A fast and efficient dynamic updating method for a secure and reliable model of internet of vehicles

Abstract

The application belongs to the technical field of vehicle networking security management, and particularly relates to a fast and efficient vehicle networking security and trust model dynamic updating method, which comprises the following steps: collecting communication data through an RSU, constructing a trust matrix, dividing the trust matrix into two categories, marking, training a trust evaluation model based on a deep neural network on the RSU as a teacher model, training a relatively simple trust evaluation model on a vehicle as a student model, judging by the vehicle whether the trust evaluation model of the vehicle is adapted to the network environment of the current area, and actively initiating a trust model updating request to the RSU in the area; after receiving the request of the vehicle, the RSU adopts a response-based knowledge distillation method to update the trust evaluation model for the vehicle. The application uses knowledge distillation to transfer knowledge, constantly updates and improves the trust evaluation model of the vehicle, enables the vehicle to have the ability of active adaptation and learning, better meets the demand of a dynamic network environment, and has a wide application prospect.

Description

Technical Field

[0001] This invention belongs to the field of vehicle network security management technology, specifically involving a fast and efficient method for dynamically updating a vehicle network security and trust model. Background Technology

[0002] In recent years, the remarkable performance of the Internet of Vehicles (IoV) in intelligent transportation systems has garnered widespread social attention and recognition. It provides users with safer, more convenient travel and more efficient and intelligent transportation services, not only improving traffic efficiency but also leading transportation services towards a more intelligent development. However, due to the openness and high dynamism of IoV, the risk of vehicles being maliciously attacked is correspondingly increased, potentially leading to problems such as vehicle communication interruptions, the sending of false messages, and data leaks. Therefore, researching the security issues of IoV has become one of the main focuses.

[0003] Trust management is an effective method for preventing cyberattacks and reducing risks by establishing trust mechanisms. However, traditional trust management models mainly use static, centralized encryption technologies to address security and privacy issues. But due to the dynamic changes in the connected vehicle environment, with new behaviors and threats constantly emerging, traditional static trust management models struggle to effectively address this dynamic and uncertain nature. Therefore, how to construct a flexible trust management architecture that can quickly adapt to the dynamic changes in the network environment is a crucial problem that urgently needs to be solved in the field of connected vehicle security.

[0004] As research on knowledge distillation in model compression and lightweight learning matures, its application in the field of trust management can be considered. Model transfer based on this technology can effectively reduce the training time of trust evaluation models while maintaining their accuracy, thereby helping vehicles adapt more quickly to dynamic changes in the network environment. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a fast and efficient method for dynamically updating a vehicle-to-everything (V2X) security and trust model, comprising the following steps:

[0006] S1: By monitoring the communication process between vehicles through the RSU, communication data is collected, used as evidence of trust, and a trust matrix is ​​constructed.

[0007] S2: The collected trust matrix is ​​divided into two categories using a clustering algorithm, and then labeled as trustworthy and malicious, respectively, to construct a training dataset;

[0008] S3: Train a trust evaluation model based on a deep neural network on RSU, keep it up-to-date at all times, and use it as the teacher model in knowledge distillation;

[0009] S4: Train a relatively simple trust evaluation model on the vehicle and use it as a student model in knowledge distillation;

[0010] S5: The vehicle determines that its trust assessment model is not suitable for the current network environment and actively initiates a trust model update request to the RSU in the area.

[0011] S6: After receiving a request from a vehicle, the RSU uses a response-based knowledge distillation method to update the trust assessment model for the vehicle.

[0012] The beneficial effects of this invention are:

[0013] This invention firstly transfers knowledge from the teacher model on the RSU to the student model on the vehicle using knowledge distillation, helping the student model converge faster without needing to train it from scratch, thus reducing training time and saving significant computational resources for the vehicle. Secondly, under the guidance of the teacher model, the vehicle's trust assessment model learns to adapt to new situations, enabling the vehicle to adapt to new network environments more quickly. In summary, this invention uses knowledge distillation to transfer knowledge, continuously updating and improving the vehicle's trust assessment model, giving the vehicle the ability to proactively adapt and learn, better meeting the needs of dynamic network environments, and has broad application prospects. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating a fast and efficient method for dynamically updating a secure and reliable vehicle network model according to the present invention.

[0015] Figure 2 This is a schematic diagram illustrating a scenario of a fast and efficient method for dynamically updating a secure and reliable vehicle network model according to the present invention.

[0016] Figure 3 This is a schematic diagram of the response-based knowledge distillation process of the present invention. Detailed Implementation

[0017] 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] A fast and efficient method for dynamically updating the security and trust model of vehicle-to-everything (V2X) networks, such as... Figure 1 As shown, it includes:

[0019] S1: RSU monitors the communication process between vehicles, collects communication data, uses it as trust evidence, and constructs a trust matrix;

[0020] Vehicle Support Unit (RSU) collects vehicle-to-vehicle communication data through a sliding time window, including instances of packet loss, tampering, and forwarding delays. To obtain the latest communication data, the RSU continuously updates the time window, tracking and recording the latest communication activity. The RSU then converts the communication data collected within this time window into three binary sequences. These three sequences represent the occurrence of packet loss, tampering, and forwarding delays, respectively, using binary encoding where 1 indicates the occurrence of the corresponding event and 0 indicates its absence. These three binary sequences are then integrated to construct a trust matrix, which serves as input to the trust assessment model. The trust matrix is ​​a two-dimensional data structure where rows and columns correspond to communication data within the time window, and the elements of the matrix indicate whether the corresponding communication data experienced packet loss, tampering, or forwarding delay.

[0021] S2: The collected trust matrix is ​​divided into two categories using a clustering algorithm, and then labeled as trustworthy and malicious, respectively, to construct a training dataset;

[0022] As the network continues to operate, interactions between vehicles gradually accumulate rich trust evidence. This trust evidence is further labeled and constructed into a dataset for subsequent trust model training. Specifically, this invention first uses Principal Component Analysis (PCA) to reduce the dimensionality of the trust matrix, obtaining a dimensionality-reduced feature matrix. Next, this invention uses K-means clustering to classify all feature matrices. In this process, the feature matrix is ​​treated as a data point, and its category is determined by the Euclidean distance between the data point and the cluster centroid. In this invention, all feature matrices are divided into two categories, corresponding to the trusted category and the malicious category, respectively. Based on the clustering results, each category is labeled accordingly, where "1" represents the trusted category and "0" represents the malicious category. The training dataset constructed in this way contains data from both trusted and malicious categories, providing a foundation for subsequent trust model training.

[0023] S3: Train a trust evaluation model based on a deep neural network on RSU, keep it up-to-date at all times, and use it as the teacher model in knowledge distillation;

[0024] Because of its abundant computational resources, the RSU can collect and process large-scale data in real time. This invention utilizes the collected global trust evidence to train a deep neural network-based trust evaluation model on the RSU. Global trust evidence provides more comprehensive vehicle communication data, which helps improve the model's accuracy, robustness, and generalization ability. The model has N convolutional layers and one fully connected layer, which can be selected from VGG, LeNet, or AlexNet. This model is used as a teacher model for knowledge distillation, distilling its output knowledge into a student model for it to learn.

[0025] S4: training a relatively simple trust evaluation model on the vehicle, and using it as the student model in knowledge distillation;

[0026] The trust evaluation model trained on the vehicle is similar to the model on the RSU, but has a more simplified structure, with M convolutional layers and one fully connected layer, wherein M<N. Due to the limited computing resources of the vehicle, it cannot process large-scale training data like the RSU, so the trust evidence of the vehicle mainly comes from local trust evidence obtained from interaction with surrounding vehicles. The vehicle model serves as the student model in knowledge distillation, and learns and improves its trust evaluation performance by using the teacher model on the RSU. This student-teacher knowledge distillation based method can effectively improve the trust evaluation performance of the vehicle model, while adapting to the environment of limited computing resources of the vehicle.

[0027] S5: the vehicle determines that its model does not adapt to the network environment of the current area, and actively initiates a trust model update request to the RSU in the area where it is located;

[0028] As shown in Figure 2 , in the scenario of Figure 2 , the vehicle 1 drives into a new RSU coverage area, and the network environment of the new area is different from that of the previous area. This change may be caused by a variety of reasons, including building occlusion, network load fluctuation, electromagnetic interference and other factors, leading to turbulence in the regional network environment. When coping with this kind of network environment turbulence, problems such as local network congestion, network delay and packet loss attack occur, which in turn leads to fluctuations in communication data between vehicles, causing corresponding changes in trust evidence. The changes caused by this turbulence will further affect the current trust evaluation model of the vehicle, which may lead to misjudgment. Specifically, vehicles that were originally considered trustworthy may be mistakenly identified as malicious vehicles, causing the phenomenon that the number of malicious vehicles detected by the system suddenly increases. In addition, those vehicles wrongly determined to be malicious may face the measure of being prohibited from communicating, which undoubtedly causes a waste of network resources. Because in this case, these misjudged vehicles will not be able to continue to effectively communicate and interact with other vehicles in the network, thereby adversely affecting the effectiveness and normal operation of the entire network.

[0029] In view of this, the vehicle will independently determine whether its trust model adapts to the network environment of the area it is located in.

[0030] First, the present invention fuses the binary trust evidence sequences collected by the vehicle into a single evidence sequence, the sequence size is l ω1 . During the fusion process, as long as any one bit value is 1 in any of the original evidence sequences, the bit value at the corresponding position of the fused sequence is also set to 1. Further, a length of l is introduced ω2A sliding window is used to extract data from the fused individual evidence sequences and calculate the attack probability. When the network environment is stable, the calculated attack probability is relatively stable each time; conversely, if the attack probability changes significantly, it indicates that the current network environment may have changed. By setting an appropriate attack probability threshold, the probability value calculated by the vehicle is compared with the threshold. When the probability value exceeds the threshold, the vehicle determines that the current network environment has changed.

[0031] Calculating the probability of an attack includes:

[0032]

[0033] Where, p attack Indicates the probability of attack, l ω2attack Indicates in l ω2 The number of bits with a value of "1" in the time window, l ω2all Indicates in l ω2 The total number of all elements in the time window.

[0034] Attack probability threshold:

[0035]

[0036] Among them, Thr attack p represents the attack probability threshold. atack-now p represents the currently calculated attack probability. atack-previous This represents the attack probability calculated in the previous stage.

[0037] Secondly, in a length of l ω3 Within a sliding time window, the number of malicious results detected by the vehicle trust model is recorded. By setting an appropriate threshold, the number of malicious results in the current window is compared with the previously recorded number. When the number of malicious results in the current window exceeds the set threshold, the vehicle determines that the current trust model's detection result has an error.

[0038] Quantity threshold:

[0039]

[0040] Among them, Thr num n represents the quantity threshold. malicious-now n represents the number of malicious nodes detected so far. malicious-previous This indicates the number of malicious nodes detected in the previous stage.

[0041] Finally, when the vehicle determines that the current network environment has changed and simultaneously identifies errors in the trust model's detection results, it indicates that the current trust model is no longer suitable for the new network environment. In this case, the vehicle proactively initiates a trust assessment model update request to the RSU. This proactive feedback mechanism highlights the vehicle's intelligent decision-making when it perceives dynamic changes in the network and identifies inaccuracies in the trust model. By initiating an update request to the RSU, the vehicle demonstrates its need for adaptive trust management to ensure that the trust model matches changes in the current network environment.

[0042] S6: After receiving the vehicle's request, the RSU uses a response-based knowledge distillation method to update the vehicle's trust assessment model, such as... Figure 3 As shown.

[0043] The trust assessment models for RSUs and vehicles are respectively used as the teacher model and the student model, denoted as . and

[0044] To prevent responding to model update requests from malicious vehicles, the RSU, upon receiving a model update request from a vehicle, considers the evaluation results of its neighboring vehicles. Specifically, during trust management, each vehicle can obtain information about its neighboring vehicles and output a trust evaluation result. The RSU will only respond to a vehicle's model update request if more than 50% of its neighboring vehicles evaluate it as trustworthy.

[0045] Once the RSU decides to respond to the model update request sent by the vehicle, it uses a response-based knowledge distillation method to update the vehicle's trust evaluation model. This knowledge distillation method is an effective knowledge transfer technique that, by... knowledge and experience transfer to Help Quickly adapt to new environments. During the distillation process, The knowledge mainly includes its predictive ability and reasoning process, that is, the ability to make accurate predictions when faced with new input samples. By fitting This knowledge is learned by analyzing the output probability distribution. This method helps improve the accuracy and robustness of vehicle trust assessment models.

[0046] The details are as follows:

[0047] First, the teacher-student model uses a softmax output layer to generate classification probabilities:

[0048]

[0049] Among them, z iz represents the logits belonging to the i-th class of output. j p represents the logits belonging to the j-th class of output; i Let represent the probability of being classified into the i-th class, and T represent the distillation temperature. T is primarily used to adjust the softness or hardness of the output labels; the higher T is, the smoother the output probability distribution of the softmax layer. In response-based knowledge distillation, the logits of the softmax output refer to... or The output consists of raw numerical values ​​that have not been processed by the softmax function. These values ​​represent the model's confidence scores for each possible output class. Logits are used as input to the loss function during training, and then transformed into a probability distribution (soft labels) by the softmax function for calculating the loss.

[0050] Secondly, both teacher and student models are trained using the training dataset. At high temperature T... high Next, and The probability distribution output by the softmax layer is used as the soft label, and then... soft tags as The training objective. In order to... The soft label should be as close as possible to Soft labels use cross-entropy loss as the loss function, denoted as distillation loss L. soft .

[0051] Distillation loss L soft The expression is as follows:

[0052]

[0053] in, express Under temperature T, the softmax layer outputs the probability of the i-th class. express The softmax layer outputs the probability of class i at temperature T. The expression is as follows:

[0054]

[0055]

[0056] Among them, v i express logits, z i express The logits, where N represents the total number of labels.

[0057] Secondly, during the distillation process, the teacher model also has a certain error rate. Hard labels (also known as true labels) can effectively reduce the possibility of the student model learning incorrect knowledge. Hard labels here play a role similar to the standard answer in real-world educational scenarios. It will adjust its predictions based on hard labels, thereby better learning the correct knowledge and skills. Therefore, at T=1, the calculation... The cross-entropy loss between the output soft labels and hard labels is denoted as the student loss L. hard The expression is as follows:

[0058]

[0059] Among them, c i c represents the true label value in the i-th class. i For ∈{0,1}, positive labels are 1 and negative labels are 0. The specific expression is as follows:

[0060]

[0061] Finally, the total loss function of the distillation process is obtained by weighting the distillation loss and the student loss, as shown in the following expression:

[0062] L+αL soft +βL hard

[0063] Where α and β are equilibrium L soft and L hard The weights are given by α+β=1.

[0064] In the process of updating the vehicle trust assessment model, the smaller the loss function L, the more it means... and The more similar the test results are, The more knowledge you learn, the better. When L converges, it indicates... Training is complete, meaning the vehicle trust assessment model has been updated.

[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fast and efficient method for dynamically updating a vehicle-to-everything (V2X) security and trust model, characterized in that, include: S1: By monitoring the communication process between vehicles through the RSU, communication data is collected, used as evidence of trust, and a trust matrix is ​​constructed. By monitoring communication between vehicles through the RSU (Roadside Unit), communication data is collected, used as evidence of trust, and a trust matrix is ​​constructed, including: The RSU collects vehicle-to-vehicle communication data through a sliding time window, including data on packet loss, tampering, and forwarding delays. To obtain the latest communication data, the RSU continuously updates the time window, tracks, and records the latest communication data. The RSU will convert the packet loss, tampering, and forwarding delay data collected within the time window into three binary sequences; wherein the binary sequences are encoded in binary, with 1 indicating that the corresponding situation has occurred and 0 indicating that it has not occurred; These three binary sequences are integrated to construct a trust matrix, which serves as the input to the trust assessment model. The trust matrix is ​​a two-dimensional data structure, where the rows and columns correspond to the communication data within the time window, and the elements of the matrix indicate whether the corresponding communication data has experienced packet loss, tampering, or forwarding delay. S2: The collected trust matrix is ​​divided into two categories using a clustering algorithm, and then labeled as trustworthy and malicious, respectively, to construct a training dataset; S3: Train a trust evaluation model based on a deep neural network on RSU, keep it up-to-date at all times, and use it as the teacher model in knowledge distillation; S4: Train a relatively simple trust evaluation model on the vehicle and use it as a student model in knowledge distillation; S5: The vehicle determines that its trust assessment model is not suitable for the current network environment and actively initiates a trust model update request to the RSU in the area. If a vehicle determines that its model is not suitable for the current network environment, it proactively initiates a trust model update request to the RSU in the area, including: The binary trust evidence sequences collected by the vehicle are fused into a single evidence sequence, with a sequence size of [size missing]. During the fusion process, if any bit in any original evidence sequence is 1, the corresponding bit in the fused sequence will also be set to 1. Introduce a length of A sliding window is used to extract data from a fused single evidence sequence and calculate the attack probability; When the network environment is stable, the calculated attack probability is relatively stable each time; conversely, if the attack probability changes significantly, it indicates that the current network environment may have changed. An attack probability threshold is set, and when the attack probability exceeds the attack probability threshold, the vehicle determines that the current network environment has changed. In a length of Within the sliding time window, the number of malicious results detected by the vehicle trust model is recorded. A threshold is set, and the number of malicious results in the current window is compared with the number recorded previously. When the number of malicious results in the current window exceeds the set threshold, the vehicle judges that the detection result of the trust model in the current stage has an error. When a vehicle determines that the current network environment has changed and at the same time determines that the detection result of the trust model has an error, it indicates that the current trust model is no longer suitable for the new network environment. In this case, the vehicle will actively send a trust assessment model update request to the RSU. S6: After receiving a request from a vehicle, the RSU uses a response-based knowledge distillation method to update the trust assessment model for the vehicle.

2. The fast and efficient method for dynamically updating a vehicle-to-everything (V2X) security and trust model according to claim 1, characterized in that, The collected trust matrix was divided into two categories using a clustering algorithm, and then labeled as trustworthy and malicious, respectively, to construct a training dataset, including: Principal component analysis (PCA) algorithm is used to reduce the dimensionality of the trust matrix, resulting in the dimensionality-reduced feature matrix. The K-means clustering algorithm is used to classify all the dimensionality-reduced feature matrices: the feature matrix is ​​regarded as a data point, and its category is determined by the Euclidean distance between the data point and the cluster centroid; where "1" represents the trustworthy category and "0" represents the malicious category.

3. The fast and efficient method for dynamically updating the vehicle network security and trust model according to claim 1, characterized in that, A trust evaluation model based on a deep neural network is trained on the RSU and kept up-to-date at all times, which is used as the teacher model in knowledge distillation, comprising: The trust evaluation model based on a deep neural network is formed by N convolutional layers and one fully connected layer, The RSU uses the collected global trust evidence as the input of the trust evaluation model, and the output of the trust evaluation model is the trust evaluation classification for vehicles.

4. The fast and efficient method for dynamically updating the vehicle network security and trust model according to claim 1, characterized in that, A relatively simple trust evaluation model is trained on the vehicle and used as the student model in knowledge distillation, comprising: The relatively simple trust evaluation model is formed by M convolutional layers and one fully connected layer, where M<N, M is the number of convolutional layers of the relatively simple trust evaluation model, and N is the number of convolutional layers of the trust evaluation model based on deep neural network. The vehicle uses the collected local trust evidence as the input of the model, and the output of the model is the trust evaluation classification for other vehicles.

5. The fast and efficient method for dynamically updating a vehicle-to-everything (V2X) security and trust model according to claim 1, characterized in that, Calculating the attack probability, comprising: ； in, Indicates the probability of an attack. Indicates in The number of bits with a value of "1" in the time window. Indicates in The total number of all elements in the time window.

6. The fast and efficient method for dynamically updating the vehicle network security and trust model according to claim 1, characterized in that, Setting of the threshold, comprising: Attack probability threshold: ； in, This represents the attack probability threshold. This represents the currently calculated attack probability. This represents the attack probability calculated in the previous stage; Quantity threshold: ； in, Indicates the quantity threshold. This indicates the number of malicious nodes currently detected. This indicates the number of malicious nodes detected in the previous stage.

7. The fast and efficient method for dynamically updating a vehicle-to-everything (V2X) security and trust model according to claim 1, characterized in that, After the RSU receives the request from the vehicle, it adopts a response-based knowledge distillation method to update the trust evaluation model for the vehicle, comprising: The trust assessment models for RSUs and vehicles are respectively used as the teacher model and the student model, denoted as ... and ; After the RSU receives the model update request sent by the vehicle, in the trust management process, each vehicle can obtain the information of its neighboring vehicles and output the trust evaluation result. Only when more than 50% of the neighboring vehicles of a certain vehicle evaluate it as trusted, will the RSU respond to the model update request sent by it; Once the RSU decides to respond to the model update request sent by the vehicle, it uses a response-based knowledge distillation method to update the trust assessment model for the vehicle; during the distillation process, Knowledge includes its predictive ability and reasoning process, that is, the ability to make accurate predictions when faced with new input samples. By fitting We learn this knowledge by analyzing the output probability distribution.

8. The fast and efficient method for dynamically updating the vehicle network security and trust model according to claim 7, characterized in that, After the RSU decides to respond to the model update request sent by the vehicle, it adopts the response-based knowledge distillation method to update the trust evaluation model for the vehicle, comprising: Teacher Model Calculate class probabilities using the softmax output layer. ,in, Indicates the first The probability of a class Indicates belonging to the first The class outputs logits, Indicates belonging to the first The class outputs logits, Indicates the distillation temperature; High temperature Next, the teacher model and student model The probability distribution output by the softmax layer is used as a soft label, and the teacher model is employed. Soft labels as student models The training objective is to build the student model The soft labels should be as close as possible to the teacher model. Distillation loss of soft labels ,in, Teacher model Under temperature T, the softmax layer outputs the probability of the i-th class. Student model At temperature T, the softmax layer outputs the probability of the i-th class. Indicates the total number of categories; Using real labels as hard labels, student model Adjust your predictions based on hard labels. At that time, calculate the student model The cross-entropy loss between the output soft labels and hard labels is used as the student loss. ,in, This represents the true label value in the i-th class. Positive labels are assigned a value of 1, and negative labels are assigned a value of 0. The category label with the highest probability is considered a positive label, while other category labels are considered negative labels. This represents the probability of the student model outputting from the softmax layer at temperature T=1; Weighted combined distillation loss and student losses The total loss function of the distillation process is obtained. During the update process of the vehicle trust assessment model, the loss function The smaller, the better With teacher-student model The more similar the test results are, The more knowledge you acquire, the better. When convergence occurs, it indicates that Training is complete, meaning the vehicle trust assessment model update is complete; among which, and They represent balance respectively. and The weights, and .

Images