Lightweight multi-level hybrid intrusion detection method based on differential privacy in internet of vehicles

Abstract

The application discloses a lightweight multi-level hybrid intrusion detection method based on machine learning and differential privacy, adds Laplace random noise in a differential privacy paradigm into an original data set, performs data sampling on the data set, adopts a data regularization algorithm to regularize the data set, uses a feature selection method to select N features from M existing features, uses a feature extraction method to transform group measurement values of a certain mode, uses a single-class support vector machine model to detect data, and judges whether the data is normal data; uses a supervised learning model to detect known attacks; and uses an unsupervised learning model to detect unknown zero-day attacks. The application adopts multiple machine learning models to detect normal traffic, known attacks and unknown zero-day attacks in a vehicle network with high performance, and adopts a differential privacy technology to protect the privacy information of the data. The application can realize multiple detections of attacks in the Internet of Vehicles, and prevent the Internet of Vehicles from being attacked by various network attacks.

Description

Technical Field

[0001] This invention relates to vehicle networking and deep learning technologies, specifically to a lightweight, multi-level hybrid intrusion detection method based on differential privacy in vehicle networking. Background Technology

[0002] With the continuous innovation of in-vehicle technology and the gradual replacement of traditional cars by connected and autonomous vehicles, vehicular ad hoc networks (V2VNs) are also constantly developing and improving. V2VNs support vehicle-to-everything (V2X) communication, which can be categorized based on the nature of the surrounding environment, including vehicle-to-vehicle, vehicle-to-infrastructure, vehicle-to-pedestrian, and vehicle-to-network communication. Currently, V2VNs support numerous applications, primarily focusing on different aspects of traffic organization, such as driver assistance, public safety, toll collection, road traffic control, improving vehicle safety, and providing drivers with convenient and fast infotainment services.

[0003] Vehicular ad hoc networks (VANs) are highly disconnected and mobile networks where nodes possess ample computing power and storage capacity. The high mobility of vehicles, the variability in density, the diversity of communication methods, and frequent network topology changes present numerous problems and challenges in designing and developing VAN architectures. This also makes VANs vulnerable to various network attacks, such as denial-of-service (DoS), distributed denial-of-service (DDoS), and GPS spoofing attacks. These attacks severely threaten the security of VANs, reduce their stability and robustness, and can lead to vehicle malfunctions or traffic accidents.

[0004] Currently, intrusion detection-based solutions are widely used in security zones, primarily for detecting attacks and developing corresponding countermeasures. Traditional static detection technologies, such as firewalls, are passive defenses, subject to various limitations and regional influences, and have certain limitations in handling external attacks. Intrusion detection complements firewalls, proactively defending against cybersecurity threats by analyzing network and system data to determine the presence of intrusions, thus protecting the system and compensating for the shortcomings of traditional defense methods. Therefore, in vehicular ad hoc networks, using intrusion detection-based solutions can protect these networks from various network attacks.

[0005] Currently, machine learning and data mining algorithms are widely considered effective models for designing intrusion detection methods. However, training machine learning models requires large amounts of data that may contain sensitive information. Therefore, the risk of privacy breaches is high during model training, posing a significant challenge to data security and privacy. In the field of intrusion detection methods for vehicular ad hoc networks, research on protecting the privacy of training data is also scarce.

[0006] Furthermore, two problems exist in the field of intrusion detection in vehicular ad hoc networks. First, attackers can execute unknown zero-day attacks not included in known attack patterns. Such attacks may be misclassified as normal, leading to a higher proportion of false positives. Second, researchers often focus on detecting anomalous data while neglecting normal data. For example, patent CN108322445 A discloses a network intrusion detection method based on transfer learning and ensemble learning, using machine learning algorithms to filter network traffic data and construct various network traffic data containing attack behaviors, ignoring the issue of privacy leakage, and also incurring significant computational overhead. For example, patent CN 116155572A discloses an encrypted traffic network intrusion detection method based on ensemble learning, which only detects encrypted data traffic without addressing the zero-day attack problem, similarly ignoring the issue of data privacy leakage. Summary of the Invention

[0007] Purpose of the invention: The purpose of this invention is to address the shortcomings of existing technologies and provide a lightweight, multi-level hybrid intrusion detection method based on differential privacy in the Internet of Vehicles.

[0008] Technical solution: The present invention provides a lightweight multi-level hybrid intrusion detection method based on differential privacy in the Internet of Vehicles, comprising the following steps:

[0009] Step (1): Add the Laplace random noise r from the differential privacy paradigm to the original dataset X to protect the privacy information of the dataset and obtain the differential privacy dataset X. DP X DP =X+r;

[0010] Step (2), due to the differential privacy dataset X DP The dataset X contains a large amount of data, which is relevant to the differential privacy dataset. DP Data sampling is performed to reduce the amount of data and the resource cost of machine learning algorithms, resulting in a differential privacy dataset X. DP_1 ;

[0011] Step (3): Apply the data regularization algorithm - standard score - to the differential privacy dataset X. DP_1 Regularization is applied to improve data comparability but reduce data interpretability, resulting in a differential privacy dataset X. DP_2 ;

[0012] Step (4): For the differential privacy dataset X DP_2 Feature selection is performed using information gain and fast correlation filtering algorithms to choose N features from the existing M features, thereby optimizing specific system metrics and obtaining the differential privacy dataset X. DP_3 ;

[0013] Step (5): For the differential privacy dataset XDP_3 Using principal component analysis (PCA) as a feature extraction method, the group measurements of a certain pattern are transformed to highlight the representative features of the pattern, resulting in a differential privacy dataset X. DP_4 ;

[0014] Step (6): Detect differential privacy dataset X using a single-class support vector machine model. DP_4 This is used to determine whether the data is normal.

[0015] Step (7): If the differential privacy dataset X DP_4 The system contains anomalous data, and then uses a voting ensemble model integrating multiple supervised learning models to detect known attacks. The specific method is as follows:

[0016] Step (7-1), the voting ensemble model is the second-level detection model;

[0017] Step (7-2): The voting ensemble model integrates three supervised learning-based algorithms, namely gradient boosting decision tree, random forest and extra tree, to achieve a higher performance learner.

[0018] Step (7-3): Receive the differential privacy dataset X from the single-class support vector machine model. DP_4 Then, on the differential privacy dataset X DP_4 The detection probability result P' is obtained by performing the detection;

[0019] Step (7-4): If P'>P, where P is a fixed threshold, then the output result y={2,3,...,n}, where 2,3,...,n represent multiple types of known attacks;

[0020] Step (7-5): If P' is less than or equal to P, the voting ensemble model marks the data stream as an uncertain instance and submits it to the stacked ensemble model for the next level of detection.

[0021] Step (8): If the differential privacy dataset X DP_4 The abnormal data in the dataset are all unknown attack data. A stacked ensemble model integrating multiple unsupervised learning models is used to detect unknown zero-day attacks.

[0022] Furthermore, step (3) involves the differential privacy dataset X. DP_1 The specific process of regularization is as follows:

[0023] Machine learning models often perform better on standardized datasets, therefore it is necessary to use standard scoring algorithms on the differential privacy dataset X. DP_1 Perform normalization;

[0024] For the differential privacy dataset X DP_1The data x in the data can be normalized to a mean of 0 and a standard deviation of 1 using the standard score algorithm;

[0025] The implementation method of the standard score algorithm satisfies: Where x n Let x be the normalized eigenvalue, μ be the original eigenvalue, and σ be the mean and standard deviation of the eigenvalue, respectively.

[0026] Furthermore, step (4) applies the differential privacy dataset X DP_2 When performing feature selection, the information gain algorithm is first used for processing, followed by the fast correlation filtering algorithm. The specific method is as follows:

[0027] First, the information gain algorithm is used on the differential privacy dataset X. DP_2 Feature selection is performed to obtain the dataset X′ after feature selection using information gain. DP_2 ;

[0028] Then, the fast correlation filtering algorithm is used to filter X′. DP_2 Feature selection is performed to obtain the differential privacy dataset X. DP_3 .

[0029] Furthermore, the single-class support vector machine model is the first-level detection model; it learns the decision boundary for normal observations and considers some outliers. If an outlier lies within the decision boundary, the model treats it as normal; otherwise, it treats it as an anomaly. The single-class support vector machine model detects differential privacy dataset X. DP_4 The specific process is as follows:

[0030] Single-class support vector machine model for detecting differential privacy dataset X DP_4 To determine if the data is normal, obtain the detection result set y;

[0031] If only y = {1} is present, it directly outputs the data detection result as normal data, or it outputs the differential privacy dataset X. DP_4 Marked as normal dataset;

[0032] If y = {1, -1}, it represents a differential privacy dataset X. DP_4 It contains anomalous data, and then the entire differential privacy dataset X is... DP_4 Submit it to the supervised learning model for processing.

[0033] Furthermore, while the supervised learning model in step (8) cannot combat unknown attacks, it can mark unknown attacks as uncertain instances and submit them to the unsupervised learning model for detection when unknown attacks exist in anomalous data. The stacked ensemble model is a third-level detection model; the stacked ensemble model detects differential privacy dataset X. DP_4The specific process of the unknown zero-day attack in the game is as follows:

[0034] Single-class support vector machines and isolated forests are typical anomaly detection algorithms based on unsupervised learning;

[0035] The stacked ensemble model first learns the original data using two base learners: a single-class support vector machine and an isolation forest. Both base learners output data from the original data. Then, the outputs of these two models are stacked column-wise to form a new (m, p)-dimensional dataset, where m represents the number of samples, i.e., the differential privacy dataset X. DP_4 The row number, p represents the number of base learners (since this invention uses two base learners, p = 2 here);

[0036] The new sample data is then fed into the second-layer model for fitting. The second-layer model is the best performing base learner among the two base learners. According to the experimental results, the isolation forest model performs the best, so the second-layer model is the isolation forest model.

[0037] After receiving uncertain instances from the voting ensemble model, the superimposed ensemble model obtains a detection result set y = {-1}, which is a new type of attack.

[0038] The overlay and integration model labels and signs the new attack and adds it to the database of known attack patterns; this completes the transformation from an unknown attack to a known attack, and the next attack of the same type will be considered a known attack.

[0039] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0040] (1) This invention proposes a novel intrusion detection system, which includes a single-class support vector machine model, a voting ensemble model and a stacked ensemble model, and can detect normal traffic, known attacks and unknown zero-day attacks in vehicle ad hoc networks with high performance.

[0041] (2) Differential privacy technology is used to protect the privacy information in the training data. The Laplace mechanism in differential privacy technology is applied to the training dataset to obtain differential privacy datasets with different privacy budgets. The proposed novel intrusion detection system is used to verify the practicality of these differential privacy datasets.

[0042] In summary, this invention achieves more functional features, such as normal traffic detection, known attack detection, and unknown attack detection. Furthermore, it utilizes the Laplace mechanism from differential privacy to protect the privacy of training data. Attached Figure Description

[0043] Figure 1 This is a system model diagram of the present invention;

[0044] Figure 2 This is a performance graph of the present invention without the inclusion of a differential privacy mechanism;

[0045] Figure 3 Performance graph of adding differential privacy mechanism to this invention;

[0046] Figure 4 This is a diagram showing the system overhead required by the present invention. Detailed Implementation

[0047] The technical solution of the present invention will be described in detail below, but the scope of protection of the present invention is not limited to the embodiments described.

[0048] To address the issues of data privacy leaks and unknown zero-day attacks, this invention protects vehicular ad hoc networks from various network attacks while safeguarding the privacy of training data. A single-class support vector machine model is used to detect whether traffic is normal. Subsequently, multiple machine learning algorithms are used to construct a voting ensemble model and a stacked ensemble model. The voting ensemble model detects known attack types, while the stacked ensemble model detects unknown zero-day attacks. Simultaneously, the Laplace mechanism from differential privacy is used to protect the privacy of the training data. This invention not only achieves very high detection performance for normal traffic, known attacks, and unknown attacks, but also protects the privacy of the training data. Differential privacy technology effectively solves the problem of privacy leakage in training data while reducing the impact on model training. The differential privacy mechanism allows quantization algorithms to provide privacy protection for the sensitive datasets they operate on. By using differential privacy mechanisms and machine learning algorithms, the model is effectively trained while protecting sensitive information.

[0049] like Figure 1 As shown, a lightweight multi-level hybrid intrusion detection method based on differential privacy in the Internet of Vehicles (IoV) is characterized by the following steps:

[0050] Step (1): Add the Laplace random noise r from the differential privacy paradigm to the original dataset X to obtain the differential privacy dataset X. DP X DP =X+r;

[0051] Step (2): For the differential privacy dataset X DP Data sampling is performed to obtain the differential privacy dataset X. DP_1 ;

[0052] Step (3): Apply the data regularization algorithm - standard score - to the differential privacy dataset X. DP_1 Regularization is performed to obtain the differential privacy dataset X. DP_2 ;

[0053] Step (4): For the differential privacy dataset XDP_2 Feature selection is performed using information gain and fast correlation filtering algorithms to choose N features from the existing M features, thereby optimizing specific system metrics and obtaining the differential privacy dataset X. DP_3 ;

[0054] Step (5): For the differential privacy dataset X DP_3 Using principal component analysis (PCA) as a feature extraction method, the group measurements of a certain pattern are transformed to highlight the representative features of the pattern, resulting in a differential privacy dataset X. DP_4 ;

[0055] Step (6): Detect differential privacy dataset X using a single-class support vector machine model. DP_4 This is used to determine whether the data is normal.

[0056] Step (7): If the differential privacy dataset X DP_4 It contains anomalous data, and then uses a voting ensemble model integrating multiple supervised learning models to detect known attacks with high performance; specific method:

[0057] Step (7-1), the voting ensemble model is the second-level detection model;

[0058] Step (7-2): The voting ensemble model integrates three supervised learning-based algorithms, namely gradient boosting decision tree, random forest and extra tree, to achieve a higher performance learner.

[0059] Step (7-3): Receive the differential privacy dataset X from the single-class support vector machine model. DP_4 Then, on the differential privacy dataset X DP_4 The detection probability result P' is obtained by performing the detection;

[0060] Step (7-4): If P'>P, where P is a fixed threshold, then the output result y={2,3,...,n}, where 2,3,...,n represent multiple types of known attacks;

[0061] Step (7-5): If P' is less than or equal to P, the voting ensemble model marks the data stream as an uncertain instance and submits it to the stacked ensemble model for the next level of detection.

[0062] Step (8): If the differential privacy dataset X DP_4 The abnormal data in the dataset are all unknown attack data. A stacked ensemble model integrating multiple unsupervised learning models is used to detect unknown zero-day attacks.

[0063] As can be seen from the above process, this invention includes five stages: data preprocessing, feature engineering, a signature-based intrusion detection system, an anomaly-based intrusion detection system, and results. Data preprocessing includes steps such as data cleaning, data sampling, and normalization. During feature engineering, feature selection and feature extraction are performed on the dataset. Information gain-based and fast correlation-based filter methods are used for feature selection to remove irrelevant and redundant features. Principal component analysis is used for feature extraction, further reducing the dimensionality and noise features of the dataset. The signature-based intrusion detection system has a two-level detection model: a single-class support vector machine model and a voting ensemble model. As the first-level detection model, the single-class support vector machine classifier first detects whether the data is normal. If so, it directly outputs the result and marks the data as normal. Otherwise, it indicates that the data contains anomalous data and then submits it to the next-level model for processing. The voting ensemble model is the second-level detection model. It integrates three supervised learning-based algorithms: gradient-boosted decision trees, random forests, and additional trees, to achieve a higher-performance learner. After receiving a data stream from a single-class support vector machine (SVM), the system detects the data and obtains a detection probability result P'. If P' is greater than a threshold P, the output result is one or more known attack patterns. Conversely, it marks the data as an uncertain instance and submits it to an anomaly-based intrusion detection system for the next level of detection. The anomaly-based intrusion detection system uses a stacked ensemble model for detection. The stacked ensemble model integrates two unsupervised learning-based algorithms, namely a single-class SVM and an isolation forest, to obtain a higher-performance learner. After receiving uncertain instances from the signature-based intrusion detection system, the anomaly-based intrusion detection system detects unknown attacks in the uncertain instances and then adds the unknown attack type to the database of known attack patterns, which will be treated as known attacks in the future.

[0064] To facilitate understanding of the technical solution in this embodiment, the meanings of the relevant variables are explained first, as shown in Table 1.

[0065] Table 1

[0066] X Dataset <![CDATA[X DP ]]> Datasets under differential privacy protection ε Privacy Budget Parameters r Laplace random noise, with values ​​ε = 1, 10, 100, 200 <![CDATA[x n ]]> Normalized eigenvalues μ mean σ Standard deviation y Results obtained from model testing P threshold P′ Detection probability results

[0067] Example:

[0068] This embodiment is constructed using a large number of machine learning algorithms, including but not limited to: Bayesian optimization algorithm for hyperparameter tuning; information gain algorithm and fast correlation filtering algorithm for feature selection; principal component analysis for feature extraction; and single-class support vector machine, voting ensemble, and stacked ensemble for detection models.

[0069] This invention achieves the following performance in processing datasets without adding Laplace noise: Figure 2As shown, the detection performance of this invention is very high, generally above 95%. After adding Laplace noise, the processing performance on privacy-sensitive datasets is as follows: Figure 3 As shown, the performance is very high, proving that differential privacy technology can effectively protect the privacy information of the dataset while having a minimal impact on data detection.

[0070] The technical solution of this invention is also applicable to processing differential privacy datasets X. DP A survey was conducted on the CPU overhead consumed during the process, and the results are as follows: Figure 4 As shown, we found that the CPU resources consumed by this invention only account for a maximum of 6%, which is very lightweight.

[0071] As can be seen from the above embodiments, this invention considers an intrusion detection system for detecting abnormal data. It focuses on dataset processing, primarily using machine learning algorithms to filter network traffic data, thereby constructing various network traffic datasets containing attack behaviors. The use of differential privacy technology prevents privacy information leakage and protects the privacy of training data. Furthermore, the system overhead of the intrusion detection system is also considered.

Claims

1. A lightweight, multi-level hybrid intrusion detection method based on differential privacy in the Internet of Vehicles (IoV), characterized in that, Includes the following steps: Step (1): Add the Laplace random noise r from the differential privacy paradigm to the original dataset X to obtain the differential privacy dataset X. DP X DP =X+r; Step (2): For the differential privacy dataset X DP Data sampling is performed to obtain the differential privacy dataset X. DP_1 ; Step (3): Apply the data regularization algorithm - standard score - to the differential privacy dataset X. DP_1 Regularization is performed to obtain the differential privacy dataset X. DP_2 ; Step (4): For the differential privacy dataset X DP_2 Feature selection is performed using information gain and fast correlation filtering algorithms to choose N features from the existing M features, thereby optimizing specific system metrics and obtaining the differential privacy dataset X. DP_3 ; Step (5): For the differential privacy dataset X DP_3 Using principal component analysis (PCA) as a feature extraction method, the group measurements of a certain pattern are transformed to highlight the representative features of the pattern, resulting in a differential privacy dataset X. DP_4 ; Step (6): Detect differential privacy dataset X using a single-class support vector machine model. DP_4 This is used to determine whether the data is normal. Step (7): If the differential privacy dataset X DP_4 The system contains anomalous data, and then uses a voting ensemble model, which integrates multiple supervised learning models, to detect known attacks. The specific method is as follows: Step (7-1), the voting ensemble model is the second-level detection model; Step (7-2): The voting ensemble model integrates three supervised learning-based algorithms, namely gradient boosting decision tree, random forest and extra tree, to achieve a higher performance learner. Step (7-3): Receive the differential privacy dataset X from the single-class support vector machine model. DP_4 Then, on the differential privacy dataset X DP_4 The detection probability result P' is obtained by performing the detection; Step (7-4): If P'>P, where P is a fixed threshold, then the output result y={2,3,...,n}, where 2,3,...,n represent multiple types of known attacks; Step (7-5): If P' is less than or equal to P, the voting ensemble model marks the data stream as an uncertain instance and submits it to the stacked ensemble model for the next level of detection. Step (8): If the differential privacy dataset X DP_4 The abnormal data in the dataset are all unknown attack data. A stacked ensemble model integrating multiple unsupervised learning models is used to detect unknown zero-day attacks.

2. The lightweight multi-level hybrid intrusion detection method based on differential privacy according to claim 1, characterized in that, Step (3) involves the differential privacy dataset X. DP_1 The specific process of regularization is as follows: For the differential privacy dataset X DP_1 The data x in the data can be normalized to a mean of 0 and a standard deviation of 1 using the standard score algorithm; The implementation method of the standard score algorithm satisfies: Where x n Let x be the normalized eigenvalue, μ be the original eigenvalue, and σ be the mean and standard deviation of the eigenvalue, respectively.

3. The lightweight multi-level hybrid intrusion detection method based on differential privacy according to claim 1, characterized in that, Step (4) involves the differential privacy dataset X. DP_2 When performing feature selection, the information gain algorithm is first used for processing, followed by the fast correlation filtering algorithm. The specific method is as follows: First, the information gain algorithm is used on the differential privacy dataset X. DP_2 Feature selection is performed to obtain the dataset X′ after feature selection using information gain. DP_2 ; Then, the fast correlation filtering algorithm is used to filter X′. DP_2 Feature selection is performed to obtain the differential privacy dataset X. DP_3 .

4. The lightweight multi-level hybrid intrusion detection method based on differential privacy according to claim 1, characterized in that: Step (6) involves using a single-class support vector machine model to detect the differential privacy dataset X. DP_4 The specific process is as follows: Single-class support vector machine model for detecting differential privacy dataset X DP_4 To determine if the data is normal, obtain the detection result set y; If only y = {1} is present, it directly outputs the data detection result as normal data, or it outputs the differential privacy dataset X. DP_4 Marked as normal dataset; If y = {1, -1}, it represents a differential privacy dataset X. DP_4 It contains anomalous data, and then the entire differential privacy dataset X is... DP_4 Submit it to the supervised learning model for processing.

5. The lightweight multi-level hybrid intrusion detection method based on differential privacy according to claim 1, characterized in that, Step (8) uses a stacked ensemble model integrating multiple unsupervised learning models to detect the differential privacy dataset X. DP_4 The specific process of the unknown zero-day attack in the game is as follows: Step (8-1): Supervised learning models cannot fight against unknown attacks, but they can mark unknown attacks as uncertain instances and submit them to unsupervised learning models for detection when unknown attacks exist in abnormal data. Stacked ensemble models are third-level detection models. Step (8-2), single-class support vector machine and isolation forest are typical anomaly detection algorithms based on unsupervised learning; Step (8-3): The stacked ensemble model first learns the original data through two base learners: a single-class support vector machine and an isolation forest. Both base learners output data from the original data. Then, the outputs of the two models are stacked column-wise to form a new (m, p)-dimensional dataset, where m represents the number of samples, i.e., the differential privacy dataset X. DP_4 The number of rows, p represents the number of base learners; Step (8-4): Then, the new sample data is fed into the second-layer model for fitting. The second-layer model is the best performing base learner among the two base learners. According to the experimental results, the isolation forest model has the best performance. Therefore, the second-layer model is the isolation forest model. Step (8-5): After receiving uncertain instances from the voting ensemble model, the superimposed ensemble model obtains the detection result set y = {-1}, which is the new type of attack. Step (8-6): The superimposed integrated model marks and signs the new attack and adds it to the database of known attack patterns; thus completing the transformation from unknown attack to known attack, and the next attack of the same type will be regarded as a known attack.

Images