Sensor spoofing attack identification method based on cross-modal consistency verification and application

Abstract

The application discloses a sensor spoofing attack identification method based on cross-modal consistency verification and application, and belongs to the field of intelligent driving safety and multi-sensor fusion. First, the method completes multi-source sensor data acquisition and space-time registration, and builds a unified cross-modal perception representation model. Second, the target state association is established relying on the observation data of each sensor, and the motion, space and environment cross-modal consistency features are extracted. Third, a consistency quantitative evaluation model is constructed to determine the matching deviation of the sensing information to capture the anomaly. Finally, the attack identification conclusion is output in combination with risk assessment, and the perception model is dynamically iterated. The scheme can accurately identify GPS, visual, radar and other multi-type sensor spoofing attacks relying on cross-modal verification, multi-source data association and dynamic risk assessment strategy, significantly improves the attack detection performance of the vehicle-mounted perception system, and strengthens the reliability, robustness and driving safety level of the intelligent driving perception layer.

Description

Technical Field

[0001] This invention belongs to the field of intelligent driving safety and multi-sensor fusion technology, and in particular relates to a sensor spoofing attack identification method and its application based on cross-modal consistency verification. Background Technology

[0002] As autonomous driving technology evolves towards Level 3 and above, the reliance of intelligent vehicle perception systems on multi-source heterogeneous sensors continues to deepen. Current mainstream intelligent vehicles typically integrate multiple types of sensors, including cameras, LiDAR, millimeter-wave radar, Global Navigation Satellite System (GNSS) receivers, inertial measurement units (IMUs), and onboard V2X communication modules, achieving comprehensive perception and accurate modeling of the surrounding environment through multimodal data fusion. While multi-sensor fusion architecture improves perception accuracy and redundancy, it also significantly expands the system's attack surface. Attackers can launch deception attacks against single or multiple sensors, injecting observational data, tampering with sensor outputs, or simulating environmental signals to cause the perception system to generate incorrect environmental understanding, thereby misleading the decision-making and planning module to engage in dangerous driving behaviors.

[0003] Sensor spoofing attacks are more covert and destructive than traditional denial-of-service or replay attacks. GPS spoofing attacks transmit satellite signals to GNSS receivers, giving vehicles offset position coordinates; visual spoofing attacks interfere with camera target detection results through adversarial samples or physical patches; radar spoofing attacks tamper with the distance and velocity measurements of millimeter-wave radar by transmitting jamming signals. Even more serious is the fact that advanced attackers can simultaneously launch coordinated spoofing attacks against multiple sensors, rendering single-modal anomaly detection mechanisms ineffective.

[0004] Existing sensor attack detection technologies have the following main shortcomings: (1) Limitations of single-modal detection: Existing methods are mostly designed for single sensor types, relying on the statistical characteristics or physical constraints of the sensor itself to make anomaly judgments, and lack cross-modal cross-verification capabilities. When the deceptive data carefully constructed by the attacker satisfies the single-modal statistical distribution constraints, such methods will produce serious false negatives; (2) Lack of consistency verification mechanism: Multi-sensor fusion systems naturally have cross-modal redundant information, but existing solutions fail to effectively utilize the physical consistency constraints that should be satisfied between different modal observation results. Information such as position, speed, and environmental semantics has a definite mapping relationship between different sensors. Deception attacks will inevitably destroy this cross-modal consistency, and existing technologies lack quantitative measurement methods for such consistency destruction; (3) Static risk assessment: Existing detection systems mostly use fixed thresholds for binary judgment, failing to comprehensively consider the degree of real-time abnormal deviation and historical abnormal behavior patterns, making it impossible to continuously quantify and classify attack risks, and difficult to support the formulation of graded response strategies; (4) Lack of adaptive evolution capability of the model: Faced with constantly evolving attack methods and dynamically changing driving environments, the detection performance of static models will continue to degrade, lacking online parameter adjustment and threshold adaptive mechanisms based on detection feedback.

[0005] Therefore, this invention proposes a sensor spoofing attack identification method and application based on cross-modal consistency verification. By constructing a unified spatiotemporal representation of multi-source sensor data, extracting cross-modal consistency features and establishing a quantitative evaluation model, integrating real-time abnormal deviations and historical abnormal behaviors for risk level assessment, and combining attack confidence judgment and impact index analysis to achieve accurate attack identification, and continuously improving the system's detection capability and adaptability to new spoofing attacks through a closed-loop adaptive update mechanism of model parameters and judgment thresholds. Summary of the Invention

[0006] The purpose of this invention is to overcome the problems of lack of cross-modal verification capability, static risk assessment and insufficient model adaptation capability in existing intelligent vehicle sensor attack detection technologies, and to propose a sensor spoofing attack identification method and application based on cross-modal consistency verification.

[0007] To achieve the above objectives, this invention provides a sensor spoofing attack identification method based on cross-modal consistency verification, comprising the following steps: Step 1) Acquire raw sensing data generated by multiple sensors, collect raw observation data, and form a raw sensing dataset. The specific method is as follows: Raw observation data is collected and organized according to sensor type to form a raw sensing dataset. Its expression is:

[0008] in, Represents the original sensory dataset, Indicates the first The raw observation data corresponding to the sensor type, Indicates the total number of sensors. Indicates the sensor number; The original observation data Represented as:

[0009] in, Indicates the first The raw observation data corresponding to the sensor type, Indicates the first Type of sensor in the first The observations obtained at each sampling time point Indicates the number of sampled data. Indicates the sampling time number; To quantify the data integrity of various sensors and calculate data efficiency :

[0010] in, Indicates the first The data efficiency of sensor-like devices Indicates the first The number of effective observation data of this type of sensor Indicates the first The total number of observation data of this type of sensor; when Greater than the preset efficiency threshold If the data is in the correct state, retain the corresponding sensor data; otherwise, mark it as an abnormal data source.

[0011] Step 2) Preprocess the raw sensor dataset obtained in Step 1) by performing time synchronization, spatial alignment, and standardization to form a preprocessed dataset. The specific method is as follows: Using the raw sensory dataset formed in step 1) Time synchronization, spatial coordinate alignment, and standardization are performed to form a preprocessed dataset. ; First, let the first... Sensor-like sensors The original observation sequence is as follows:

[0012] Indicates the first Type of sensor in the first The original timestamps of each sampling time point ( ), Indicates the number of sampled data. This indicates the sampling time number; due to inconsistent sampling frequencies and clock skew among different data sources, the sampling times of all sensors are corrected to a unified time grid. ,in To unify time grid points; Based on the timestamps of each sensor, the data correspondence under a unified time reference is calculated. First, the original timestamps are corrected sample by sample, and the time synchronization result is expressed as follows:

[0013] in, Indicates the first Type of sensor in the first The correction timestamp after synchronization of each sampling time point Indicates the first Type of sensor in the first The original timestamps of each sampling time point Indicates the first The time offset of a sensor relative to a unified time base. To unify time grid points; After timestamp correction, the first The observed values ​​of the sensor remain unchanged, only the time stamp is replaced with a calibration timestamp. The synchronized sequence is denoted as:

[0014] For the first Sensor-like devices perform time alignment, i.e., correcting timestamps. For interpolation nodes, for the synchronized sequence Construct a time-dimensional interpolation function and apply it at a unified time grid point. The value is evaluated to obtain the time-aligned data:

[0015] in, For time-dimensional interpolation operators (linear interpolation or cubic spline interpolation). To correct timestamps This is the sequence after the nodes are synchronized. To unify time grid points, Indicates the first Sensors at a unified time grid point Time alignment value at; Subsequently, the time alignment value was adjusted. Perform spatial alignment processing, let the first... There are 1 sensor, and the number of sampled data is 1 The original set of space observation points is:

[0016] Its spatial distribution needs to be aligned with a unified spatial grid. Maintain consistency;

[0017] Based on time alignment, further adjustments were made to the first... Each sensor performs spatial alignment, Mapped to a unified spatial grid point The time-space joint alignment number is obtained as follows:

[0018] in, For spatial interpolation operators, Indicates the first After time and spatial alignment, the sensor data is placed in a unified spatial grid. And unified time grid points Data values ​​at; Subsequently, the synchronized observation data were standardized, and the standardization process focused on the temporal-spatial joint alignment number after temporal and spatial alignment. ; For the first Temporal-spatial joint alignment of individual sensors Calculate its mean and standard deviation in a unified spatiotemporal domain:

[0019]

[0020] in, and The first Each sensor at a unified spatial grid point Place, across all The mean and standard deviation of each unified time step; the calculation scope of the above statistics is limited to the aligned time-space joint alignment number. superior; Based on this, Z-score standardization is performed to obtain the standardized data:

[0021] in, Indicates the first Type of sensor in the first Standardized observations corresponding to each sampling time; For the first Temporal-spatial joint alignment number of sensor-like devices; Indicates the first Mean value of sensor-like observation data; Indicates the first Standard deviation of sensor-like observation data; ; The various types of preprocessed observation data are organized in a unified manner to form a preprocessed dataset. :

[0022]

[0023] in, Indicates the first Type of sensor in the first Standardized observations corresponding to each sampling time. This indicates a preprocessed dataset. This indicates the total number of sensors.

[0024] Step 3) Based on the preprocessed dataset formed in Step 2), extract cross-modal consistency features, construct the correlation between observation results from different sensors, and form a cross-modal consistency feature set. The specific method is as follows: Using the preprocessed dataset formed in step 2) Cross-modal consistency features are extracted by analyzing the positional, motion, and state correlations among observations from different sensors to construct a cross-modal consistency feature set. ; First, calculate the feature correlation between observation data from different modalities:

[0025] in, Indicates the first Type of sensor and the first Feature correlation between sensor types Indicates the first The first type of sensor One preprocessed data, Indicates the first The first type of sensor One preprocessed data, Indicates the first Mean of preprocessed data for sensor-like devices Indicates the first Mean of preprocessed data for sensor-like devices Indicates the number of samples; Construct cross-modal consistent feature vectors based on feature correlation:

[0026] in, Indicates the first Cross-modal consistency features Indicates the first Type of sensor and the first Feature correlation between sensor types Indicates the first The preprocessed data set corresponding to the sensor type Indicates the total number of sensors; All cross-modal consistency features are combined to form a cross-modal consistency feature set. :

[0027] in, Represents a cross-modal consistency feature set. Indicates the first Cross-modal consistency features This indicates the number of cross-modal consistency features.

[0028] Step 4) Establish a cross-modal consistency evaluation model using the cross-modal consistency feature set formed in Step 3), and analyze the cross-modal consistency feature set using the cross-modal consistency evaluation model to generate consistency evaluation results. The specific method is as follows: Using the cross-modal consistency feature set formed in step 3) A cross-modal consistency evaluation model is established. By calculating the degree of matching and deviation between different modal features, the cross-modal consistency features are quantitatively evaluated, and consistency evaluation results are generated. ; First, calculate the average deviation between cross-modal consistency features:

[0029] Where δ represents the average deviation value of the cross-modal characteristics. Indicates the first Cross-modal consistency features Indicates the first Cross-modal consistency features Indicates the number of cross-modal consistency features; The consistency evaluation results are calculated based on the average deviation value of cross-modal characteristics:

[0030] in, The result represents the consistency evaluation result, where δ represents the average deviation of cross-modal characteristics, and e represents the base of the natural logarithm. When the observation deviation between different modal observation results is within the preset normal deviation range [0, 0.2], δ decreases, and the consistency evaluation result improves. Increase; when the observation deviation between different modal observation results exceeds the preset normal deviation range [0, 0.2], δ increases, and the consistency evaluation results... Decrease.

[0031] Step 5) Calculate the degree of deviation between the observation information of each sensor using the consistency evaluation results generated in Step 4), and generate a comprehensive anomaly deviation result. ; The specific method is as follows: using the consistency evaluation results generated in step 4). The degree of deviation between observation information from different modalities is calculated, the consistency status of cross-modal observation results is quantitatively analyzed, and a comprehensive anomaly deviation result is generated. ; First, based on the consistency evaluation results Calculate abnormal deviation results :

[0032] in, Indicates abnormal deviation from the result. This indicates the consistency evaluation results; To enhance the sensitivity of abnormal deviation results to local abnormal fluctuations, the abnormal deviation increment is calculated:

[0033] in, Indicates time The corresponding abnormal deviation increment, Indicates time The corresponding abnormal deviation results Indicates time Corresponding abnormal deviation results; Based on abnormal deviation results Incremental deviation from abnormality Calculate the overall anomaly deviation results:

[0034] in, This indicates the overall abnormal deviation results. Indicates abnormal deviation from the result. This indicates an abnormal deviation from the increment. This represents the weighting coefficient, and satisfies 0 < <1; When the overall abnormal deviation from the result Exceeding the preset abnormal threshold If the current cross-modal observation information shows an abnormal deviation trend, it is determined that the current cross-modal observation information remains in a normal and consistent state; otherwise, it is determined that the current cross-modal observation information remains in a normal and consistent state.

[0035] Step 6) Utilize the comprehensive anomaly deviation results generated in Step 5) to establish an attack risk analysis model, and use the attack risk analysis model to perform risk quantification analysis, generating a normalized attack risk level result. The specific method is as follows: Using the comprehensive anomaly deviation results generated in step 5) An attack risk analysis model is established, which quantifies the risk of sensor spoofing attacks by fusing real-time anomaly deviation information with historical anomaly behavior information, and generates a normalized attack risk level result. ; First, calculate the frequency of historical anomalies:

[0036] in, Indicates the frequency of historical anomalies. Indicates the number of historical anomalies. This represents the total number of historically observed events; Based on the comprehensive abnormal deviation results and frequency of historical anomalies Calculate the attack risk level result :

[0037] in, This indicates the attack risk level result. This indicates the overall abnormal deviation results. Indicates the frequency of historical anomalies. Indicates the weight of abnormal deviations from the result. Represents the weight of historical abnormal frequencies, and satisfies:

[0038] To facilitate subsequent risk level classification, the attack risk level results are... Normalization is performed:

[0039] in, This represents the normalized attack risk level result, where R represents the attack risk level result. This represents the historical maximum risk value. This represents the historical minimum risk value.

[0040] Step 7) Use the attack risk level results generated in Step 6) to identify sensor spoofing attack behavior and generate spoofing attack identification results. The specific method is as follows: Using the normalized attack risk level result generated in step 6) To identify sensor spoofing attacks, an attack determination mechanism is constructed to classify and judge the current observation state, generating spoofing attack identification results. .

[0041] First, based on the normalized attack risk level results Calculate attack confidence:

[0042] in, Indicates the attack confidence level. This indicates the normalized attack risk level result. Represents the base of the natural logarithm; Based on attack confidence Perform attack detection:

[0043] in, This indicates the attack determination result. =1 indicates the presence of a sensor spoofing attack. =0 indicates that there is no sensor spoofing attack. Indicates the attack confidence level. Indicates the attack detection threshold; To assess the extent of the attack's impact, an attack impact index is calculated:

[0044] in, Indicates the impact index of the attack. Indicates the attack confidence level. This represents the normalized attack risk level result. When the attack impact index Exceeding the preset impact threshold When the attack behavior is deemed to have a high-risk impact on the perception system, the attack impact index is determined to be high. Not exceeding the preset impact threshold At that time, it is determined that the current attack behavior has a low-risk impact on the perception system; Based on the attack determination result and the attack impact index Generate deception attack identification results ,include: Calculate the impact level of the attack :

[0045] in For indicator functions, The preset impact threshold; Calculate the deception attack identification result

[0046]

[0047] in, The set of values ​​is ,when When, it indicates that there is no sensor spoofing attack; when At times, sensor spoofing attacks exist and have a low-risk impact on the vehicle's perception system; when At that time, it indicates the existence of sensor spoofing attacks and a high-risk impact on the vehicle's perception system; ultimately, based on the attack determination results... and attack impact index Generate attack impact level Generate deception attack identification results .

[0048] Step 8) Adaptively adjust the perceptual model parameters and attack judgment threshold using the deception attack identification results generated in Step 7) to generate an updated attack identification model. The specific method is as follows: Using the deception attack identification results generated in step 7) Attack identification model Dynamic updates are performed by adjusting model parameters and attack detection thresholds to improve the attack identification model's adaptability to new deceptive attack behaviors, generating an updated attack identification model. ; First, based on the results of the deception attack identification Calculate the model parameter correction amount:

[0049] in, This represents the amount of model parameter correction. This represents the parameter adjustment coefficient. This indicates the results of the deception attack identification. Indicating attack identification model The output prediction results; Update the model parameters based on the model parameter corrections:

[0050] in, This represents the updated model parameters. This represents the model parameters before the update. This indicates the amount of model parameter correction; To enhance the model's adaptability to changes in the attack environment, the attack detection threshold is dynamically updated:

[0051] in, This indicates the updated attack detection threshold. This represents the attack detection threshold before the update, and γ represents the threshold adjustment coefficient. Indicates the current attack confidence level. This represents the historical average attack confidence level; Based on the updated model parameters and the updated attack detection threshold Build an updated attack identification model :

[0052] in, This represents the updated attack detection model. θ' represents the model building function, and θ' represents the updated model parameters. This indicates the updated attack detection threshold.

[0053] Step 9) Utilize the updated attack identification model generated in Step 8) To acquire a new round of real-time sensor data, perform time synchronization, spatial alignment, and standardization preprocessing as described in step 2) to form a real-time preprocessed dataset. ; Real-time preprocessing of datasets Substitute the updated model parameters into the cross-modal consistency verification and risk analysis process described in steps 3) to 6), and use the updated model parameters. and the updated attack detection threshold Perform real-time spoofing attack identification and obtain real-time spoofing attack identification results. The specific method is as follows: Using the updated attack identification model generated in step 8) Acquire a new round of real-time sensor data, and preprocess it according to the method described in step 2) to form a real-time preprocessed dataset. :

[0054] in, This represents a real-time preprocessed dataset. Indicates the first Type of sensor in the first Standardized observations corresponding to each real-time sampling time. Indicates the total number of sensors. Indicates the amount of real-time sampled data; Real-time preprocessing of datasets Extract the real-time cross-modal consistency feature set according to the method described in step 3). :

[0055] in, Represents the real-time cross-modal consistency feature set. Indicates the first A real-time cross-modal consistency feature. Indicates the number of real-time cross-modal consistency features; Using the updated model parameters generated in step 8) For real-time cross-modal consistency feature sets Perform consistency evaluation and anomaly deviation analysis to generate real-time comprehensive anomaly deviation results. :

[0056] in, This indicates the real-time comprehensive abnormal deviation results. Indicates real-time abnormal deviation from the result. Indicates real-time abnormal deviation from increment. Denotes the weight coefficients, and satisfies ; Real-time comprehensive abnormal deviation results Introducing the updated attack identification model In the middle, using the updated model parameters Calculate real-time attack confidence :

[0057] in, Indicates the real-time attack confidence level. This represents the updated model parameters generated in step 8). This indicates the real-time comprehensive abnormal deviation results. Represents the base of the natural logarithm; Based on real-time attack confidence Compared with the updated attack determination threshold generated in step 8), Attack detection is performed to obtain real-time deception attack identification results. :

[0058] in, This indicates the real-time identification results of spoofing attacks. This indicates that a sensor spoofing attack has been detected. This indicates that no sensor spoofing attack was detected. This represents the updated attack threshold generated in step 8). Real-time spoofing attack identification results As the final identification output, at the same time Feedback is sent to step 7) to accumulate and update the results of the next round of spoofing attack identification and to continuously and adaptively adjust the model parameters, thereby realizing online real-time identification and closed-loop continuous optimization of sensor spoofing attacks.

[0059] Compared with the prior art, the present invention has the following beneficial effects: (1) Strong cross-modal cross-validation capability: Breaking through the limitations of traditional single-modal independent detection, by constructing positional correlation, motion correlation and state correlation between observation results of different sensors, cross-validation of multi-source data is achieved by utilizing cross-modal consistency features, so that carefully constructed single-modal deception data exposes inconsistencies in cross-modal verification, significantly reducing the false negative rate; (2) Accurate consistency quantification evaluation: Introducing a consistency evaluation model based on exponential decay function, the discrete intermodal deviation is transformed into a continuous consistency metric value, realizing the fine quantification of the degree of cross-modal matching, avoiding information loss caused by fixed threshold judgment; (3) High anomaly detection sensitivity: By fusing absolute anomaly bias (3) Deviate from incremental abnormality and construct a comprehensive abnormality deviation index, which can capture global consistency destruction and maintain high sensitivity to local mutation attacks, effectively responding to gradual deception attacks and sudden deception attacks; (4) Hierarchical risk assessment: integrate real-time abnormality deviation information and historical abnormal behavior patterns, generate continuous risk level quantification values ​​through weighted fusion and normalization processing, support the formulation of hierarchical response strategies, and avoid the coarse-grained defects of binary judgment; (5) Model adaptive evolution: construct a closed-loop update mechanism based on detection feedback, and adjust model parameters and attack judgment thresholds at the same time, so that the system can continuously adapt to new attack methods and dynamic environmental changes, and ensure the stability of long-term detection performance.

[0060] In summary, the sensor spoofing attack identification method based on cross-modal consistency verification provided by this invention, by constructing a complete closed-loop system from cross-modal feature extraction, consistency quantification evaluation, abnormal deviation detection, risk level assessment to attack identification and model iteration, achieves high-precision identification and continuous adaptive defense of intelligent vehicle perception systems against multiple types of sensor spoofing attacks, and provides a solid technical guarantee for the safe and reliable operation of the intelligent driving perception layer. Attached Figure Description

[0061] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0063] Example 1: A sensor spoofing attack identification method based on cross-modal consistency verification, applied in intelligent driving. This method first acquires raw perception data from multiple sensors in an intelligent vehicle, completing spatiotemporal registration and standardization preprocessing. Then, it extracts cross-modal consistency features to construct the correlation between observation results from different sensors. Further, it establishes a cross-modal consistency evaluation model to quantify inter-modal matching deviations. Subsequently, it calculates the degree of abnormal deviation and integrates historical anomaly information to establish an attack risk analysis model. Sensor spoofing attack behaviors are identified based on the risk level results. The model parameters and judgment thresholds are adaptively updated using identification feedback. Through cross-modal verification, multi-source data association, and dynamic risk assessment strategies, accurate identification of spoofing attacks from various sensors such as GPS, vision, and radar is achieved, effectively improving the attack detection performance of the vehicle perception system and the level of driving safety. The following detailed implementation steps illustrate this method.

[0064] Step 1) Acquire raw perception data generated by the multi-source sensors of the intelligent vehicle. Collect raw observation data from cameras, LiDAR, millimeter-wave radar, global navigation satellite system sensors, inertial measurement units, and on-board perception devices to form a raw perception dataset. ;

[0065] This step aims to construct a unified framework for the acquisition and organization of multi-source heterogeneous sensor data for intelligent vehicles, providing a standardized data foundation for subsequent cross-modal consistency verification. During the operation of an intelligent vehicle, the onboard perception system collects information about the surrounding environment in real time through various sensors with different physical principles. The data output by each sensor has fundamental differences in modality, dimension, sampling frequency, and physical semantics. Therefore, a unified data organization model needs to be established to achieve structured management of multi-source observation data.

[0066] (I) Multi-source sensor data acquisition architecture: Assume the total number of sensors on the intelligent vehicle is... The system encompasses cameras, lidar, millimeter-wave radar, Global Navigation Satellite System (GNSS) sensors, inertial measurement units (IMUs), and an onboard V2X communication module. Each sensor independently generates raw observation data according to its inherent sampling frequency, which is transmitted to the central sensing and processing unit via the onboard high-speed bus. The data modes and dimensions output by each sensor are fundamentally different: cameras output two-dimensional image tensors, lidar outputs three-dimensional point cloud coordinate sets, millimeter-wave radar outputs target distance-velocity-angle feature vectors, GNSS outputs position coordinate vectors, IMU outputs six-dimensional motion state vectors, and the V2X module outputs a sequence of state information of surrounding traffic participants.

[0067] (II) Formal Definition of Raw Sensing Dataset: Raw observation data generated by cameras, lidar, millimeter-wave radar, global navigation satellite system sensors, inertial measurement units, and vehicle communication modules are collected and organized uniformly according to sensor categories to form a raw sensing dataset. Its expression is:

[0068] : Raw perception dataset, representing the collection of raw observation data collected by all sensors of the intelligent vehicle; : No. Raw observation data corresponding to the sensor type; The total number of sensors depends on the sensor configuration scheme of the intelligent vehicle; Sensor number, value range is: .

[0069] Furthermore, the raw observation data Represented as:

[0070] : No. Raw observation data corresponding to the sensor type; : No. Type of sensor in the first The dimensions of the observations obtained at each sampling time depend on the sensor type and measurement mode. The amount of sampled data is determined by both the sensor sampling frequency and the observation duration. : Sampling time number, value range is .

[0071] (III) Quantitative Assessment of Data Integrity: During data acquisition, sensor malfunctions, communication interruptions, or electromagnetic interference may lead to the loss or invalidation of some observation data. To ensure the reliability of subsequent cross-modal consistency analysis, a quantitative assessment of the data integrity of various sensors is necessary. This includes calculating the data validity rate. :

[0072] : No. The data efficiency of this type of sensor has a range of values. This reflects the proportion of valid data from the sensor during the observation period; : No. The number of valid observation data from the sensor, i.e. the number of data points that pass integrity and format checks; : No. The total number of observation data points for a sensor type, i.e., the theoretical number of data points that the sensor should generate within the observation period. When Greater than the preset efficiency threshold If the data is valid, retain the corresponding sensor data; otherwise, mark it as an abnormal data source and exclude it from subsequent cross-modal analysis. Efficiency threshold. The setting of the parameter needs to take into account both data availability and analysis reliability, and its value should ensure that the sensor data participating in cross-modal consistency verification has sufficient statistical representativeness.

[0073] (iv) Multi-source observation data aggregation: Multi-source observation data that meet the data integrity requirements are aggregated to form the original sensing dataset. During the aggregation process, it is necessary to record the category identifier, sampling frequency, data dimension, and timestamp information of each sensor to provide metadata support for subsequent time synchronization and spatial alignment. For cases where multiple sensors have inconsistent sampling frequencies, the original timestamps of each sensor are retained in the original dataset for unified time registration in step 2). Finally, the original sensing dataset is... Step 2) is used for preprocessing.

[0074] Step 2) Preprocess the raw sensor dataset obtained in Step 1) by performing time synchronization, spatial alignment, and standardization to form a preprocessed dataset.

[0075] This step aims to eliminate inconsistencies in time reference, spatial reference frame, units of measurement, and data quality among multi-source sensor data, and to construct a standardized data representation suitable for cross-modal consistency feature extraction. Since different sensors differ in clock sources, installation locations, measurement units, and noise characteristics, directly performing cross-modal correlation analysis will introduce systematic biases. Therefore, a rigorous preprocessing procedure is necessary to eliminate these non-aggressive interference factors.

[0076] (a) Time synchronization processing: First, let the first... Sensor-like sensors The original observation sequence is as follows:

[0077] Indicates the first Type of sensor in the first The original timestamps of each sampling time point ( ), Indicates the number of sampled data. This indicates the sampling time number. Because the sampling frequencies of different data sources are inconsistent and there is clock skew, the sampling times of all sensors need to be corrected to a unified time grid. ,in To unify time grid points.

[0078] Based on the timestamps of each sensor, the data correspondence under a unified time reference is calculated. First, the original timestamps are corrected sample by sample, and the time synchronization result is expressed as follows:

[0079] in, Indicates the first Type of sensor in the first The correction timestamp after synchronization of each sampling time point Indicates the first Type of sensor in the first The original timestamps of each sampling time point Indicates the first The time offset of a sensor relative to a unified time base. To unify the time grid points. After timestamp correction, the [number]th [time grid point]. The observed values ​​of the sensor remain unchanged, only the time stamp is replaced with a calibration timestamp. The synchronized sequence is denoted as:

[0080] For the Sensor-like devices perform time alignment, i.e., correcting timestamps. For interpolation nodes, for the synchronized sequence Construct a time-dimensional interpolation function and apply it to a unified time grid point. The value is evaluated to obtain the time-aligned data:

[0081] in, For time-dimensional interpolation operators (linear interpolation or cubic spline interpolation). To correct timestamps This is the sequence after the nodes are synchronized. To unify time grid points, Indicates the first Sensors at a unified time grid point The time alignment value at that location.

[0082] (ii) Spatial alignment processing: Let the first... The original set of spatial observation points for each sensor is:

[0083] Its spatial distribution and unified spatial grid Inconsistent.

[0084]

[0085] Based on time alignment, further adjustments were made to the first... Each sensor performs spatial alignment, Mapped to a unified spatial grid point The time-space joint alignment number is obtained as follows:

[0086] in, For spatial interpolation operators (bilinear interpolation or inverse distance weighted interpolation); Indicates the first After time and spatial alignment, the sensor data is placed in a unified spatial grid. And unified time grid points The data value at that location.

[0087] (ii) Standardization Processing: To eliminate dimensional and numerical range differences between different sensors, the synchronized observation data undergoes standardization processing. The standardization process targets the time-space joint aligned data after time and spatial alignment. Instead of any original observations. For the first... Temporal-spatial joint alignment of individual sensors Calculate its mean and standard deviation in a unified spatiotemporal domain:

[0088]

[0089] in, and The first Each sensor at a unified spatial grid point Place, across all The mean and standard deviation of each uniform time step (number of sampled data points). The calculation of the above statistics is strictly limited to the aligned time-space joint alignment number. Above, it is ensured that the statistical characteristics reflect the data distribution under a unified spatiotemporal reference. Based on this, Z-score standardization is performed to obtain the standardized data:

[0090] in, Indicates the first Type of sensor in the first Standardized observations corresponding to each sampling time; For the first Temporal-spatial joint alignment number of sensor-like devices; Indicates the first Mean value of sensor-like observation data; Indicates the first Standard deviation of sensor-like observation data; To prevent extremely small positive numbers from being divided by zero ( ).

[0091] Standardization maps the data from each sensor to a unified space with zero mean and unit variance, so that subsequent cross-modal correlation calculations are not affected by dimensions and numerical range.

[0092] (III) Construction of Preprocessed Dataset: The various types of preprocessed observation data are organized in a unified manner to form a preprocessed dataset. :

[0093] : Preprocessed dataset, containing a collection of data from all sensors after time synchronization, spatial alignment, outlier filtering, missing value compensation, and standardization; : No. The preprocessed data set corresponding to the sensor type; Total number of sensors. Finally, the preprocessed dataset... Input step 3), used to construct a cross-modal consistent feature set. .

[0094] Step 3) Based on the preprocessed dataset formed in Step 2), extract cross-modal consistency features, construct the correlation between observation results from different sensors, and form a cross-modal consistency feature set.

[0095] This step aims to extract consistency features from preprocessed multi-source sensor data that reflect the inherent correlations between observations from different modalities. Under normal operating conditions, there are definite physical constraints between the observations of the same physical scene from different sensors: position information exhibits geometric consistency between GNSS, LiDAR, and vision; velocity information exhibits kinematic consistency between IMU and radar; and environmental semantics exhibits spatial consistency between camera and LiDAR. Spoofing attacks will disrupt these cross-modal consistency constraints; therefore, extracting and quantifying these consistency features constitutes the core foundation for attack identification.

[0096] (a) Calculation of cross-modal feature correlation: using the preprocessed dataset formed in step 2). Cross-modal consistency features are extracted by analyzing the positional, motion, and state correlations among observations from different sensors to construct a cross-modal consistency feature set. First, calculate the feature correlation degree between observation data from different modalities:

[0097] Indicates the first Type of sensor and the first Feature correlation between sensor types; Indicates the first The first type of sensor One preprocessed data; Indicates the first The first type of sensor One preprocessed data; Indicates the first Mean value of preprocessed data for sensor-like devices; Indicates the first The mean of the preprocessed data for the sensor-like system; m represents the number of samples.

[0098] Feature correlation Physical meaning: When When the value is close to 1, it indicates a strong linear correlation between the observation data of the two types of sensors, and their observation results are tightly coupled at the physical level; when... A value close to 0 indicates that the observation data from the two types of sensors are independent and there is no significant linear correlation. Under normal operating conditions, sensor pairs with physical coupling (such as GNSS and IMU in the position dimension, and camera and lidar in the target detection dimension) should exhibit a high degree of feature correlation.

[0099] (II) Construction of Cross-Modal Consistent Feature Vectors: Further, cross-modal consistent feature vectors are constructed based on feature correlation:

[0100] Indicates the first A cross-modal consistency feature; Indicates the first Type of sensor and the first Feature correlation between sensor types; Indicates the first The preprocessed data set corresponding to the sensor type; This indicates the total number of sensors.

[0101] Cross-modal consistent feature vector The construction principle is based on feature correlation. As weighting coefficients, the preprocessed data from all sensors are weighted and summed. Under normal operating conditions, since the observations from each sensor satisfy physical consistency constraints, the weighted feature vector should exhibit stable statistical properties. However, when a sensor is subjected to a spoofing attack, its observation data deviates from normal physical constraints, leading to changes in the correlation of the corresponding features. Changes occur, leading to abnormal fluctuations in the cross-modal consistency feature vector.

[0102] (III) Formation of cross-modal consistency feature set: Combine all cross-modal consistency features to form a cross-modal consistency feature set. :

[0103] Cross-modal consistency feature set, which contains the set of all cross-modal consistency features; : No. Each cross-modal consistency feature corresponds to the cross-modal correlation state of a specific sensor pair or a specific physical quantity dimension; The number of cross-modal consistency features is determined by the number of sensor pair combinations and the feature dimension.

[0104] Cross-modal consistency feature set This comprehensive feature set encompasses multiple dimensions, including location consistency features (GNSS-LiDAR-visual location correlation), motion consistency features (IMU-radar velocity correlation), and environmental consistency features (visual-LiDAR semantic correlation), forming a complete representation of the cross-modal consistency of multi-source sensing data. Finally, this cross-modal consistency feature set... Input step 4), used to construct the cross-modal consistency evaluation model and generate consistency evaluation results. .

[0105] Step 4) Establish a cross-modal consistency evaluation model using the cross-modal consistency feature set formed in Step 3), and analyze the cross-modal consistency feature set using the cross-modal consistency evaluation model to generate consistency evaluation results. ;

[0106] This step aims to establish a quantitative evaluation model for cross-modal consistency characteristics, compressing multi-dimensional cross-modal correlation states into a single consistency metric. The evaluation model should have the following characteristics: (1) assigning high evaluation values ​​to normal consistency states and low evaluation values ​​to consistency violations; (2) the evaluation results have a non-linear response to the degree of deviation, with small deviations causing gradual evaluation changes and large deviations causing drastic evaluation changes; (3) the evaluation results have clear physical interpretability, facilitating subsequent anomaly determination and risk assessment.

[0107] (a) Calculation of the average deviation value of cross-modal features: using the cross-modal consistency feature set formed in step 3). A cross-modal consistency evaluation model is established to quantitatively evaluate cross-modal consistency features by calculating the degree of matching and deviation between different modal features. First, the average deviation value between cross-modal consistency features is calculated:

[0108] The average deviation value of cross-modal features is a non-negative real number that reflects the average degree of dispersion among all cross-modal consistency features. : No. A cross-modal consistency feature, for the feature set The first in One element; : No. A cross-modal consistency feature, for the feature set The first in One element; : Number of cross-modal consistency features; : Normalization factor, which is the number of all unique feature pairs, ensuring that the average deviation value is not affected by the number of features; : No. The and the first The absolute deviation between cross-modal consistency features.

[0109] Cross-modal characteristic mean deviation value Physical meaning: Under normal operating conditions, the consistency characteristics of each cross-modal mode should fluctuate slightly around their expected values, and the average deviation value of the cross-modal characteristics is... The deviation is relatively small; when a sensor is subjected to a spoofing attack, the cross-modal consistency characteristics involving that sensor will deviate from the normal range, resulting in an average deviation value of the cross-modal characteristics. Significantly increased.

[0110] (ii) Calculation of Consistency Evaluation Results: Further, the consistency evaluation results are calculated based on the average deviation value of cross-modal characteristics:

[0111] Consistency evaluation results, with a value range of: This reflects the overall level of cross-modal consistency of current multi-source sensor data; : Average deviation value of cross-modal characteristics; The base of the natural logarithm.

[0112] Consistency evaluation results Mathematical characteristics analysis: (1) When hour, , indicating that all cross-modal consistency features are completely consistent, and multi-source sensing data are in an ideal consistency state; (2) when When it increases, It decreases monotonically according to an exponential law. (3) The exponential decay characteristic makes It has high sensitivity to small deviations and saturation suppression effect on large deviations, avoiding excessive influence of extreme outliers on the evaluation results; (4) The range of values It has a clear probabilistic interpretation and can be regarded as a confidence level of consistency among multi-source sensing data.

[0113] When the observation deviation between different modal observation results is within the preset normal deviation range [0, 0.2], δ decreases, and the consistency evaluation results improve. Increase; when the observation deviation between different modal observation results exceeds the preset normal deviation range [0, 0.2], δ increases, and the consistency evaluation results... Reduce. Finally, the consistency evaluation results will be... Input step 5), used to calculate the comprehensive anomaly deviation result. .

[0114] Step 5) Calculate the degree of deviation between the observation information of each sensor using the consistency evaluation results generated in Step 4), and generate a comprehensive anomaly deviation result.

[0115] This step aims to transform the consistency evaluation results into anomaly deviation metrics and introduce time-series incremental information to enhance the detection sensitivity for local anomalies. The simple consistency evaluation results... It reflects the global consistency state at the current moment, while deception attacks may manifest as gradual deviations or sudden jumps. It is necessary to combine absolute deviations and incremental deviations to capture both types of attack modes simultaneously.

[0116] (a) Calculation of abnormal deviation results: using the consistency evaluation results generated in step 4). The degree of deviation between observation information from different modalities is calculated, and the consistency status of cross-modal observation results is quantitatively analyzed. First, based on the consistency evaluation results... Calculate abnormal deviation results :

[0117] : Abnormal deviation from the result, the value range is This reflects the degree to which current cross-modal observation information deviates from the normal consistency state; Consistency evaluation results (from step 4). Abnormal deviation results. Consistency evaluation results They are complementary: The larger the value, the more severe the violation of cross-modal consistency. The smaller the value, the better the consistency of multi-source sensor data.

[0118] (II) Calculation of Abnormal Deviation Increment: Further, to enhance the sensitivity of abnormal deviation results to local abnormal fluctuations, the abnormal deviation increment is calculated:

[0119] :time The corresponding abnormal deviation increment reflects the rate of change of the degree of abnormal deviation between adjacent time points; :time Corresponding abnormal deviation results; :time The corresponding abnormal deviation results.

[0120] Abnormal deviation increment Physical meaning: (1) When And when the value is large, it indicates that the degree of abnormal deviation is rapidly increasing, which may indicate the initial stage of a sudden deception attack; (2) when and Maintaining a high level indicates the presence of a persistent steady-state deception attack; (3) when When this occurs, it indicates that the degree of abnormal deviation is easing, the attack may have stopped, or the system is recovering.

[0121] (III) Calculation of Comprehensive Abnormal Deviation Results: Based on the abnormal deviation results Incremental deviation from abnormality Calculate the overall anomaly deviation results:

[0122] By combining the results of abnormal deviations and fusing the absolute deviation level and the rate of change of deviation, the simultaneous detection of steady-state attacks and transient attacks can be achieved. Abnormal deviation results reflect the absolute deviation level at the current moment; Abnormal deviation increment reflects the rate of change of the degree of deviation; : Weighting coefficients, and satisfying The relative importance of controlling absolute deviation and incremental deviation.

[0123] Weighting coefficient Setting principles: When When the deviation is large, the overall abnormal deviation result focuses more on the absolute deviation level, resulting in higher sensitivity for detecting persistent steady-state attacks; when When the value is small, the overall anomaly deviation result focuses more on the rate of deviation change, resulting in higher sensitivity for detecting sudden transient attacks. When the overall anomaly deviation result... Exceeding the preset abnormal threshold If the current cross-modal observation information shows an abnormal deviation trend, it is determined that the current cross-modal observation information remains in a normal and consistent state; otherwise, it is determined that the current cross-modal observation information remains in a normal and consistent state. Abnormal threshold. The settings need to be determined based on the statistical distribution of comprehensive abnormal deviation results under normal operating conditions to ensure that no misjudgment occurs within the normal fluctuation range.

[0124] Finally, the results of the comprehensive abnormal deviations will be analyzed. Input step 6), used to build the attack risk analysis model and generate the normalized attack risk level results. .

[0125] Step 6) Utilize the comprehensive anomaly deviation results generated in Step 5) to establish an attack risk analysis model, and use the attack risk analysis model to perform risk quantification analysis, generating a normalized attack risk level result.

[0126] This step aims to integrate real-time anomaly deviation information with historical anomaly behavior patterns to establish a quantitative analysis model for attack risk. Simple real-time anomaly deviations may be caused by non-attack factors such as sudden environmental changes or momentary sensor malfunctions, while historical anomaly behavior patterns can reflect the statistical regularity of sensor attacks. By integrating these two types of information, it is possible to more accurately distinguish between offensive and non-offensive anomalies and generate continuous quantitative values ​​for risk levels.

[0127] (a) Calculation of historical anomaly frequency: using the comprehensive anomaly deviation results generated in step 5). An attack risk analysis model is established to quantify the risk of sensor spoofing attacks by fusing real-time anomaly deviation information with historical anomaly behavior information. First, the frequency of historical anomalies is calculated:

[0128] : Frequency of historical anomalies, with a range of values. This reflects the probability of anomalies occurring within the historical observation period; : Number of historical abnormal events, i.e., the total number of events in the historical observation period whose overall abnormal deviation results exceed the abnormal threshold; The total number of historical observation events, i.e., the total number of sampling moments within the historical observation period.

[0129] Frequency of historical anomalies The introduction of this feature enables risk assessment to have a time-dimensional memory capability: if a sensor frequently exhibits abnormal deviations within a historical period, its current anomaly is more likely to originate from a sustained attack rather than occasional interference, and should be assigned a higher risk weight.

[0130] (ii) Calculation of attack risk level results: Further, based on the comprehensive abnormal deviation results and frequency of historical anomalies Calculate the attack risk level result :

[0131] The attack risk level result is a weighted fusion value of real-time abnormal deviation and historical abnormal frequency. The comprehensive abnormal deviation results (from step 5) reflect the real-time anomaly level at the current moment. The frequency of historical anomalies reflects the statistical patterns of abnormal behavior within historical cycles. The weight of abnormal deviation results reflects the importance of real-time abnormal information in risk assessment; Historical anomaly frequency weighting reflects the importance of historical behavioral patterns in risk assessment; And it satisfies the normalization constraint:

[0132] Weighting coefficient and The settings need to balance real-time detection sensitivity and historical statistical stability: When the value is large, the system responds to the current anomaly more quickly, but this may increase the false alarm rate. When the scale is large, the system relies more on historical statistical patterns and has a stronger ability to suppress occasional anomalies, but its response to new types of attacks may be delayed.

[0133] (III) Risk Level Normalization: To facilitate subsequent risk level classification, the attack risk level results are normalized. Normalization is performed:

[0134] The normalized attack risk level result has a value range of [value missing]. ; Attack risk level results; The historical maximum risk value is the highest value of the attack risk level results within the historical observation period. The historical minimum risk value is the minimum attack risk level result within the historical observation period.

[0135] Normalization maps the attack risk level results to a unified standard. The interval is used to eliminate differences in the absolute level of risk values ​​across different operational phases, ensuring comparability and consistency of subsequent attack judgment thresholds across cycles. Finally, the normalized attack risk level results are... Input step 7), used to generate spoofing attack identification results. .

[0136] Step 7) Use the normalized attack risk level result generated in Step 6) to identify sensor spoofing attack behavior and generate spoofing attack identification results.

[0137] This step aims to accurately identify sensor spoofing attacks by constructing an attack confidence calculation and judgment mechanism based on quantified attack risk levels. Simultaneously, an attack impact index is introduced to assess the degree of harm the attack poses to the sensing system, providing a basis for subsequent tiered response.

[0138] (a) Attack confidence calculation: using the normalized attack risk level result generated in step 6). To identify sensor spoofing attacks, an attack determination mechanism is constructed to classify and judge the current observation state. First, based on the normalized attack risk level results... Calculate attack confidence:

[0139] Attack confidence level, with a value range of [value missing]. This represents the probability estimate of the presence of sensor spoofing attacks in the current observation state; : The normalized attack risk level result, from step 6); The base of the natural logarithm.

[0140] Attack confidence The Sigmoid function is used to map continuous risk level values ​​to probabilistic confidence scores: (1) when hour, , indicating that the attack and the normal state have equal probability; (2) when And when it increases, Monotonically increasing and approaching 1 indicates an increase in attack confidence; (3) when hour, A monotonically decreasing trend approaching 0 indicates an increased confidence level in the normal state. The smoothing property of the Sigmoid function avoids the boundary discontinuity problem caused by hard thresholding.

[0141] (ii) Attack determination: Further, based on the attack confidence level Perform attack detection:

[0142] Attack determination result; This indicates the presence of a sensor spoofing attack. This indicates that there is no sensor spoofing attack behavior; Attack confidence; Attack detection threshold, with a value range of: The attack detection threshold is set according to system security requirements and false alarm rate tolerance. The settings need to strike a balance between detection sensitivity and false alarm rate: At lower levels, the system is more sensitive to potential attacks, but the false alarm rate increases; At higher levels, the false alarm rate decreases, but weak attacks may be missed.

[0143] (III) Calculation of Attack Impact Index: To assess the degree of attack impact, the attack impact index is calculated:

[0144] Attack Impact Index: This index comprehensively reflects the combined impact of attack confidence and risk level. Attack confidence; : Normalized attack risk level result.

[0145] Attack Impact Index Physical meaning: This index simultaneously considers the probability estimate of the existence of an attack (attack confidence level). ) and the severity of the attack (normalized attack risk level results) The product of these two factors results in high-confidence and high-risk attacks receiving the largest impact index value, while low-confidence or low-risk attacks receive a smaller impact index value.

[0146] When the attack impact index Exceeding the preset impact threshold When the attack is deemed to have a high-risk impact on the vehicle's perception system, the attack impact index is determined to be high. Not exceeding the preset impact threshold If the attack is deemed to have a low-risk impact on the vehicle's perception system, a high-risk impact assessment triggers a strong defense response; a low-risk impact assessment triggers a weak defense response or continuous monitoring.

[0147] (iv) Generating deception attack identification results: It should be noted that the attack determination results are... This represents the "existence" dimension of the attack, and the attack impact index. This represents the "severity" dimension of the attack; the two belong to different dimensions of information. If the attack judgment result is directly... As the final output, the attack impact index The severity information carried will not be able to be passed to subsequent steps, and discussing the degree of attack impact in the absence of an attack is logically untenable. Therefore, this embodiment uses the attack determination result as the basis for its interpretation. As a gating mechanism, it is applied only if a sensor spoofing attack is detected, based on the attack impact index. The impact is classified, and the two dimensions are merged into a unified result for identifying deception attacks. .

[0148] Specifically, define the attack impact level. for:

[0149] in, The level of attack impact; For indicator functions; The preset impact threshold is set according to system security requirements. When... hour, It is always 0, thus avoiding misjudgments about the extent of an attack's impact when no attack exists; when hour, according to Does it exceed The value can be 0 or 1.

[0150] Furthermore, the deception attack identification result is defined. for:

[0151] in, The set of values ​​for the deception attack identification result is as follows: .when When, it indicates that there is no sensor spoofing attack; when At times, sensor spoofing attacks exist and have a low-risk impact on the vehicle's perception system; when This indicates the presence of sensor spoofing attacks and poses a high-risk impact on the vehicle's perception system.

[0152] It is understandable that the results of the deception attack identification are... The attack determination result can be recovered without loss. With respect to the level of attack impact ,Right now ,when hour Therefore, the result of the deception attack identification. It simultaneously and comprehensively characterizes both the existence of the attack and the severity of its impact, thus eliminating the limitations of relying solely on... The problem of overlapping output with intermediate values ​​and information loss exponentially increases the impact of attacks. The information is then fed into the subsequent processing chain.

[0153] Ultimately, based on the attack determination results and attack impact index Generate attack impact level This leads to the generation of deception attack identification results. and the results of the deception attack identification Input step 8), used to update the attack identification model and generate the updated attack identification model. .

[0154] Step 8) Adaptively adjust the perceptual model parameters and attack judgment threshold using the deception attack identification results generated in Step 7) to generate an updated attack identification model.

[0155] This step aims to construct a closed-loop model update mechanism based on detection feedback, enabling the attack identification system to continuously adapt to new deception attack methods and dynamic operating environments. Faced with the continuous evolution of attack techniques and the constant changes in the operating environment, static model parameters and fixed judgment thresholds will lead to a gradual degradation in detection performance. By utilizing the deviation between actual identification results and model prediction results to drive parameter updates, the attack identification model can achieve online self-evolution.

[0156] (a) Calculation of model parameter correction: using the deception attack identification results generated in step 7). Attack identification model Dynamic updates are performed by adjusting model parameters and attack detection thresholds to improve the adaptability of the attack identification model to new deception attack behaviors. First, based on the deception attack identification results... Calculate the model parameter correction amount:

[0157] Model parameter correction amount reflects the direction and magnitude of parameter adjustment due to the deviation between model prediction and actual recognition results; : Parameter adjustment coefficient, which controls the magnitude of the impact of a single update on the model parameters. ; The result of the deception attack identification is the actual attack label determined. or ); Attack identification model The output prediction result is the model's prediction of the attack state under the current parameters.

[0158] Model parameter correction The update logic is: (1) When and hour, The model parameters are adjusted to enhance the sensitivity of attack detection; (2) when and hour, The model parameters are adjusted in the direction of suppressing false alarms; (3) when hour, The model parameters remain unchanged, indicating that the current prediction is accurate.

[0159] (ii) Model parameter update: Further, update the model parameters based on the model parameter correction amount:

[0160] Updated model parameters; The model parameters before the update, i.e., the current attack identification model. The parameters used; : Model parameter correction amount, derived from the previous calculation.

[0161] Parameter updates enable the model to gradually adapt to shifts in the distribution of attack features and changes in sensor performance.

[0162] (III) Dynamic Update of Attack Detection Threshold: To enhance the model's adaptability to changes in the attack environment, the attack detection threshold is dynamically updated.

[0163] Updated attack detection threshold; Attack detection threshold before the update; Threshold adjustment coefficient: controls the step size for threshold updates. ; : Current attack confidence, derived from the calculation result in step 7); Historical average attack confidence level, which is the average of all attack confidence levels within the historical observation period.

[0164] Adaptive logic for dynamic threshold update: (1) When When the current attack confidence level is higher than the historical average, the threshold should be appropriately raised to avoid excessive alarms in high-risk environments; (2) when When the current attack confidence level is lower than the historical average, the threshold is appropriately lowered to enhance the detection sensitivity for potential weak attacks; (3) when When the threshold remains unchanged, the system is in a steady state.

[0165] (iv) Construction of the updated attack identification model: based on the updated model parameters and the updated attack detection threshold Build an updated attack identification model :

[0166] The updated attack identification model has enhanced detection capabilities against new types of deception attacks. : Model building function, which encapsulates model parameters and judgment thresholds into a complete attack identification model; Updated model parameters; : Updated attack detection threshold.

[0167] (v) Closed-loop iteration mechanism: Finally, the updated attack identification model will be... This will be used for the next cycle's sensor spoofing attack detection task, enabling continuous optimization and iterative updates of the attack detection model. The updated attack detection model... It will replace the previous attack detection model in the next detection cycle. It participates in the entire process of cross-modal consistency feature extraction, consistency evaluation, anomaly deviation calculation, risk analysis, and attack determination. Through continuous closed-loop iteration, the system can gradually accumulate experience in recognizing new attack patterns and continuously improve detection accuracy and robustness.

[0168] Step 9) Utilize the updated attack identification model generated in Step 8) Acquire a new round of real-time sensor data, and preprocess it according to the method described in step 2) to form a real-time preprocessed dataset. ; Real-time preprocessing of datasets Substitute the updated model parameters into the cross-modal consistency verification and risk analysis process described in steps 3) to 6), and use the updated model parameters. and the updated attack detection threshold Perform real-time spoofing attack identification and obtain real-time spoofing attack identification results. .

[0169] This step aims to update the attack identification model generated in step 8). This is then deployed in a real-world online identification application, achieving a closed-loop implementation from "model training and updating" to "real-time attack identification." Steps 1) to 8) complete the construction of the attack identification model and adaptive parameter updates, while step 9) utilizes the updated model parameters. Compared with the updated attack detection threshold The new round of real-time sensor data is subjected to end-to-end spoofing attack identification, the final real-time identification result is output, and the identification result is fed back to step 7) to drive the continuous iterative optimization of the model.

[0170] (I) Real-time sensing data acquisition and preprocessing

[0171] After updating the attack identification model in step 8), the system enters the online identification phase. A new round of real-time sensor data is acquired. Raw observation data from cameras, lidar, millimeter-wave radar, global navigation satellite system sensors, and inertial measurement units are collected according to the method described in step 1). Preprocessing is then performed strictly according to the time synchronization, spatial alignment, and standardization processing flow described in step 2), forming a real-time preprocessed dataset. :

[0172] in: Real-time preprocessed dataset: This represents the standardized set of preprocessed observation data from each sensor in the new round of online identification phase. : No. Type of sensor in the first The standardized observations corresponding to each real-time sampling time have the same preprocessing procedure as in step 2). Completely consistent, including timestamp correction, temporal interpolation alignment, spatial interpolation alignment, and Z-score normalization; Sensor category number, with a value range of: This should be consistent with the sensor numbering system in step 1); : Real-time sampling time number, with a value range of ; The total number of sensors is the same as the total number of sensors in step 1). The amount of real-time sampled data is determined by the sampling duration and sampling frequency during the online identification phase.

[0173] Real-time preprocessing datasets The preprocessed dataset output from step 2) They have the exact same data structure and physical semantics, the difference lies in This corresponds to the real-time data in the new round of online recognition phase, and This corresponds to data from the historical training phase. This design ensures data consistency between the model training phase and the online recognition phase, allowing the updated model parameters to... It can be directly applied to the analysis and processing of real-time data.

[0174] (ii) Real-time cross-modal consistency feature extraction

[0175] Real-time preprocessing of datasets Following the method described in step 3), calculate the feature correlation degree between real-time observation data of different modalities, and construct a real-time cross-modal consistent feature set. .

[0176] First, calculate the feature correlation between observation data of different modalities under real-time data:

[0177] in: Real-time data Type of sensor and the first The feature correlation between sensor types is calculated using the same method as in step 3). Consistent, but the data source has been changed to a real-time preprocessed dataset. ; : No. Type of sensor in the first Standardized observations at each real-time sampling time point, derived from the real-time preprocessed dataset. ; : No. Type of sensor in the first Standardized observations at each real-time sampling time point, derived from the real-time preprocessed dataset. ; , : respectively the first Class and First Sensor-like devices in real-time preprocessing of datasets The mean of the above; The amount of real-time sampled data. Based on the real-time feature correlation, a real-time cross-modal consistent feature vector is constructed, forming a real-time cross-modal consistent feature set. :

[0178] in: Real-time cross-modal consistency feature set, containing all cross-modal consistency features under real-time data; : No. A real-time cross-modal consistency feature, constructed in the same way as in step 3). Consistent; : Number of real-time cross-modal consistency features.

[0179] (III) Real-time consistency evaluation and comprehensive anomaly deviation analysis: real-time cross-modal consistency feature set Substitute the cross-modal consistency evaluation and anomaly deviation analysis processes described in steps 4) and 5) in sequence.

[0180] First, calculate the real-time cross-modal characteristic average deviation value according to the method in step 4). :

[0181] in, This represents the average deviation value of real-time cross-modal features. This represents the mean of the real-time cross-modal consistency characteristics. This indicates the number of real-time cross-modal consistency features.

[0182] The real-time consistency evaluation result is calculated based on the average deviation value of real-time cross-modal characteristics. :

[0183] in, This indicates the real-time consistency evaluation results. This represents the average deviation value of real-time cross-modal features. It represents the base of the natural logarithm.

[0184] Subsequently, following the method in step 5), based on the real-time consistency evaluation results... Calculate real-time anomaly deviation results Deviation increment from real-time anomalies And generate real-time comprehensive anomaly deviation results. :

[0185]

[0186]

[0187] in: The real-time comprehensive anomaly deviation results comprehensively reflect the degree of anomaly deviation of cross-modal observation information under real-time data; Real-time abnormal deviation results are derived from real-time consistency evaluation results. Obtained through conversion; Real-time anomaly deviation increment measures the magnitude of change in anomaly deviation results between adjacent real-time moments; , These represent the abnormal deviation results between the current real-time moment and the previous real-time moment, respectively. The weighting coefficients are consistent with those in step 5) and satisfy the following conditions: .

[0188] Real-time comprehensive abnormal deviation results This involves applying the analysis process of steps 4) and 5) to the output of real-time data. It integrates real-time consistency deviation information and real-time temporal mutation information, providing a basis for subsequent use of the updated attack identification model. The attack determination provides core input features.

[0189] (iv) Real-time attack confidence calculation based on the updated model: real-time comprehensive anomaly deviation results Substitute into the updated attack identification model generated in step 8) In the middle, using the updated model parameters Calculate real-time attack confidence :

[0190] in: Real-time attack confidence level, with a value range of [value missing]. The larger the value, the higher the possibility that the current real-time observation data is subject to sensor spoofing attacks; Step 8) generates updated model parameters that include the results of identifying historical deception attacks. The corrected attack identification knowledge is compared to the original model parameters used in step 7). It has a stronger ability to detect attacks; Real-time comprehensive abnormal deviation results, derived from the calculation output of the previous step; The base of the natural logarithm.

[0191] The formula uses the Sigmoid activation function to integrate real-time abnormal deviation results. Updated model parameters Weighted mapping to Interval. When Large (i.e., real-time observation data shows significant cross-modal anomaly deviation) and When attack patterns have been accurately captured through learning, When the value is large, the Sigmoid output approaches 1, indicating a high confidence level for real-time attacks. High; conversely, when real-time observation data has good cross-modal consistency, The value is relatively small, the Sigmoid output is close to 0, and the real-time attack confidence is low. Low.

[0192] (v) Generation of real-time attack determination and identification results based on updated thresholds: based on real-time attack confidence. Compared with the updated attack determination threshold generated in step 8), By comparing and judging, the real-time spoofing attack identification result is obtained. :

[0193] in: The real-time spoofing attack identification results serve as the final identification output of this method. This indicates that a sensor spoofing attack has been detected in the real-time observation data. The system should immediately trigger a security alert and notify the downstream decision-making module to take defensive measures. This indicates that no sensor spoofing attacks were detected in the real-time observation data, and the multi-sensor sensing system is currently operating normally. Real-time attack confidence, derived from the calculation output of the previous step; Step 8) generates the updated attack threshold, compared to the original attack threshold used in step 7). Based on the current attack confidence level Compared with historical average attack confidence The relative relationships have been adaptively adjusted to better adapt to the dynamic changes in the current attack environment.

[0194] Updated attack detection threshold Its role in real-time identification: When the historical attack environment becomes more complex. The sensitivity is adaptively lowered to improve detection sensitivity and reduce the false negative rate; when the historical attack environment is relatively stable, The sensitivity is adaptively increased to suppress false alarms and ensure system stability. This mechanism enables real-time attack detection to achieve a dynamic balance between sensitivity and specificity.

[0195] (vi) Identification result feedback and closed-loop continuous optimization mechanism: Real-time spoofing attack identification results As the final identification output, at the same time Feedback is sent to step 7) for the cumulative update of the next round of deception attack identification results and the continuous adaptive adjustment of model parameters. The specific feedback path is as follows: 1. Feedback to step 7): Real-time spoofing attack identification results The results of the deception attack identification (included in step 7) In this process, the attack identification sample library is expanded to provide richer training signals for subsequent model parameter correction; 2. Re-execution of step 8): Accumulate and update the results of deception attack identification. This will trigger the model parameter correction and threshold update process in step 8) again, generating a new round of updated attack identification model. Make the model parameters Attack detection threshold Continuous evolution; 3. Closed-loop iteration: Updated attack identification model This will be applied again in the next round of real-time recognition in step 9), forming "real-time recognition". Results Feedback Model update A closed-loop continuous optimization mechanism for "real-time identification".

[0196] Through the aforementioned closed-loop iterative mechanism, the system can continuously accumulate new attack identification experience during online operation and continuously improve the updated model parameters. The ability to detect new deception attack patterns, while also adjusting the updated attack detection threshold. Continuously adapt to the dynamic evolution of the attack environment to ensure the accuracy, robustness, and environmental adaptability of the sensor spoofing attack identification system in long-term operation.

Claims

1. A sensor spoofing attack identification method based on cross-modal consistency verification, characterized in that, Includes the following steps: Step 1) Acquire raw sensing data generated by multiple sensors, collect raw observation data, and form a raw sensing dataset. ; Step 2) Preprocess the raw sensor dataset obtained in Step 1) by performing time synchronization, spatial alignment, and standardization to form a preprocessed dataset. ; Step 3) Based on the preprocessed dataset formed in Step 2), extract cross-modal consistency features, construct the correlation between observation results from different sensors, and form a cross-modal consistency feature set. ; Step 4) Establish a cross-modal consistency evaluation model using the cross-modal consistency feature set formed in Step 3), and analyze the cross-modal consistency feature set using the cross-modal consistency evaluation model to generate consistency evaluation results. ; Step 5) Calculate the degree of deviation between the observation information of each sensor using the consistency evaluation results generated in Step 4), and generate a comprehensive anomaly deviation result. ; Step 6) Utilize the comprehensive anomaly deviation results generated in Step 5) to establish an attack risk analysis model, and use the attack risk analysis model to perform risk quantification analysis, generating a normalized attack risk level result. ; Step 7) Use the attack risk level results generated in Step 6) to identify sensor spoofing attack behavior and generate spoofing attack identification results. ; Step 8) Adaptively adjust the perceptual model parameters and attack judgment threshold using the deception attack identification results generated in Step 7) to generate an updated attack identification model. ; Step 9) Utilize the updated attack identification model generated in Step 8) To acquire a new round of real-time sensor data, perform time synchronization, spatial alignment, and standardization preprocessing as described in step 2) to form a real-time preprocessed dataset. ; Real-time preprocessing of datasets Substitute the updated model parameters into the cross-modal consistency verification and risk analysis process described in steps 3) to 6), and use the updated model parameters. and the updated attack detection threshold Perform real-time spoofing attack identification and obtain real-time spoofing attack identification results. .

2. The sensor spoofing attack identification method based on cross-modal consistency verification according to claim 1, characterized in that, The specific method in step 1) is as follows: Raw observation data is collected and organized according to sensor type to form a raw sensing dataset. Its expression is: in, Represents the original sensory dataset, Indicates the first The raw observation data corresponding to the sensor type, Indicates the total number of sensors. Indicates the sensor number; The original observation data Represented as: in, Indicates the first The raw observation data corresponding to the sensor type, Indicates the first Type of sensor in the first The observations obtained at each sampling time point Indicates the number of sampled data. Indicates the sampling time number; To quantify the data integrity of various sensors and calculate data efficiency : in, Indicates the first The data efficiency of sensor-like devices Indicates the first The number of effective observation data of this type of sensor Indicates the first The total number of observation data of this type of sensor; when Greater than the preset efficiency threshold If the data is in the correct state, retain the corresponding sensor data; otherwise, mark it as an abnormal data source.

3. The sensor spoofing attack identification method based on cross-modal consistency verification according to claim 1, characterized in that, The specific method in step 2) is as follows: Using the raw sensory dataset formed in step 1) Time synchronization, spatial coordinate alignment, and standardization are performed to form a preprocessed dataset. ; First, let the first... Sensor-like sensors The original observation sequence is as follows: Indicates the first Type of sensor in the first The original timestamps of each sampling time point ( ), Indicates the number of sampled data. This indicates the sampling time number; due to inconsistent sampling frequencies and clock skew among different data sources, the sampling times of all sensors are corrected to a unified time grid. ,in To unify time grid points; Based on the timestamps of each sensor, the data correspondence under a unified time reference is calculated. First, the original timestamps are corrected sample by sample, and the time synchronization result is expressed as follows: in, Indicates the first Type of sensor in the first The correction timestamp after synchronization of each sampling time point Indicates the first Type of sensor in the first The original timestamps of each sampling time point Indicates the first The time offset of a sensor relative to a unified time base. To unify time grid points; After timestamp correction, the first The observed values ​​of the sensor remain unchanged, only the time stamp is replaced with a calibration timestamp. The synchronized sequence is denoted as: For the first Sensor-like devices perform time alignment, i.e., correcting timestamps. For interpolation nodes, for the synchronized sequence Construct a time-dimensional interpolation function and apply it at a unified time grid point. The value is evaluated to obtain the time-aligned data: in, For time-dimensional interpolation operators (linear interpolation or cubic spline interpolation). To correct timestamps This is the sequence after the nodes are synchronized. To unify time grid points, Indicates the first Sensors at a unified time grid point Time alignment value at; Subsequently, the time alignment value was adjusted. Perform spatial alignment processing, let the first... There are 1 sensor, and the number of sampled data is 1 The original set of space observation points is: Its spatial distribution needs to be aligned with a unified spatial grid. Maintain consistency; Based on time alignment, further refine the first... Each sensor performs spatial alignment, Mapped to a unified spatial grid point The time-space joint alignment number is obtained as follows: in, For spatial interpolation operators, Indicates the first After time and spatial alignment, the sensor data is placed in a unified spatial grid. And unified time grid points Data values ​​at; Subsequently, the synchronized observation data were standardized, and the standardization process focused on the temporal-spatial joint alignment number after temporal and spatial alignment. ; For the first Temporal-spatial joint alignment of individual sensors Calculate its mean and standard deviation in a unified spatiotemporal domain: in, and The first Each sensor at a unified spatial grid point Place, across all The mean and standard deviation of each unified time step; the calculation scope of the above statistics is limited to the aligned time-space joint alignment number. superior; Based on this, Z-score standardization is performed to obtain the standardized data: in, Indicates the first Type of sensor in the first Standardized observations corresponding to each sampling time; For the first Temporal-spatial joint alignment number of sensor-like devices; Indicates the first Mean value of sensor-like observation data; Indicates the first Standard deviation of sensor-like observation data; ; The various types of preprocessed observation data are organized in a unified manner to form a preprocessed dataset. : in, Indicates the first Type of sensor in the first Standardized observations corresponding to each sampling time. This indicates a preprocessed dataset. This indicates the total number of sensors.

4. The sensor spoofing attack identification method based on cross-modal consistency verification according to claim 1, characterized in that, The specific method in step 3) is as follows: Using the preprocessed dataset formed in step 2) Cross-modal consistency features are extracted by analyzing the positional, motion, and state correlations among observations from different sensors to construct a cross-modal consistency feature set. ; First, calculate the feature correlation between observation data from different modalities: in, Indicates the first Type of sensor and the first Feature correlation between sensor types Indicates the first The first type of sensor One preprocessed data, Indicates the first The first type of sensor One preprocessed data, Indicates the first Mean of preprocessed data for sensor-like devices Indicates the first Mean of preprocessed data for sensor-like devices Indicates the number of samples; Construct cross-modal consistent feature vectors based on feature correlation: in, Indicates the first Cross-modal consistency features Indicates the first Type of sensor and the first Feature correlation between sensor types Indicates the first The preprocessed data set corresponding to the sensor type Indicates the total number of sensors; All cross-modal consistency features are combined to form a cross-modal consistency feature set. : in, Represents a cross-modal consistency feature set. Indicates the first Cross-modal consistency features This indicates the number of cross-modal consistency features.

5. The sensor spoofing attack identification method based on cross-modal consistency verification according to claim 1, characterized in that, The specific method in step 4) is as follows: Using the cross-modal consistency feature set formed in step 3) A cross-modal consistency evaluation model is established. By calculating the degree of matching and deviation between different modal features, the cross-modal consistency features are quantitatively evaluated, and consistency evaluation results are generated. ; First, calculate the average deviation between cross-modal consistency features: Where δ represents the average deviation value of the cross-modal characteristics. Indicates the first Cross-modal consistency features Indicates the first Cross-modal consistency features Indicates the number of cross-modal consistency features; The consistency evaluation results are calculated based on the average deviation value of cross-modal characteristics: in, The result represents the consistency evaluation result, where δ represents the average deviation of cross-modal characteristics, and e represents the base of the natural logarithm. When the observation deviation between different modal observation results is within the preset normal deviation range [0, 0.2], δ decreases, and the consistency evaluation result improves. Increase; when the observation deviation between different modal observation results exceeds the preset normal deviation range [0, 0.2], δ increases, and the consistency evaluation results... Decrease.

6. The sensor spoofing attack identification method based on cross-modal consistency verification according to claim 1, characterized in that, The specific method in step 5) is as follows: Using the consistency evaluation results generated in step 4) The degree of deviation between observation information from different modalities is calculated, the consistency status of cross-modal observation results is quantitatively analyzed, and a comprehensive anomaly deviation result is generated. ; First, based on the consistency evaluation results Calculate abnormal deviation results : in, Indicates abnormal deviation from the result. This indicates the consistency evaluation results; To enhance the sensitivity of abnormal deviation results to local abnormal fluctuations, the abnormal deviation increment is calculated: in, Indicates time The corresponding abnormal deviation increment, Indicates time The corresponding abnormal deviation results Indicates time Corresponding abnormal deviation results; Based on abnormal deviation results Incremental deviation from abnormality Calculate the overall anomaly deviation results: in, This indicates the overall abnormal deviation results. Indicates abnormal deviation from the result. This indicates an abnormal deviation from the increment. This represents the weighting coefficient, and satisfies 0 < <1; When the overall abnormal deviation from the result Exceeding the preset abnormal threshold If the current cross-modal observation information shows an abnormal deviation trend, it is determined that the current cross-modal observation information remains in a normal and consistent state; otherwise, it is determined that the current cross-modal observation information remains in a normal and consistent state.

7. The sensor spoofing attack identification method based on cross-modal consistency verification according to claim 1, characterized in that, The specific method in step 6) is as follows: Using the comprehensive anomaly deviation results generated in step 5) An attack risk analysis model is established, which quantifies the risk of sensor spoofing attacks by fusing real-time anomaly deviation information with historical anomaly behavior information, and generates a normalized attack risk level result. ; First, calculate the frequency of historical anomalies: in, Indicates the frequency of historical anomalies. Indicates the number of historical anomalies. This represents the total number of historically observed events; Based on the comprehensive abnormal deviation results and frequency of historical anomalies Calculate the attack risk level result : in, This indicates the attack risk level result. This indicates the overall abnormal deviation results. Indicates the frequency of historical anomalies. Indicates the weight of abnormal deviations from the result. Represents the weight of historical abnormal frequencies, and satisfies: To facilitate subsequent risk level classification, the attack risk level results are... Normalization is performed: in, This represents the normalized attack risk level result, where R represents the attack risk level result. This represents the historical maximum risk value. This represents the historical minimum risk value.

8. The sensor spoofing attack identification method based on cross-modal consistency verification according to claim 1, characterized in that, The specific method in step 7) is as follows: Using the normalized attack risk level result generated in step 6) To identify sensor spoofing attacks, an attack determination mechanism is constructed to classify and judge the current observation state, generating spoofing attack identification results. ; First, based on the normalized attack risk level results Calculate attack confidence: in, Indicates the attack confidence level. This indicates the normalized attack risk level result. Represents the base of the natural logarithm; Based on attack confidence Perform attack detection: in, This indicates the attack determination result. =1 indicates the presence of a sensor spoofing attack. =0 indicates that there is no sensor spoofing attack. Indicates the attack confidence level. Indicates the attack detection threshold; To assess the extent of the attack's impact, an attack impact index is calculated: in, Indicates the impact index of the attack. Indicates the attack confidence level. This represents the normalized attack risk level result. When the attack impact index Exceeding the preset impact threshold When the attack behavior is deemed to have a high-risk impact on the perception system, the attack impact index is determined to be high. Not exceeding the preset impact threshold At that time, it is determined that the current attack behavior has a low-risk impact on the perception system; Based on the attack determination result and the attack impact index Generate deception attack identification results ,include: Calculate the impact level of the attack : in For indicator functions, This is a preset impact threshold; Calculate the deception attack identification result in, The set of values ​​is ,when When, it indicates that there is no sensor spoofing attack; when At times, sensor spoofing attacks exist and have a low-risk impact on the vehicle's perception system; when At that time, it indicates the existence of sensor spoofing attacks and a high-risk impact on the vehicle's perception system; ultimately, based on the attack determination results... and attack impact index Generate attack impact level Generate deception attack identification results .

9. The sensor spoofing attack identification method based on cross-modal consistency verification according to claim 1, characterized in that, The specific method in step 8) is as follows: Using the deception attack identification results generated in step 7) Attack identification model Dynamic updates are performed by adjusting model parameters and attack detection thresholds to improve the attack identification model's adaptability to new deceptive attack behaviors, generating an updated attack identification model. ; First, based on the results of the deception attack identification Calculate the model parameter correction amount: in, This represents the amount of model parameter correction. This represents the parameter adjustment coefficient. This indicates the results of the deception attack identification. Indicating attack identification model The output prediction results; Update the model parameters based on the model parameter corrections: in, This represents the updated model parameters. This represents the model parameters before the update. This indicates the amount of model parameter correction; To enhance the model's adaptability to changes in the attack environment, the attack detection threshold is dynamically updated: in, This indicates the updated attack detection threshold. This represents the attack detection threshold before the update, and γ represents the threshold adjustment coefficient. Indicates the current attack confidence level. This represents the historical average attack confidence level; Based on the updated model parameters and the updated attack detection threshold Build an updated attack identification model : in, This indicates the updated attack detection model. θ' represents the model building function, and θ' represents the updated model parameters. This indicates the updated attack detection threshold.

10. The application of the sensor spoofing attack identification method based on cross-modal consistency verification according to any one of claims 1-9 in intelligent vehicles, characterized in that, In step 1), raw perception data generated by the multi-source sensors of the intelligent vehicle is acquired, including raw observation data from cameras, LiDAR, millimeter-wave radar, global navigation satellite system sensors, inertial measurement units, and on-board perception devices, forming a raw perception dataset. .

Images