Internet of vehicles CAN bus intrusion detection method based on noise perception active learning

Abstract

The invention discloses an Internet of Vehicles CAN bus intrusion detection method based on noise perception active learning, and belongs to the technical field of Internet of Vehicles safety and machine learning. The invention aims to solve the technical problems of false label noise interference, high manual labeling cost, high attack missing report rate caused by class imbalance and the like. The core of the method is to execute a noise sensing mixed query strategy in an iterative loop: firstly, generating a pseudo tag through clustering and correcting by using an integrated noise detector; secondly, calculating uncertainty scores and noise probabilities of the samples, fusing the uncertainty scores and the noise probabilities to obtain a comprehensive score, and preferentially selecting the samples with high uncertainty and low noise probabilities; and then adaptively selecting a sampling strategy according to the model performance and applying category balance constraint. The query batch is used to iteratively update the model while dynamically adjusting the classification threshold to reduce the missing report rate. According to the method, the influence of pseudo label noise can be effectively suppressed, and the attack detection precision and generalization capability are remarkably improved with extremely low labeling cost.

Description

Technical Field

[0001] This invention belongs to the field of vehicle network security and machine learning technology, specifically relating to a high-efficiency abnormal traffic classification method based on noise perception active learning for intrusion detection in vehicle CAN bus networks. Applicable fields include, but are not limited to, intelligent connected vehicles, autonomous driving systems, communication security of onboard electronic control units (ECUs), V2X security protection for vehicle networks, and CAN network security monitoring for industrial control systems. Background Technology

[0002] In the context of the rapid development of modern intelligent connected vehicles, the Controller Area Network (CAN), as the core bus protocol for communication between electronic control units (ECUs) within the vehicle, carries data exchange for critical functions such as engine control, braking system, and steering system. However, the CAN bus protocol was not designed with network security in mind, lacking security mechanisms such as authentication, data encryption, and access control, making it highly vulnerable to malicious attacks. Common CAN bus attacks include denial-of-service (DoS) attacks, fuzzy attacks, replay attacks, and message forgery attacks. These attacks can lead to loss of vehicle control, malfunctions, and even endanger the lives of passengers.

[0003] Currently, intrusion detection technology based on deep learning algorithms is being gradually applied to the field of CAN bus security. High-precision anomaly detection models are built through supervised learning on a large amount of labeled normal and attack traffic data. While it has achieved some success in detection accuracy, the "large-scale data labeling" model upon which its success depends also constitutes the biggest bottleneck in its actual deployment in the field of vehicle network security, specifically manifested in the following common challenges: (1) High annotation costs and reliance on experts: Training a high-performance CAN bus intrusion detection model requires massive amounts of traffic data that are precisely annotated by cybersecurity experts. CAN bus data is characterized by high speed and high traffic volume, and millions of message records can be generated in a single drive, making manual annotation of each message almost impossible. At the same time, there may be high similarity in features between different attack types, or the attack behavior may be extremely covert, all of which require annotators to have in-depth knowledge of vehicle network security, resulting in extremely high manpower and time costs; (2) Common class imbalance problem: During normal vehicle operation, the number of legitimate CAN messages usually far exceeds the total number of attack messages, and the ratio of normal traffic to attack traffic may reach 100:1 or even higher. At the same time, the frequency of different attack types also varies greatly. This inherent data imbalance will cause the model to be severely biased towards the dominant normal class during training, resulting in poor ability to identify rare but critical attack classes, leading to a high false negative rate. A single missed attack may have catastrophic consequences. (3) Noise from false labels: To reduce labeling costs, existing methods often use unsupervised learning techniques such as clustering to generate false labels as training data. However, clustering algorithms divide data based on assumptions about the data distribution, which cannot guarantee that the generated false labels are completely consistent with the real labels, inevitably introducing label noise. These noisy labels can mislead the model's learning, leading to a decrease in detection performance. In particular, when attack samples are incorrectly labeled as normal, the system's false negative risk will be significantly increased. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a noise-aware active learning-based intrusion detection method for CAN bus in the Internet of Vehicles (IoV). This method first uses a clustering algorithm to generate pseudo-labels for each attack type dataset. Then, an integrated noise detector is constructed to correct the pseudo-labels. Finally, a hybrid query strategy integrating "uncertainty" and "noise awareness" dimensions is designed for active learning. This hybrid query strategy first uses a classification model to calculate uncertainty scores based on the predicted probabilities of unlabeled samples. Simultaneously, it uses an integrated noise detector to evaluate the reliability of sample pseudo-labels, prioritizing samples with high uncertainty and low noise probability for training. The sampling strategy is dynamically adjusted based on the current model performance, and class balance constraints are applied during the sampling process to alleviate class imbalance.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] A method for intrusion detection on a vehicle-to-everything (V2X) CAN bus based on noise perception active learning includes the following steps: Step 1: Acquisition and processing of CAN bus data sets; Step 2: Generation of pseudo-labels based on clustering; Step 3: Construct an integrated noise detector; Step 4: Seed pool initialization and noise correction; Step 5: Construct an initial active learning classification model; Step 6: Select samples using a noise-aware hybrid query strategy; Step 7: Model retraining and dynamic threshold adjustment; Step 8: Iterate repeatedly to achieve optimal detection performance.

[0007] The specific details of step one are as follows: CAN bus communication data is collected in a real vehicle operating environment through the on-board diagnostic interface or by directly accessing the CAN bus network. The collected data covers four typical attack scenarios: denial-of-service attack (DoS), fuzzy attack, gear spoofing attack, and RPM spoofing attack, and also includes communication data under normal driving conditions. Each CAN message record includes original characteristic fields such as timestamp, CAN identifier (CAN ID), data length code (DLC), and 8-byte data payload (DATA[0-7]) to ensure data integrity and temporal continuity.

[0008] The specific details of step two are as follows: Preprocessing and pseudo-label generation are performed on the datasets for each attack type. First, the original CAN bus data is normalized using Min-Max to scale each feature value to the [0,1] interval, eliminating the influence of differences in feature dimensions. Then, standardization is performed to make the data conform to a zero-mean, unit-variance distribution. Next, principal component analysis (PCA) is used for dimensionality reduction, retaining principal components with a cumulative variance contribution rate of 95%. Finally, the K-Means clustering algorithm is used to perform unsupervised clustering on the dimensionality-reduced data. Based on the maximum cluster mapping strategy, the clustering results are converted into binary pseudo-labels, where the cluster with the most samples is labeled as normal traffic, and the remaining clusters are labeled as attack traffic. After completing the above processing on the datasets for the four attack types, they are merged to generate a unified training dataset with pseudo-labels.

[0009] The specific details of step three are as follows: A weighted integrated noise detector is constructed to evaluate the reliability of fake tags and identify noise samples. This detector integrates four complementary sub-detectors: a CAN protocol consistency detector detects abnormal samples that violate protocol rules by learning the time interval, data length, and data field pattern of normal CAN communication; an isolated forest detector identifies outliers in the feature space based on the idea of ​​random partitioning; a denoising autoencoder learns the low-dimensional representation of normal data through a neural network and uses reconstruction error to identify abnormal samples; and a K-nearest neighbor distance detector identifies abnormal points that are far away from normal samples based on the Euclidean distance relationship between samples. The abnormal scores of the four sub-detectors are weighted and fused to obtain the comprehensive noise probability of each sample.

[0010] The specific details of step four are as follows: A certain number of samples are randomly selected from the training dataset with pseudo-labels to construct an initial Seed pool. An ensemble noise detector is used to correct the pseudo-labels in the Seed pool. The correction strategy is as follows: When the ensemble detector determines that a sample is an attack with a probability higher than a set confidence threshold, but its pseudo-label is normal, the label of the sample is corrected to attack to deal with the false negative noise generated by clustering; when the ensemble detector determines that a sample is normal with a probability higher than a set confidence threshold, but its pseudo-label is attack, the label of the sample is corrected to normal to deal with the false positive noise generated by clustering. Samples are extracted from the remaining data according to a set ratio to construct an unlabeled Pool for subsequent active learning queries.

[0011] The specific details of step five are as follows: the initial classification model is trained using the corrected Seed pool data, and an active learner is constructed. The classification model supports multiple classifiers such as random forest, convolutional neural network, long short-term memory network or TabNet; the active learning query strategy supports uncertainty sampling, marginal sampling or entropy sampling; at the same time, a performance monitor is initialized to track the trend of the model's precision, recall and F1 score on the validation set. When the performance is detected to continuously decline beyond the set threshold, the strategy is adjusted or the model is reinitialized.

[0012] The specific details of step six are as follows: A hybrid query strategy integrating "uncertainty" and "noise perception" is designed for sample selection. First, the uncertainty score is calculated based on the predicted probability of unlabeled samples in the Pool using the current classification model, and the candidate samples with the most ambiguous judgment by the model are selected. Then, an integrated noise detector is used to evaluate the reliability of the pseudo-labels of the candidate samples and calculate the noise probability. The uncertainty score and noise probability are fused to calculate a comprehensive score, and samples with high uncertainty and low noise probability are selected first to ensure that the selected samples can both resolve the current cognitive ambiguity of the model and have high label reliability. At the same time, the sampling strategy is dynamically adjusted according to the current model performance: when the false negative rate is too high, an attack enhancement sampling strategy is adopted, and boundary attack samples are selected first. When the performance continues to decline, a conservative sampling strategy is adopted. When the overall performance is poor, a balanced backoff sampling strategy is adopted. During the sampling process, a class balance constraint is applied to ensure the minimum proportion of attack samples and alleviate the class imbalance problem. When the false negative rate exceeds a set threshold, a false negative-targeted learning mechanism is activated to find samples with similar features to the false negative samples in the Pool and add them to the training batch first.

[0013] The specific details of step seven are as follows: high-value samples selected by the hybrid query strategy are added to the existing training set to form an expanded training set. The classification model is retrained using the expanded complete training set so that the model can learn more discriminative feature representations. At the same time, the classification threshold is dynamically adjusted based on the precision-recall curve of the validation set to find the optimal threshold that maximizes the F1 score, thereby reducing the false negative rate while ensuring precision. After training is completed, the selected samples are removed from the pool.

[0014] The specific details of step eight are as follows: The entire method is driven by a closed-loop iterative cycle of "query-correction-retraining-evaluation". In each round of the cycle, the model's performance metrics such as precision, recall, F1 score, accuracy, false positive rate, and false negative rate are first evaluated on the validation set. The performance monitor continuously tracks changes in the F1 score. When the performance is detected to have decreased continuously for more than a set number of times, the model is triggered to reinitialize in order to escape the local optimum. The iteration continues until the preset number of iterations is reached or the pool of samples is exhausted, and finally, a binary classification model with optimal intrusion detection performance is output.

[0015] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: (1) Significantly reduce labeling costs: Pseudo-labels are automatically generated through clustering algorithms, and the most valuable samples are selected by active learning. This greatly reduces the reliance on manual labeling. The detection performance can be similar to or even higher with a much smaller amount of labeling required by traditional supervised learning methods, directly reducing the manpower and time costs of deploying vehicle network safety systems. (2) Effectively suppressing false label noise: By constructing an integrated noise detector that integrates CAN protocol consistency detection, isolated forest, denoising autoencoder and K nearest neighbor distance detection, it can effectively identify and correct noise labels generated by clustering, avoid noise labels misleading model learning, and significantly improve the reliability of the detection model; (3) Effectively alleviate class imbalance problem: By applying class balance constraints during the sampling process, adopting attack enhancement sampling strategy and false negative targeted learning mechanism, it is ensured that the attack samples are fully learned during the training process, effectively reducing the false negative rate and improving the detection capability of attack traffic; (4) Adaptive performance optimization: The performance monitor tracks the changes in model performance in real time and dynamically adjusts the sampling strategy and classification threshold according to the current precision, recall and false negative rate, so that the system can adaptively optimize the detection performance and maintain good detection results under different data distributions. Attached Figure Description

[0016] Figure 1 This is an overall flowchart of the present invention; Figure 2This is the active learning structure of the noise-aware hybrid query strategy of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. This invention provides a CAN bus intrusion detection method based on noise perception active learning. This method significantly improves the labeling quality and cost-effectiveness of pseudo-label data, enhances the model's generalization ability and robustness in noisy label environments, and also improves the model's detection accuracy for vehicle network attacks.

[0018] Figure 1 The flowchart illustrates the method of this invention, which fully describes a systematic, closed-loop intelligent intrusion detection method. The method begins with standardized data acquisition and preprocessing of raw CAN bus data. Pseudo-labels are generated using the K-Means clustering algorithm, and then the pseudo-labels are corrected using an integrated noise detector to train an initial classification model capable of distinguishing between normal and attack traffic. This classification model is then placed in a core iterative optimization process. In this process, the method employs an innovative noise-aware hybrid query strategy: first, candidate samples that the model finds most confusing are selected through uncertainty measurement; then, a comprehensive score is fused using the noise probability calculated by the integrated noise detector; and finally, a sampling strategy is adaptively selected based on the current model performance to choose a query batch that is both representative and balanced in terms of class, with the highest information value. This query batch is used to iteratively retrain the classification model, thereby continuously enhancing its ability to automatically identify attack traffic. The entire "query-sampling-retraining" iterative cycle continues until the model classification performance reaches a preset indicator or the query budget is exhausted, ultimately achieving efficient and accurate CAN bus intrusion detection at the most economical labeling cost.

[0019] The specific implementation steps are as follows: Step 1.1 By deploying data acquisition equipment on the vehicle CAN bus network, large-scale, standardized continuous data acquisition of CAN bus messages is carried out. The acquisition content includes timestamp, CAN ID, data length code (DLC) and 8-byte data payload (DATA[0-7]). Step 1.2 Strictly maintain consistency in data collection conditions to ensure that the collection process covers all known attack types (including DoS attacks, Fuzzy attacks, RPM spoofing attacks, Gear spoofing attacks, etc.) as well as legitimate CAN messages under normal driving conditions, laying a solid foundation for building a comprehensive and unbiased dataset in the future.

[0020] Step 2.1 The Min-Max normalization technique is used to uniformly convert the collected raw CAN bus data to the [0,1] range. This operation can effectively preserve the key structural features of CAN messages, such as timing features and identifier features, while eliminating the impact of differences in the dimensions of different features. Step 2.2 Further standardize the normalized data to make it conform to a standard normal distribution, which will facilitate subsequent PCA dimensionality reduction processing; Step 2.3 Principal Component Analysis (PCA) is used for feature dimensionality reduction. The cumulative variance contribution rate threshold is set to 95%, and the number of principal components to be retained is automatically determined. This significantly reduces the data dimensionality and computational complexity of subsequent model processing while retaining the main information of the data.

[0021] Step 3.1 From the preprocessed dataset, use the K-Means clustering algorithm to perform unsupervised clustering on all samples, dividing the data into K clusters (in this embodiment, K=2, corresponding to normal traffic and attack traffic respectively). Step 3.2 Select K initial cluster centers using a random initialization strategy, and select the best clustering result through multiple runs to avoid getting trapped in local optima; Step 3.3 Based on the clustering results, the cluster labels are converted into binary pseudo-labels using the maximum cluster mapping strategy: the cluster with the most samples is labeled as normal traffic (label 0), and the remaining clusters are labeled as attack traffic (label 1), thus generating the initial pseudo-label dataset.

[0022] Step 4.1 Construct a weighted integrated noise detector, which consists of four complementary sub-detectors: (1) Denoising Autoencoder: It learns a low-dimensional representation of data through a neural network and uses reconstruction error to identify abnormal samples. Its network structure includes an encoder and a decoder. Noise is added during training to enhance the noise resistance of the model. (2) Isolation Forest: An anomaly detection algorithm based on random forest, which evaluates the degree of anomaly by calculating the average path length required for a sample to be isolated, and sets the pollution ratio to 0.1; (3) CAN Consistency Checker: Based on the characteristics of the CAN bus protocol, it learns the normal message pattern, including message time interval, data length code validity, data field value range, etc. It does not rely on tag information and is completely based on data statistical characteristics for detection. (4) Distance-based detector: The K-nearest neighbor algorithm is used to calculate the average distance from a sample to its nearest neighbor. The larger the distance, the more likely the sample is to be an anomaly. Step 4.2 Assign weights to the four sub-detectors and obtain the final noise probability score through weighted fusion.

[0023] Step 5.1 From the preprocessed dataset, randomly select a certain number of samples (50,000 in this example) as the initial seed pool. Step 5.2 Use an integrated noise detector to evaluate the noise probability of the samples in the Seed pool and obtain the integrated noise score for each sample; Step 5.3 Correct the pseudo-labels based on the noise probability and the correction confidence threshold (0.8 in this embodiment): When there is a high confidence conflict between the prediction probability of the ensemble detector and the original pseudo-label (i.e., the ensemble score is greater than the correction confidence and the original label is 0, or the ensemble score is less than 1 minus the correction confidence and the original label is 1), the pseudo-labels are corrected to the prediction result of the ensemble detector, thereby purifying the noisy labels in the Seed pool.

[0024] Step 6.1 Select Random Forest as the base classifier and leverage its ensemble learning properties to improve the model's generalization ability and robustness to noise. Step 6.2 Use the purified Seed pool data to train the initial classification model, use the modAL active learning framework to build the ActiveLearner, and use the uncertainty sampling query strategy; Step 6.3 Randomly select samples from the remaining data in proportion (20% in this embodiment) to construct a candidate sample pool (Pool) as the source of query samples in the active learning iteration process.

[0025] Figure 2This is a detailed flowchart of the noise-aware hybrid query strategy of this invention. It filters the most perplexing candidate samples through uncertainty measurement, then combines this with the noise probability of an integrated noise detector for comprehensive scoring and fusion, and adaptively selects a sampling strategy based on the current model performance to choose a query batch that is both highly informative and maintains class balance. This process is the core innovation of this invention, aiming to efficiently and unbiasedly select the most valuable samples from the candidate sample pool. It is achieved through a multi-stage screening process, ensuring that the final selected samples possess both "high uncertainty" and "low noise probability."

[0026] The specific implementation steps are as follows: Step 1.1 First Stage: Uncertainty-Based Candidate Sample Evaluation. Firstly, this invention utilizes the currently trained classification model to evaluate the uncertainty of all unlabeled samples in the candidate sample pool. Prediction entropy is used as the core metric for uncertainty. Specifically, for each unlabeled sample... The probability distribution of a sample belonging to each category is predicted using a classification model. The uncertainty score of this sample can be calculated using the following formula: ; in: This represents the total number of all preset categories (normal traffic and attack traffic). Representative model judges samples Category The probability of a sample being classified into any one category. Entropy measures the degree of disorder in the probability distribution predicted by a model. A higher entropy value means that the model cannot definitively classify a sample into any category, i.e., it has higher uncertainty and higher learning value. Step 1.2 Second Stage: Noise Probability Assessment Based on Integrated Detector. While calculating the uncertainty score, this method introduces an integrated noise detector to evaluate the label reliability of candidate samples. The integrated noise detector consists of four complementary sub-detectors: a denoising autoencoder, an Isolation Forest, a CAN Consistency Checker, and a distance-based detector. For each candidate sample... The integrated detector outputs its noise probability. This represents the probability that the sample label may be incorrect. The final noise probability is obtained by weighted fusion of the outputs of each sub-detector: ; in: Representing the Each sub-detector on the sample Abnormal scores, The weights represent the weights of the sub-detector. Step 1.3 Third Stage: Comprehensive Score Fusion and Adaptive Strategy Selection. After obtaining the uncertainty score and noise probability, this method fuses the two to generate a comprehensive score. The core idea of ​​the fusion strategy is to select samples with high uncertainty and low noise probability, i.e., samples that have both learning value and reliable labels. The formula for calculating the comprehensive score is: ; in: These are noise weighting coefficients, dynamically adjusted based on the current model performance. When the false negative rate (FNR) is high, Increase to 0.4 to enhance quality control; when the false alarm rate is high, Reduce to 0.1 to avoid over-filtering; when performance is balanced, Take 0.25; Meanwhile, this method adaptively selects the sampling strategy based on the current model's performance metrics (precision, recall, F1 score, and false negative rate): (1) When the false negative rate (FNR) > 0.7, an attack-enhanced sampling strategy is adopted, focusing on boundary attack samples; (2) When performance continues to decline, a conservative sampling strategy is adopted, and sampling is performed only based on uncertainty; (3) When the overall performance is poor (F1 < 0.2), a balanced backsampling strategy is adopted to ensure class balance; (4) Under normal circumstances, a performance-aware noise-guided sampling strategy is adopted; Step 1.4 Fourth Stage: Class Balance Constraints and Final Batch Construction. To directly address the class imbalance problem caused by the scarcity of attack samples in CAN bus data, this method applies class balance constraints during the sampling process. Specifically, candidate samples are divided into normal sample sets according to their labels. and attack sample set Samples are selected from two sets, sorted by their overall scores. Let the batch size be... The minimum attack sample ratio is (Default 0.25), then: ; From attack sample set Select the one with the highest overall score One sample, from the normal sample set Select the one with the highest overall score Each sample is merged to form the final query batch. : ; This hierarchical sampling mechanism forcibly ensures that representative samples can be selected even from the minority class (attack samples), thus directly and effectively solving the sampling bias problem caused by data imbalance in traditional active learning methods. At the same time, it ensures the reliability of the labels of the selected samples through noise probability filtering.

Claims

1. A method for intrusion detection of a vehicle-to-everything (V2X) CAN bus based on noise perception active learning, characterized in that... Includes the following steps: Step 1: Acquisition and processing of CAN bus data sets; Step 2: Generation of pseudo-labels based on clustering; Step 3: Construct an integrated noise detector; Step 4: Seed pool initialization and noise correction; Step 5: Construct the initial active learning classification model; Step 6: Sample selection based on a noise-aware hybrid query strategy; Step 7: Model retraining and dynamic threshold adjustment; Step 8: Iterate repeatedly until optimal performance is achieved.

2. The intrusion detection method for CAN bus in vehicle networking based on noise perception active learning according to claim 1, characterized in that: The specific process in Step 1 is as follows: CAN bus communication data is collected in a real vehicle operating environment through the on-board diagnostic interface or by directly accessing the CAN bus network. The collected data covers four typical attack scenarios: denial-of-service attack (DoS), fuzzy attack, gear spoofing attack, and RPM spoofing attack, and also includes communication data under normal driving conditions. Each CAN message record contains original characteristic fields such as timestamp, CAN identifier (CANID), data length code (DLC), and 8-byte data payload (DATA[0-7]) to ensure data integrity and temporal continuity.

3. The intrusion detection method for CAN bus in vehicle networking based on noise perception active learning according to claim 1, characterized in that: The specific process in Step 2 is as follows: Preprocessing and pseudo-label generation are performed on the datasets for each attack type. First, the original CAN bus data is normalized using Min-Max, scaling each feature value to the [0,1] interval to eliminate the influence of differences in feature dimensions. Then, standardization is performed to make the data conform to a zero-mean, unit-variance distribution. Next, Principal Component Analysis (PCA) is used for dimensionality reduction, retaining principal components with a cumulative variance contribution rate of 95%. Finally, the K-Means clustering algorithm is used to perform unsupervised clustering on the dimensionality-reduced data. Based on the maximum cluster mapping strategy, the clustering results are converted into binary pseudo-labels, where the cluster with the most samples is labeled as normal traffic, and the remaining clusters are labeled as attack traffic. After completing the above processing on the datasets for the four attack types, they are merged to generate a unified training dataset with pseudo-labels.

4. The intrusion detection method for CAN bus in vehicle networking based on noise perception active learning according to claim 1, characterized in that: The specific process in Step 3 is as follows: A weighted integrated noise detector is constructed to evaluate the reliability of fake tags and identify noise samples. This detector integrates four complementary sub-detectors: a CAN protocol consistency detector detects abnormal samples that violate protocol rules by learning the time interval, data length and data field pattern of normal CAN communication; an isolated forest detector identifies outliers in the feature space based on the idea of ​​random partitioning. Denoising autoencoders learn low-dimensional representations of normal data through neural networks and use reconstruction errors to identify abnormal samples. The K-nearest neighbor detector identifies outliers that are far from normal samples based on the Euclidean distance between samples. It then weights and fuses the outlier scores from the four sub-detectors to obtain the overall noise probability for each sample.

5. The intrusion detection method for CAN bus in vehicle networking based on noise perception active learning according to claim 1, characterized in that: The specific process in Step 4 is as follows: A certain number of samples are randomly selected from the training dataset with pseudo-labels to construct an initial Seed pool. An ensemble noise detector is used to correct the pseudo-labels in the Seed pool. The correction strategy is as follows: When the ensemble detector determines that a sample is an attack with a probability higher than the set confidence threshold, but its pseudo-label is normal, the label of the sample is corrected to attack to deal with the false negative noise generated by clustering; when the ensemble detector determines that a sample is normal with a probability higher than the set confidence threshold, but its pseudo-label is attack, the label of the sample is corrected to normal to deal with the false positive noise generated by clustering. Samples are extracted from the remaining data according to a set ratio to construct an unlabeled Pool for subsequent active learning queries.

6. The intrusion detection method for CAN bus in vehicle networking based on noise perception active learning according to claim 1, characterized in that: The specific process in Step 5 is as follows: Train the initial classification model using the corrected Seed pool data, build an active learner, and the classification model supports multiple classifiers such as random forest, convolutional neural network, long short-term memory network or TabNet; Active learning query strategies support uncertainty sampling, marginal sampling, or entropy sampling; Simultaneously, a performance monitor is initialized to track the trend of changes in the model's precision, recall, and F1 score on the validation set. When a continuous decline in performance is detected exceeding a set threshold, a policy adjustment or model reinitialization is triggered.

7. The intrusion detection method for CAN bus in vehicle networking based on noise perception active learning according to claim 1, characterized in that: The specific process in Step 6 is as follows: A hybrid query strategy integrating "uncertainty" and "noise perception" is designed for sample selection. First, the uncertainty score is calculated based on the prediction probability of unlabeled samples in the Pool using the current classification model, and the candidate samples with the most ambiguous judgment by the model are selected. Then, an integrated noise detector is used to evaluate the reliability of the pseudo-labels of the candidate samples and calculate the noise probability. The uncertainty score and noise probability are fused to calculate a comprehensive score, and samples with high uncertainty and low noise probability are selected first to ensure that the selected samples can solve the current cognitive ambiguity of the model and have high label reliability. At the same time, the sampling strategy is dynamically adjusted according to the current model performance: when the false negative rate is too high, an attack enhancement sampling strategy is adopted, and boundary attack samples are selected first. When the performance continues to decline, a conservative sampling strategy is adopted. When the overall performance is poor, a balanced backoff sampling strategy is adopted. Class balance constraints are applied during the sampling process to ensure the minimum proportion of attack samples and alleviate the class imbalance problem. When the false negative rate exceeds the set threshold, a false negative-targeted learning mechanism is activated to find samples with similar features to the false negative samples in the Pool and add them to the training batch first.

8. The intrusion detection method for CAN bus in vehicle networking based on noise perception active learning according to claim 1, characterized in that: The specific process in Step 7 is as follows: the high-value samples selected by the hybrid query strategy are added to the existing training set to form an expanded training set. The classification model is retrained using the expanded complete training set so that the model can learn more discriminative feature representations. At the same time, the classification threshold is dynamically adjusted based on the precision-recall curve of the validation set to find the optimal threshold that maximizes the F1 score, thereby reducing the false negative rate while ensuring precision. After training is completed, the selected samples are removed from the pool.

9. The intrusion detection method for CAN bus in vehicle networking based on noise perception active learning according to claim 1, characterized in that: The specific process in Step 8 is as follows: The entire method is driven by a closed-loop iterative cycle of "query-correction-retraining-evaluation". In each round of the cycle, the model's performance metrics such as precision, recall, F1 score, accuracy, false positive rate, and false negative rate are first evaluated on the validation set. The performance monitor continuously tracks the changes in the F1 score. When the performance is detected to have decreased continuously for more than a set number of times, the model is triggered to reinitialize in order to escape the local optimum. The iteration continues until the preset number of iterations is reached or the pool of samples is exhausted, and finally, a binary classification model with the best intrusion detection performance is output.

Images