control reference. The speed reference is linearly mapped to the [0,1] interval to obtain the speed gating factor. At the same time, the control reference is normalized to obtain the control gating factor. The maximum value of the two is taken as the motion gating factor.

[0031] The time consistency features, visual-kinematic cross-consistency features, motion gating factors, and GNSS-radar consistency residuals mentioned above are Z-score normalized and then concatenated in a preset order to form a consistency verification feature vector.

[0032] In one possible implementation, the initial anomaly score is calculated based on a frequency-domain enhanced variational autoencoder (FVAE), as follows:

[0033] First, a causal dilated convolutional network (TCN) is used to encode the historical multimodal feature sequence of length L at time t, extracting temporal context features that characterize local temporal changes and global evolution trends. Simultaneously, a fast Fourier transform is performed on the historical multimodal feature sequence to extract frequency domain features that characterize overall spectral regularity and local frequency abrupt changes. Then, the temporal context features and frequency domain features are fused, and the fused context representation is used as conditional input to construct the prior and posterior distributions of the latent variables. The prior distribution is predicted from the historical multimodal feature sequence, while the posterior distribution is inferred from the historical multimodal feature sequence and the current multimodal feature vector.

[0034] Then, a hierarchical decoding strategy is adopted to divide the latent variables into physical state sub-vectors and temporal consistency sub-vectors, and reconstruct them through the physical state decoder and temporal consistency decoder respectively to obtain the reconstructed distribution parameters of the observed features and the reconstructed values ​​of the global spectral features.

[0035] Subsequently, an initial anomaly score is constructed. .in, To reconstruct the negative log-likelihood in the hierarchical time domain; This represents the divergence between the posterior and prior distributions. for The weight parameters, It represents the mean square error between the reconstructed values ​​of the frequency domain features of the historical sequence and the global spectral features, and is used to characterize frequency domain structural anomalies.

[0036] In one possible implementation, the initial anomaly score is calculated based on Deep Support Vector Data Description (Deep SVDD), as follows:

[0037] First, the multimodal feature vectors are used as input to obtain the latent representation through a deep encoding network.

[0038] Then, we define the sphere center vector, construct the training loss function, and constrain the normal samples to form a compact distribution in the latent space.

[0039] Finally, the distance from the sample representation to the center of the hypersphere is calculated in the latent space as the initial anomaly score. In one possible implementation, a detection network based on prototype representation and memory update mechanism maps multimodal feature vectors to the latent space through an encoder, constructs a normal behavior prototype library, and calculates the anomaly score during the detection phase by calculating the distance between the sample latent representation and the prototype.

[0040] In one possible implementation, a multi-channel integrated detection framework is constructed based on three models: a frequency domain variational autoencoder (FVAE), a deep support vector description model (Deep SVDD), and a detection network based on prototype representation and memory update mechanism (MemoryBank).

[0041] First, the three models described above are used to analyze the input time series data and output the anomaly score at the current time. Then, the anomaly scores output by the three models are normalized using Min-Max. A heterogeneous dual-channel judgment mechanism is designed to generate corresponding anomaly scores for time series structure anomalies and amplitude tampering anomalies, constructing a time series structure anomaly bypass channel and an amplitude tampering anomaly channel. The time series structure anomaly bypass channel is for delay-type and time series structure disruption-type anomalies, and its anomaly score is directly given by the normalized FVAE anomaly score; the amplitude tampering anomaly channel is for amplitude tampering-type anomalies, and its anomaly score is obtained by weighted fusion of the normalized anomaly scores from the Deep SVDD and MemoryBank models. Finally, the anomaly scores of the two channels are weighted and fused to obtain the comprehensive anomaly score.

[0042] In one possible implementation, the adaptive threshold adjustment mechanism based on the Beta distribution is as follows:

[0043] A Beta distribution is used to statistically model the obtained composite anomaly scores to determine the anomaly detection threshold. Historical composite anomaly score sequences recorded within a sliding time window are utilized. The sample fits the Beta distribution Set confidence quantiles. The threshold is defined as: in Let quantile function be the Beta distribution quantile function. An online adaptive adjustment of the threshold is employed at the current time. Collect recent A window is formed by a score that provides normal feedback. ,right Refitting Beta distribution parameters And calculate the dynamic threshold When the integrated detection framework receives feedback that the current sample is normal, it adds the corresponding comprehensive anomaly score to the historical comprehensive anomaly score sequence and updates the threshold; if there is no normal feedback, it does not update.

[0044] Secondly, embodiments of this application provide an electronic device, including a processor and a memory;

[0045] The memory is used to store computer programs.

[0046] When the processor executes the program stored in the memory, it implements any of the adaptive vehicle sensor attack detection methods described in this application.

[0047] Thirdly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the adaptive vehicle sensor attack detection methods described in this application.

[0048] Fourthly, embodiments of this application provide a computer program product containing instructions that, when run on a computer, cause the computer to execute any of the adaptive vehicle sensor attack detection methods described in this application.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] This invention constructs a cross-modal verification mechanism through multi-sensor spatiotemporal alignment and feature consistency checks, thereby enabling more accurate location of attacked timestamps and reducing false positive rates under complex interference and incomplete information conditions.

[0051] Based on vehicle kinematics and multi-sensor alignment, this invention constructs physical constraint features that reflect timing consistency, such as timestamp step ratio, multi-vehicle synchronization gap, and sensor hysteresis. These features can effectively identify abnormal behaviors such as replay attacks, time tampering, and short-term delay injection.

[0052] This invention employs a frequency-domain enhanced conditional VAE to construct an anomaly scoring mechanism. It extracts global and segmented spectral features through FFT and fuses them with time-domain features to form latent variable distribution constraints, enabling the model to utilize both time-domain evolution patterns and frequency-domain structural information. This significantly enhances the ability to identify periodic disturbances, spectral morphology anomalies, and sudden anomalies.

[0053] This invention further adopts the concept of ensemble learning, which complements and integrates the sensitivity advantages of different detection models on different attack types, avoiding the missed detections or false alarms caused by the performance degradation of a single model under specific working conditions or specific attack modes, and improving the cross-scenario generalization ability.

[0054] In the threshold determination stage, this invention models the statistical distribution of normalized anomaly scores using a Beta distribution and adaptively adjusts the threshold based on an online update mechanism. This effectively reduces the false alarm rate while ensuring detection sensitivity and improves the availability of long-term online deployment. Attached Figure Description

[0055] The accompanying drawings are provided to enhance understanding of the invention and form part of the specification. These drawings illustrate the invention but should not be construed as limiting the invention in any way. In the drawings:

[0056] Figure 1 This is a flowchart of a method according to an embodiment of the present invention. Detailed Implementation

[0057] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings. The accompanying drawings are simplified schematic diagrams, intended to illustrate the basic structure and working principle of the present invention. Only the main components related to the present invention are shown, and the remaining unshown parts do not constitute a limitation on the present invention.

[0058] refer to Figure 1 An adaptive vehicle sensor attack detection method based on multimodal fusion specifically includes:

[0059] S1. Obtain raw observation data collected by onboard sensors from connected autonomous driving platoons during operation. The vehicle control input from the previous moment The raw observation data includes vehicle motion and operational status information acquired by each vehicle in the convoy via inertial sensors, a global navigation satellite system, and a vehicle bus, as well as distance information between adjacent vehicles. Then, based on a data-driven state estimation model, the vehicle state evolution process is modeled to obtain state estimates robust to sensor anomalies and noise disturbances. .

[0060] This implementation uses the CAN bus to collect multi-dimensional raw observation data for each vehicle in the convoy, including global coordinate position, attitude, control input, speed, and vehicle spacing. Due to timestamp offsets between sensors, time alignment processing is performed on each sensor.

[0061] S2. Construct a time-sliding window, input the original observation data, state estimates, and control input sequences within the window into a causal dilated convolutional network (TCN), and extract local temporal dependency features. Simultaneously, the residual characteristics between the original observation data and the state estimate within the window are calculated, and consistency verification features are constructed based on the consistency of vehicle kinematics and the consistency relationship between visually perceived state. .

[0062] S3. Calculate the statistics of the original state data and local temporal dependency features within the sliding window, including mean, variance, maximum, and minimum values. Then, concatenate and fuse the local temporal dependency features, residual features, consistency check features, and statistics to generate a multimodal feature vector. .

[0063] S4. Training the anomaly detection model

[0064] Frequency domain augmented variational autoencoder (FVAE), deep support vector data description (Deep SVDD), and a detection network based on prototype representation and memory update mechanism (MemoryBank) were trained separately. After training, the multimodal feature vectors were processed by each model. Perform anomaly assessment and generate initial anomaly scores.

[0065] S5. A multi-channel integrated detection framework is constructed based on the trained frequency-domain enhanced variational autoencoder, deep support vector data description, and a detection network based on prototype representation and memory update mechanism. Each detection channel corresponds to an independent anomaly detection model. The anomaly scores of the multimodal feature vectors of each channel are calculated when subjected to different types of network attacks. Finally, dynamic weighted fusion is performed to obtain the comprehensive anomaly score. .

[0066] S6. Based on the obtained comprehensive anomaly score, construct an adaptive threshold adjustment mechanism based on the Beta distribution. Record historical comprehensive anomaly scores within a time sliding window, model the comprehensive anomaly score of unattacked data using the Beta distribution, and then determine the dynamic threshold by calculating the 99th percentile of the anomaly score using the cumulative distribution function of the Beta distribution. Perform anomaly detection. When If the condition is met, it is considered abnormal; otherwise, it is considered normal.

[0067] In one possible implementation, the data-driven state estimation model employs a prediction-update two-stage architecture, specifically including:

[0068] First, construct the vehicle state vector. With observation vector The state vector is represented as:

[0069]

[0070] in, For the vehicle's global coordinate information, For heading angle, For longitudinal velocity, Let be the yaw rate. It is expressed using sin and cos trigonometric functions to avoid abrupt changes in angle.

[0071] Observation vector This represents the raw observation data collected by the vehicle's onboard sensors at time t. When the vehicle is attacked at time t, the observation vector may exhibit anomalies such as sensor zeroing, increased measurement noise, or drift.

[0072] Then, a predictor generates a priori state estimate based on the historical multi-sensor sequence. This historical multi-sensor sequence consists of multi-source sensor observation data collected by the vehicle at multiple consecutive time steps. The multi-source sensor observation data includes triaxial acceleration and triaxial angular velocity data provided by inertial sensors (inertial sensor data), latitude and longitude information provided by the Global Navigation Satellite System (GNSS) (GNSS sensor data), and vehicle position and velocity information provided by the vehicle bus (VBS) (VBS data). The predictor uses a Long Short-Term Memory (LSTM) network to perform feature fusion on the above multi-source sensor observation data and uses a masking mechanism to suppress abnormal or malfunctioning sensor features. Subsequently, the predictor outputs the vehicle's motion increment in the body coordinate system and transforms this motion increment to the global coordinate system based on the vehicle's current heading angle to obtain the vehicle's prior state in the global coordinate system. .

[0073] in, This is a historical sequence of inertial sensor data collected within the time interval t-H+1 to t. This is a historical sequence of Global Navigation Satellite System sensor data collected within the time interval t-H+1 to t. This is a historical sequence of vehicle bus data collected within the time interval t-H+1 to t, where H is the length of the time window. Let be the posterior state estimate at time t-1. For predicting network models.

[0074] Secondly, the prior state is corrected by an updater to obtain the posterior state estimate of the vehicle. The specific operation of the updater is as follows: First, a feature vector including the state prediction residual and the observation residual is constructed, where the state prediction residual is the difference between the prior state and the vehicle state vector at the current time, and the observation residual is the difference between the current observation value and the prior state. Then, a Kalman filter Kalman gain matrix is ​​dynamically generated from the above feature vector using a gated recurrent unit (GRU). And calculate the posterior state estimate based on the Kalman filter update formula. , ,in For state prediction residuals, To observe the residuals, This is the hidden state from the previous moment.

[0075] Finally, the accuracy of vehicle state estimation is improved by minimizing the loss function. This loss function mainly consists of three parts: first, the weighted mean square error between the posterior state and the observation vector; second, the weighted mean square error between the prior state and the observation vector; and third, a smoothing constraint term for the heading angle change.

[0076]

[0077] WMSE is a 6-dimensional weighted mean square error. It eliminates gradient update bias caused by different magnitudes of dependent variables by assigning differentiated weights to variables of different dimensions, thereby improving the convergence stability of the model. Here, λ is the smoothing constraint term for the heading angle change, and λ is the balance coefficient. After the model is trained, a vehicle state estimation model is obtained, and the state estimate of the nth vehicle in the platoon at time t is calculated based on this model. = ,in Let represent the posterior state estimate of the nth vehicle at time t.

[0078] This implementation method, by degrading sensor data and utilizing a time window, can accurately reconstruct the vehicle's true robust state estimate when a certain sensor fails, laying the foundation for subsequent anomaly detection.

[0079] In one possible implementation, the construction of consistency verification features The specific operations include:

[0080] Firstly, under normal communication conditions, the sensor timestamps should advance synchronously with the system time, and the time difference between multiple vehicles in the formation should remain relatively stable. To this end, time consistency characteristics are constructed from three dimensions, specifically composed of "stepping ratio characteristics, multi-vehicle synchronization gap deviation characteristics, and sensor hysteresis characteristics."

[0081] (1) Step ratio characteristics: Define the step ratio Where i represents the sampling index in the sliding window, The difference in timestamps between adjacent sensors. This represents the time difference between adjacent systems. The step ratio measures the rate at which the sensor timestamp progresses relative to system time. Extract the mean offset, minimum offset, standard deviation, and stall ratio (representing the proportion of samples with a step ratio below a stall threshold within the sliding window) of the step ratio within the sliding window. The time is determined to be a time stamp standstill, where (The threshold for determining stagnation) is used to splice together a four-dimensional timestamp step ratio feature.

[0082] (2) Multi-vehicle synchronization gap deviation characteristics: The synchronization gap between the timestamps of sensors of two adjacent vehicles is defined as... in This represents the sensor timestamp at the i-th sampling time while following the vehicle. This represents the sensor timestamp of the preceding vehicle at the i-th sampling moment. Under normal communication conditions, this synchronization gap should remain relatively stable, but it will exhibit abnormal jumps when a vehicle suffers a delay or replay attack. The aforementioned synchronization gap between two vehicles is used to characterize the time synchronization relationship between any two vehicles. For a convoy system containing multiple vehicles, the synchronization gap between each vehicle and the preceding vehicle is calculated separately, thus forming a multi-vehicle synchronization gap sequence. Then, within a sliding time window, the current synchronization gap is compared with the median of the historical synchronization gap sequence within the window, and its deviation is calculated as a multi-vehicle synchronization gap deviation feature, used to characterize abnormal changes in the time synchronization relationship between vehicles. This deviation is close to 0 under normal circumstances, but increases significantly when an attack occurs.

[0083] (3) Sensor hysteresis characteristics: defining the hysteresis quantity This reflects the difference between the system time and the sensor timestamp, where... This represents the timestamp of the corresponding sensor. Under normal conditions, due to continuous sensor updates, the hysteresis is approximately constant. When sensor data transmission is obstructed, system time continues to advance while the sensor timestamp remains stagnant, causing the hysteresis to increase linearly with time. The hysteresis at consecutive moments is represented as... Construct a time series of lag values, and extract the mean, standard deviation, and difference in lag values ​​between the first and last lag values ​​of the window from the lag window series. ,in This represents the lag of the latest time within the sliding window. (This represents the hysteresis at the start of the window), and the hysteresis characteristics of the sensor are obtained by splicing them together.

[0084] Then, based on the image data collected by the onboard camera of the following vehicle, combined with the temporal state vector and bounding box geometric features, the state of the preceding vehicle is inferred through a multimodal fusion regression model. This model consists of three parts: an image branch that crops and normalizes the image of the preceding vehicle, and then extracts visual features through ResNet18. Capture the appearance and relative scale information of the vehicle in front; extract the vehicle motion state sequence from the past T=10 frames using the state sequence branch. , where single-frame vector The state features are 9-dimensional, including the position, speed, acceleration of the following vehicle and the vehicle in front, as well as the relative distance between the two vehicles. These are then encoded into temporal state features using LSTM. Capture motion context. The bounding box sequence branch extracts a sequence of 6D bounding box geometric features from the past T=10 frames. ,in The 6-dimensional bounding box set features include the coordinates of the center point of the detection box, its width and height, area, and aspect ratio. This sequence is encoded into geometric temporal features using LSTM. The output features of the three branches are concatenated. The 3D visual inference value was then obtained through multilayer perceptron regression. ,in For estimating the relative speed of the vehicle in front, For the estimation of the longitudinal acceleration of the vehicle, For vehicle spacing estimation, the visually inferred values ​​are compared with the observed vectors, and the residuals are calculated. Statistical features of the residual sequence, including the window mean and maximum deviation, are extracted within the sliding time window as visual-kinematic cross-consistency features.

[0085] Secondly, based on the position coordinates of the preceding and following vehicles obtained from the Global Navigation Satellite System, the GNSS-calculated distance is converted to metric distance using an equidistant rectangular approximation. ,in For the Earth's radius, Latitude difference This is the difference in longitude. The average latitude is used. Then, the GNSS-estimated distance is compared with the radar-measured distance, the absolute difference between the two is calculated, and this difference is used as the GNSS-radar consistency residual characteristic.

[0086] Finally, considering that the distance between vehicles and the relative speed naturally change very little when the vehicles are stationary or at low speeds, the average absolute value of the speed of the two vehicles within the sliding window is calculated as a speed reference. And the average accelerator pedal opening value as a control reference The velocity reference is linearly mapped to the [0,1] interval to obtain the velocity gating factor. Meanwhile, the control reference is normalized to obtain the control gating factor. The maximum value of the two is taken as the motion gating factor. .when (in When the threshold for determining motion state is reached, the vehicle is determined to be stationary or at low speed, and the time consistency feature is set to zero; when... The time indicates that the vehicle is in motion, maintaining the time consistency feature unchanged. The aforementioned time consistency features, visual-kinematic cross-consistency features, motion gating factor, and GNSS-radar consistency residuals are Z-score normalized. After normalization, they are concatenated in a preset order to form a consistency verification feature vector. .

[0087] This implementation constructs consistency verification features based on the constraint relationship between vehicle kinematic consistency and visual perception state consistency. The system extracts three types of features from the time dimension—timestamp step consistency, multi-vehicle synchronization gap consistency, and sensor lag consistency—using a sliding window to characterize communication anomalies such as time progression anomalies, synchronization abrupt changes, and lag accumulation. Simultaneously, a multimodal fusion regression model is established based on following vehicle camera images, historical state sequences, and target bounding box geometric sequences to infer the key states of the preceding vehicle. The visual inference values ​​are compared with onboard sensor measurements to obtain residual statistics, forming a visual-kinematic cross-consistency feature. Furthermore, satellite positioning is used to calculate vehicle-to-vehicle distance, which is compared with radar ranging to obtain residual statistics, forming a GNSS-radar cross-consistency feature. Finally, a motion gating mechanism is introduced to suppress easily frozen physical quantity-related features in stationary or low-speed scenarios and to maintain original features in moving scenarios. This improves anomaly detectability and reduces false alarms in low-speed conditions under complex communication disturbances and multi-source sensing errors.

[0088] In one possible implementation, the initial anomaly score is calculated based on a frequency-domain enhanced variational autoencoder (FVAE). The details are as follows:

[0089] At any moment , with a length of The historical multimodal feature sequence is denoted as First, we utilize a causal dilated convolutional network (TCN) to... Encoding is performed to extract temporal contextual features that characterize local temporal changes and global evolutionary trends; simultaneously, A Fast Fourier Transform (FFT) is performed to extract frequency domain features representing the overall spectral regularity and local frequency abrupt changes. Subsequently, the time-domain context features and frequency domain features are concatenated to obtain a fused context representation, which is used as conditional information to constrain the distribution of latent variables. This fused context representation serves as the conditional input. Based on this, latent variables are constructed. prior distribution With posterior distribution The prior distribution consists of historical multimodal feature sequences. The prediction is obtained, while the posterior distribution is derived from the historical multimodal feature sequence. With the current multimodal feature vector This was deduced through joint inference.

[0090] Then, a layered decoding strategy is adopted to decode the latent variables. The data is divided into two sub-vectors: a physical state sub-vector and a temporal consistency sub-vector, each reconstructed using an independent decoder. The physical state decoder reconstructs the current multimodal observation features, and its output is the reconstructed distribution parameters of the observation features, including the reconstructed mean. With variance The time-consistency decoder is used to reconstruct the frequency domain structure features of a time series, and its output is the reconstructed value of the global spectral features. .

[0091] Subsequently, an initial anomaly score is constructed. .in, The negative log-likelihood is reconstructed for hierarchical temporal domain and used to measure the current multimodal features. With reconstructed mean The reconstruction error between them (the calculation formula is: ,in and These correspond to the reconstruction errors of the physical state subvector and the temporal consistency subvector, respectively. and (For the corresponding weight parameters). This represents the divergence between the posterior and prior distributions. for The weight parameters, It represents the mean square error between the reconstructed values ​​of the frequency domain features of the historical sequence and the global spectral features, and is used to characterize frequency domain structural anomalies.

[0092] Finally, the calculated comprehensive anomaly score Compared with a preset fixed threshold, if If the threshold is exceeded, an anomaly is determined to exist at the current moment.

[0093] In one possible implementation, the initial anomaly score is calculated based on Deep Support Vector Data Description (Deep SVDD). The details are as follows:

[0094] First, the multimodal feature vectors As input, via parameters Deep coding networks yield latent representations: in This represents the dimension of the potential space. It consists of a multi-layer fully connected network and a residual connection structure.

[0095] Then define the center vector of the sphere. in, This represents the number of training samples and is updated using momentum during training.

[0096] A training loss function is then constructed to constrain normal samples to form a compact distribution in the latent space. The normal sample loss term minimizes the squared distance from the sample to the center of the sphere. Where B represents the batch sample size.

[0097] Finally, the initial outlier score is calculated by taking the distance from the sample representation to the center of the hypersphere in the latent space. The calculated initial anomaly score Compared with a preset fixed threshold, if If the threshold is exceeded, an anomaly is determined to exist at the current moment.

[0098] In one possible implementation, a detection network based on prototype representation and memory update mechanisms uses an encoder to process multimodal feature vectors. Mapping to the latent space and constructing a normal behavior prototype library, and calculating anomaly scores during the detection phase by calculating the distance between the latent representation of the sample and the prototype.

[0099] The specific operation is as follows: input multimodal feature vector via encoder Mapping to latent space and through the decoder Reconstructing input The mean squared error between the input features and the reconstructed features is calculated as the reconstruction error. Constructing a space containing A prototype library of normal prototypes Each prototype is statistically derived from the latent representation of normal samples during the training phase and represents a typical normal behavior pattern. The prototype distance is obtained by calculating the Euclidean distance between the latent representation and the nearest prototype. Finally, the prototype distance and reconstruction error are weighted and fused to obtain the anomaly score. ,in, These are the weighting coefficients. The calculated initial anomaly scores... Compared with a preset fixed threshold, if If the threshold is exceeded, an anomaly is determined to exist at the current moment.

[0100] In this implementation, the calculation of the anomaly score employs three optional mechanisms: First, based on frequency domain enhancement conditional variational autoencoders, for lengths of... First, historical multimodal feature sequences are temporally encoded and frequency-enhanced. A comprehensive anomaly score is obtained under the joint constraints of three types of errors: temporal reconstruction, prior and posterior differences, and frequency reconstruction. Second, based on deep support vector data description, a sliding window feature is mapped to the latent space through a deep encoding network, using a hypersphere formed by normal samples as the description boundary, and the distance from the sample to the center of the sphere as the anomaly score. Third, based on a memory prototype network, features are encoded into the latent space and a normal behavior prototype library is established. Simultaneously, the input reconstruction error and its distance from the nearest normal prototype are calculated, and the two are weighted and fused to obtain the anomaly score. Finally, the anomaly score is compared with a preset fixed threshold; if it exceeds the threshold, an anomaly is determined to exist at the current time.

[0101] In one possible implementation, a multi-channel integrated detection framework is constructed based on three models: a frequency-domain variational autoencoder (FVAE), a deep support vector description model (Deep SVDD), and a detection network based on prototype representation and memory update mechanisms (MemoryBank). First, the input time-series data is analyzed using these three models, and corresponding anomaly scores are output at the current time t. Then, the anomaly scores output by the three models are normalized using Min-Max. Because network attacks may manifest as different forms of anomalies in the data, a heterogeneous dual-channel judgment mechanism is designed to generate corresponding anomaly scores for potential time-series structure anomalies and amplitude tampering anomalies, constructing a time-series structure anomaly bypass channel and an amplitude tampering anomaly channel. The time-series structure anomaly bypass channel targets delay-type and time-series structure disruption-type anomalies, and its anomaly score is directly given by the normalized FVAE anomaly score. The amplitude tampering anomaly channel targets amplitude tampering-type anomalies, and its anomaly score is obtained by weighted fusion of the normalized anomaly scores from the Deep SVDD and MemoryBank models. Finally, the anomaly scores from the two channels are weighted and fused to obtain a comprehensive anomaly score. .

[0102] In one possible implementation, the adaptive threshold adjustment mechanism based on the Beta distribution is as follows:

[0103] The comprehensive anomaly score obtained using the Beta distribution Statistical modeling is performed to determine the anomaly detection threshold. The sequence of historical composite anomaly scores recorded within a sliding time window is defined as follows: The Beta distribution is fitted using samples from this sequence. Based on this, confidence quantiles are set. The threshold is defined as: in Let quantile function be the Beta distribution quantile function. An online adaptive adjustment of the threshold is employed at the current time. Collect recent A window is formed by a score that provides normal feedback. ,right Refitting Beta distribution parameters And calculate the dynamic threshold When the integrated detection framework receives feedback that the sample at that moment is normal, it adds the corresponding comprehensive anomaly score to the historical comprehensive anomaly score sequence and updates the threshold; if there is no normal feedback, it does not update.

[0104] This invention mainly comprises two phases: model training and test deployment. In the training phase, a data-driven approach is first used to achieve robust estimation of vehicle states. Then, using historical data collected from the fleet, three heterogeneous anomaly detectors are trained separately, and an integrated detection framework is constructed to fuse the outputs of each detector. In the testing phase, the system is deployed in a real-world environment. Due to the sliding window prediction mechanism, the system undergoes a warm-start process, starting prediction only after the sliding window sequence is constructed. Subsequently, the integrated model is used to analyze anomaly attacks on data with the current timestamp. Simultaneously, based on a Beta distribution fitted to historical anomaly scores, the decision thresholds of each base predictor in the integrated framework are dynamically and adaptively adjusted to improve detection accuracy and robustness.

[0105] To fully evaluate the system's performance, this invention constructs a two-tiered evaluation system covering both window-level and event-level metrics, ensuring horizontal comparability. At the window-level, metrics such as recall, precision, and F1 score are used to comprehensively measure the model's classification performance and robustness in imbalanced scenarios, supplemented by AUROC and AUPRC to evaluate the model's inherent discriminative ability. At the event-level, the focus is on examining event recall and detection time-to-delay (TTD) to ensure timely detection and coverage of attack events.

[0106] The traditional method specifically refers to a type of support vector machine, and uses the same extraction method as the present invention in the feature extraction process.

[0107] Table 1 Comparison of performance between the inventive method and the conventional method

[0108]

[0109] This application also provides an electronic device, including a processor and a memory.

[0110] The memory is used to store computer programs.

[0111] When the processor executes a program stored in the memory, it implements any of the methods described in this application.

[0112] In one possible implementation, the electronic device of this application embodiment further includes a communication interface and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0113] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc.

[0114] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0115] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0116] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0117] In another embodiment provided in this application, a computer-readable storage medium is also provided, which stores a computer program that, when executed by a processor, implements any of the methods described in this application.

[0118] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to perform any of the methods described in this application.

[0119] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0120] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0121] The various embodiments in this specification are described in a related manner. Each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other.

[0122] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. An adaptive vehicle sensor attack detection method based on multimodal fusion, characterized in that, Includes the following steps: S1. Obtain the raw observation data of connected autonomous driving platoon vehicles and the vehicle control input of the previous moment; then, based on the data-driven state estimation model, model the vehicle state evolution process to obtain state estimates. S2. Construct a time sliding window, input the original observation data, state estimate and control input sequence within the window into a causal dilated convolutional network to extract local temporal dependency features; at the same time, calculate the residual features between the original observation data and the state estimate within the window, and construct consistency verification features based on the consistency of vehicle kinematics and the consistency relationship between visual perception state. The specific steps for constructing the consistency verification feature are as follows: First, time consistency features are constructed from three dimensions, specifically composed of "step ratio features, multi-vehicle synchronization gap deviation features, and sensor hysteresis features". Then, based on the image data collected by the vehicle-mounted camera following the vehicle, combined with the temporal state vector and the geometric features of the bounding box, the state of the preceding vehicle is inferred through a multimodal fusion regression model to obtain the visual inferred value. The visual inferred value is compared with the observation vector, and after calculating the residual, the statistical features of the residual sequence are extracted within the sliding time window, including the window mean and the maximum deviation, as the visual-kinematic cross-consistency feature. Secondly, based on the position coordinates of the preceding and following vehicles obtained by the Global Navigation Satellite System, the GNSS-estimated distance is converted to metric distance using an equidistant rectangle approximation. Then, the GNSS-estimated distance is compared with the radar-measured distance, the absolute difference between the two is calculated, and this difference is used as the GNSS-radar consistency residual feature. Finally, the average absolute value of the speed of the two vehicles within the sliding window is calculated as the speed reference and the average value of the accelerator pedal opening is calculated as the control reference. The speed reference is linearly mapped to the [0,1] interval to obtain the speed gating factor. At the same time, the control reference is normalized to obtain the control gating factor. The maximum value of the two is taken as the motion gating factor. The time consistency features, visual-kinematic cross-consistency features, motion gating factors and GNSS-radar consistency residuals are Z-score normalized and then spliced ​​in a preset order to form a consistency verification feature vector. S3. Calculate the statistics of the original observation data and local temporal dependency features within the sliding window; then concatenate and fuse the local temporal dependency features, residual features, consistency check features and statistics to generate a multimodal feature vector. S4. Training the anomaly detection model A frequency-domain enhanced variational autoencoder, a deep support vector data description, and a detection network based on prototype representation and memory update mechanism were trained separately. After training, anomaly evaluation of multimodal feature vectors was performed by each model, and an initial anomaly score was generated. S5. Based on the trained anomaly detection model, construct a multi-channel integrated detection framework, calculate the anomaly score of the multimodal feature vector of each channel when subjected to different types of network attacks, and finally perform dynamic weighted fusion to obtain a comprehensive anomaly score. S6. Based on the obtained comprehensive anomaly score, construct an adaptive threshold adjustment mechanism based on Beta distribution, determine the dynamic threshold, and perform anomaly judgment.

2. The adaptive vehicle sensor attack detection method based on multimodal fusion according to claim 1, characterized in that, The data-driven state estimation model adopts a prediction-update two-stage architecture, specifically including: First, construct the vehicle state vector. With observation vector The state vector is represented as: in, For the vehicle's global coordinate information, For heading angle, For longitudinal velocity, The yaw rate is expressed using sin and cos trigonometric functions to avoid abrupt angle changes. Observation vector This represents the raw observation data collected by the vehicle's onboard sensors at time t; then, a predictor generates a priori state estimate based on historical multi-sensor sequences; next, an updater corrects the priori state to obtain the posterior state estimate of the vehicle. Finally, the accuracy of vehicle state estimation is improved by minimizing the loss function. This loss function consists of the weighted mean square error between the posterior state and the observation vector, the weighted mean square error between the prior state and the observation vector, and a smoothing constraint term for the change in heading angle. After the model is trained, a vehicle state estimation model is obtained, and the state estimates of vehicles in the platoon are calculated accordingly.

3. The adaptive vehicle sensor attack detection method based on multimodal fusion according to claim 1, characterized in that, The specific time consistency features are as follows: (1) Step ratio characteristics: Define the step ratio Where i represents the sampling index in the sliding window, The difference in timestamps between adjacent sensors. The time difference between adjacent systems is used to extract the mean offset, minimum offset, standard deviation, and stagnation ratio of the step ratio within the sliding window, and these are then combined to form a four-dimensional timestamp step ratio feature. (2) Multi-vehicle synchronization gap deviation characteristics: The synchronization gap between the timestamps of sensors of two adjacent vehicles is defined as... ;in This represents the sensor timestamp at the i-th sampling time while following the vehicle. This represents the sensor timestamp of the preceding vehicle at the i-th sampling time. For a convoy system containing multiple vehicles, the synchronization gap between each vehicle and the preceding vehicle is calculated to form a multi-vehicle synchronization gap sequence. Then, within a sliding time window, the current synchronization gap is compared with the median of the historical synchronization gap sequence within the window, and the degree of deviation is calculated as the multi-vehicle synchronization gap deviation feature. (3) Sensor hysteresis characteristics: defining the hysteresis quantity This reflects the difference between the system time and the sensor timestamp, where... This represents the timestamp of the corresponding sensor. Under normal conditions, the hysteresis is approximately constant due to continuous sensor updates. When sensor data transmission is blocked, system time continues to advance while the sensor timestamp stagnates, causing the hysteresis to increase linearly with time. The hysteresis at consecutive moments is... The hysteresis time series is constructed, and the mean, standard deviation and the difference in hysteresis at the beginning and end of the window are extracted from the hysteresis window series. The results are then spliced ​​together to obtain the sensor hysteresis characteristics. The time consistency feature is ultimately composed of step ratio characteristics, multi-vehicle synchronization gap deviation characteristics, and sensor hysteresis characteristics.

4. The adaptive vehicle sensor attack detection method based on multimodal fusion according to claim 1, characterized in that, The initial anomaly score is calculated based on a frequency-domain enhanced variational autoencoder, as follows: First, a causal dilated convolutional network is used to encode the historical multimodal feature sequence of length L at time t to extract temporal context features. Simultaneously, a fast Fourier transform is performed on the historical multimodal feature sequence to extract frequency domain features. Then, the temporal context features and frequency domain features are fused, and the fused context representation is used as a conditional input to construct the prior and posterior distributions of the latent variables. The prior distribution is predicted from the historical multimodal feature sequence, while the posterior distribution is inferred from the historical multimodal feature sequence and the current multimodal feature vector. Then, a hierarchical decoding strategy is adopted to divide the latent variables into physical state sub-vectors and temporal consistency sub-vectors, and reconstruct them through the physical state decoder and temporal consistency decoder respectively to obtain the reconstructed distribution parameters of the observed features and the reconstructed values ​​of the global spectral features. Subsequently, an initial anomaly score is constructed. ;in, To reconstruct the negative log-likelihood in the hierarchical time domain; This represents the divergence between the posterior and prior distributions. for The weight parameters, It represents the mean square error between the reconstructed values ​​of the frequency domain features of the historical sequence and the global spectral features.

5. The adaptive vehicle sensor attack detection method based on multimodal fusion according to claim 1, characterized in that, The initial anomaly score is calculated based on deep support vector data description, as follows: First, the multimodal feature vectors are used as input, and a latent representation is obtained through a deep encoding network; Then, define the sphere center vector, construct the training loss function, and constrain the normal samples to form a compact distribution in the latent space; Finally, the initial outlier score is calculated by taking the distance from the sample representation to the center of the hypersphere in the latent space.

6. The adaptive vehicle sensor attack detection method based on multimodal fusion according to claim 1, characterized in that, The detection network based on prototype representation and memory update mechanism maps multimodal feature vectors to the latent space through encoder, constructs a normal behavior prototype library, and calculates the anomaly score by calculating the distance between the latent representation of the sample and the prototype during the detection stage.

7. The adaptive vehicle sensor attack detection method based on multimodal fusion according to claim 1, characterized in that, A multi-channel integrated detection framework is constructed based on three models: Frequency Domain Variational Autoencoder (FVAE), Deep Support Vector Description (Deep SVDD), and MemoryBank, a detection network based on prototype representation and memory update mechanism. First, the three models mentioned above are used to analyze the input time series data and output the anomaly score at the current time. Then, the anomaly scores output by the three models are normalized using Min-Max. A heterogeneous dual-channel judgment mechanism is designed to generate corresponding anomaly scores for time series structure anomalies and amplitude tampering anomalies, constructing a time series structure anomaly bypass channel and an amplitude tampering anomaly channel. The time series structure anomaly bypass channel is for delay-type and time series structure destruction-type anomalies, and its anomaly score is directly given by the normalized FVAE anomaly score. The amplitude tampering anomaly channel is for amplitude tampering-type anomalies, and its anomaly score is obtained by weighted fusion of the normalized anomaly scores from the Deep SVDD and MemoryBank models. Finally, the anomaly scores of the two channels are weighted and fused to obtain the comprehensive anomaly score.

8. The adaptive vehicle sensor attack detection method based on multimodal fusion according to claim 1, characterized in that, The adaptive threshold adjustment mechanism based on the Beta distribution is as follows: The Beta distribution is used to statistically model the obtained composite anomaly scores to determine the anomaly detection threshold; the historical composite anomaly score sequence recorded within a time sliding window is utilized. The sample fits the Beta distribution Set confidence quantiles The threshold is defined as: in The threshold is the quantile function of the Beta distribution; an online adaptive adjustment of the threshold is adopted at the current time. Collect recent A window is formed by a score that provides normal feedback. ,right Refitting Beta distribution parameters And calculate the dynamic threshold When the integrated detection framework receives feedback that the current sample is normal, it adds the corresponding comprehensive anomaly score to the historical comprehensive anomaly score sequence and updates the threshold. If there is no normal feedback, the update will not be performed.

9. An electronic device, characterized in that, Including processor and memory; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the adaptive vehicle sensor attack detection method according to any one of claims 1-8.

Images