Transfer learning-combined security-enhanced Internet of Vehicles trust evaluation method

Abstract

The invention discloses a safety-enhanced Internet of Vehicles trust evaluation method in combination with transfer learning, and relates to the field of Internet of Vehicles safety, and the method comprises the steps: constructing an urban road Internet of Vehicles model based on the urban road condition; the method comprises the following steps: designing an interactive authentication module based on an elliptic curve Diffie-Hellman protocol and an elliptic curve digital signature algorithm on the basis of an urban road vehicle networking model; based on an interaction authentication module, introducing a true and false message identification method of a three-factor trust model improved based on a transfer learning method, and designing a transfer learning model TransMegaCRN in speed expectation sub-factor analysis to obtain the prediction speed of a message sender; based on a transfer learning model TransferMegaCRN, designing a vehicle trust value calculation method by using an improved Bayesian method and a density-based clustering screening algorithm, and constructing an Internet of Vehicles trust management model; and vehicle trust evaluation is completed based on the Internet of Vehicles trust management model.

Description

Technical Field

[0001] The present invention relates to the field of vehicle networking security, and particularly to a security-enhanced vehicle networking trust evaluation method combined with transfer learning. Background Art

[0002] In recent years, the rapid development of Vehicular Ad Hoc Networks (VANET) has brought many opportunities to intelligent transportation systems, improving road safety to a certain extent and optimizing traffic flow management. In a VANET environment, there is a large amount of communication and data exchange between vehicles. Establishing a reliable trust mechanism can ensure the security of communication and the integrity of data. However, Vehicular Ad Hoc Networks are vulnerable to malicious attacks, data tampering, and information leakage threats, and traditional security protection mechanisms are difficult to effectively cope with. Moreover, due to the characteristics of Vehicular Ad Hoc Networks such as high-speed vehicle movement, complex communication environments, and fast dynamic changes in network topologies, the difficulty of trust management is increased, which makes trust management between vehicles still a key issue to be solved urgently. Therefore, it is crucial to build an efficient and accurate vehicle networking trust management mechanism, which can guarantee the security and stability of the system, help vehicles effectively identify and suppress malicious behaviors in complex network environments, and thus improve the overall security and credibility.

[0003] To ensure the security of the vehicle networking system, the trust management system can use technical means such as encryption verification, information digest, and timestamp to ensure the integrity and accuracy of data, thereby improving the credibility of communication between vehicles. In addition, the trust management mechanism can also help managers evaluate the reputation of each vehicle. By continuously tracking and analyzing the behavior records, service quality, and historical performance of vehicles, managers can obtain data-driven evaluation results, providing a basis for implementing targeted reward and punishment measures. This can not only suppress malicious behaviors but also encourage vehicles to provide high-quality services, further optimizing the overall operation efficiency and stability of the vehicle networking system. Summary of the Invention

[0004] To solve the technical problems in the above background, the present invention proposes a security-enhanced vehicle networking trust management model TSE-TMM combined with transfer learning. The present invention examines the trust management architecture in urban road scenarios. Vehicles drive on the road, collect and send information. To identify potential attackers and ensure road safety, a clear metric is needed to distinguish normal vehicles from malicious vehicles. Due to the limited capabilities of vehicles, roadside units and control centers are needed to identify attackers on the road. The vehicle networking trust management model designed by the present invention uses an improved Bayesian method and a density-based clustering screening algorithm to calculate vehicle trust values to quantify the degree of vehicle trust, thereby distinguishing potential attackers and achieving the purpose of vehicle networking system security; and an interactive authentication module based on the Elliptic Curve Diffie-Hellman protocol and Elliptic Curve Digital Signature Algorithm is incorporated into the framework to achieve anonymous communication between vehicles, traceability and identity authentication of vehicles, and an improved trust model of three factors of expectation, risk, and confidence is introduced to identify the authenticity of messages transmitted between vehicles. In the analysis of the expectation factor, a transfer learning model TranferMegaCRN is designed to obtain the predicted speed of the message sender to determine the speed expectation sub-factor, making the framework more secure and reliable.

[0005] To achieve the above object, the present invention provides a security-enhanced vehicle networking trust evaluation method combined with transfer learning. The steps include:

[0006] Based on the urban road conditions, construct an urban road vehicle networking model;

[0007] Based on the urban road vehicle networking model, design an interactive authentication module based on the Elliptic Curve Diffie-Hellman protocol and Elliptic Curve Digital Signature Algorithm;

[0008] Based on the interactive authentication module, introduce a method for identifying true and false messages based on a three-factor trust model improved by transfer learning, and design a transfer learning model TranferMegaCRN in the analysis of the speed expectation sub-factor to obtain the predicted speed of the message sender;

[0009] Based on the transfer learning model TranferMegaCRN, use an improved Bayesian method and a density-based clustering screening algorithm to design a vehicle trust value calculation method, and construct a vehicle networking trust management model;

[0010] Based on the vehicle networking trust management model, complete vehicle trust evaluation.

[0011] Preferably, the constructed urban road vehicle networking model includes four entities: vehicles equipped with on-board units OBU, roadside units RSU, trusted control centers TA, and road speed monitors.

[0012] Preferably, the steps of designing the vehicle trust value calculation method include: in the direct trust value calculation, designing a sliding window method with a time decay factor and a risk adjustment mechanism based on variance to optimize the calculation method based on Bayes' formula; in the indirect trust calculation, designing a density-based clustering screening algorithm to extract trust feature vectors by RSU and eliminate potential malicious recommendation nodes within high-density clusters.

[0013] Preferably, the method for constructing the vehicle network trust management model includes:

[0014] Based on the vehicle trust value calculation method, constructing a management model framework;

[0015] Introducing an interactive authentication module based on the Elliptic Curve Diffie-Hellman protocol and Elliptic Curve Digital Signature Algorithm into the management model framework;

[0016] Introducing a three-factor message recognition method improved by transfer learning method into the management model framework to complete the construction.

[0017] Preferably, the steps of introducing the interactive authentication module based on the Elliptic Curve Diffie-Hellman protocol and Elliptic Curve Digital Signature Algorithm include: when the vehicle registers, negotiating a shared key with the control center TA through the Elliptic Curve Diffie-Hellman protocol, encrypting the real ID to achieve anonymous communication, and ensuring that TA can trace the identity; during the communication process, using the Elliptic Curve Digital Signature Algorithm to sign and verify all messages, and if the receiver's verification fails, marking the sender as a malicious node.

[0018] Preferably, combining the transfer learning method with a time series model based on the Meta Graph Convolutional Recurrent Network MegaCRN to construct a transfer learning model TransferMegaCRN to predict the speed of the message sender and obtain the speed expectation sub-factor.

[0019] Preferably, the transfer learning steps based on TransferMegaCRN include: in the input stage, the original speed data x and the time feature covariate y1 are respectively passed through their respective fully connected layers and then input to the MegaCRN model; in the MegaCRN model, freeze the parameters of all encoder modules in MegaCRN pre-trained by the traffic flow dataset METR-LA; adopt a modular progressive thawing strategy to thaw the decoder and memory modules. During the thawing process, gradually thaw the deep layers of the encoder starting from the 5th round, continue to thaw the intermediate layers after the 10th round, and make a progressive adjustment from fixing the bottom layer to dynamically releasing the upper layer; in the output stage, the output of the MegaCRN model is further processed through a fully connected layer.

[0020] Preferably, based on the TransferMegaCRN speed prediction model, the improved true and false message recognition method with three factors of expectation, risk, and confidence includes: the receiving vehicle calculates the expected value, risk value, and confidence value determined by the speed and position of the sending vehicle to comprehensively evaluate the message credibility. When the dynamic balance condition is met, the message is determined to be true; otherwise, it is marked as untrustworthy.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] The secure enhanced vehicle network trust management model combining transfer learning of the present invention, on the basis of optimizing the vehicle trust calculation model, integrates the elliptic curve cryptography methods (ECDH and ECDSA) and the improved three-factor message recognition method. The improved Bayesian method is used to calculate the direct trust value between vehicles. The real-time interaction data is recorded based on a sliding window and the historical weight is optimized through a time decay factor. A variance adjustment mechanism is designed to reduce the trust score in high-uncertainty scenarios. A density-based clustering screening algorithm is designed. The RSU extracts the trust feature vector and eliminates the potential malicious recommendation nodes within the high-density cluster. Finally, the purified indirect trust and direct trust are weighted and aggregated to generate the local trust value. The ECDH is used to establish a secure key between the vehicle and the control center to achieve anonymous privacy communication between vehicles and identity tracking of vehicles by the control center. The ECDSA is used to achieve identity authentication and message integrity protection for vehicle-to-vehicle communication. In the improved three-factor message recognition method, the TransferMegaCRN spatio-temporal prediction model optimized by transfer learning is innovatively introduced to replace the traditional speed expectation calculation, and a modular progressive thawing strategy is adopted to improve the accuracy of message authenticity recognition in complex traffic scenarios. The security of the vehicle network system is realized. Description of the Drawings

[0023] In order to more clearly illustrate the technical solutions of the present invention, the drawings required for use in the embodiments are briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0024] Figure 1 Schematic diagram of the vehicle network trust model for urban roads in the embodiment of the present invention;

[0025] Figure 2 TransferMegaCRN model diagram of the predicted speed in the embodiment of the present invention;

[0026] Figure 3 Frame diagram of the secure enhanced vehicle network trust management model TSE-TMM combining transfer learning in the embodiment of the present invention. Detailed Embodiments

[0027] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0028] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0029] Before the description, first, the terms appearing in the present invention will be explained.

[0030] (1) Vehicle: The vehicle is equipped with an on-vehicle unit OBU, has a certain computing ability, can also sense the environment and collect data, and helps the vehicle to interact.

[0031] (2) Roadside unit RSU: The RSU is a device with base station functions deployed on both sides of the road, provides communication support, is used to assist in computing, and helps information dissemination as an intermediary for interaction.

[0032] (3) Control center TA: It mainly provides computing support, can execute a large number of computing tasks and help the vehicle or RSU make further decisions.

[0033] Embodiment:

[0034] This embodiment provides a security-enhanced vehicle networking trust evaluation method combined with transfer learning. The steps include:

[0035] S1. Based on the urban road conditions, construct an urban road vehicle networking model.

[0036] As Figure 1 shown, there are mainly four entities in the urban road vehicle networking model constructed in this embodiment, including: vehicles equipped with on-vehicle units OBU, roadside units RSU, trusted control centers TA, and road speed monitors.

[0037] (1) Vehicle movement model:

[0038] To demonstrate the complexity of urban roads, in the road movement model, the Manhattan Mobility Model is adopted. The speed of vehicles is restricted by the speed of the vehicle in front, reflecting the continuity of traffic flow. And the nodes must follow the road direction and cannot move freely away from the road. Under the general Manhattan Mobility Model, in a normal environment, the moving nodes, that is, vehicles, will move along the grid of horizontal and vertical streets on the map. When a vehicle moves until it reaches the intersection of the next cross street and vertical street, the moving node can turn left, turn right, or go straight with a certain probability. The default turning probabilities in the Manhattan model are: the probability of going straight is 0.5, and the probabilities of turning left or right are each 0.25. Whenever a vehicle reaches an intersection, a road choice will be made. Therefore, even if all vehicles start from the same starting point, after a certain period of time, the vehicles will be spread all over the place, similar to the complex urban environment in reality. However, generally, to prevent vehicle collisions, the road is at least a two-way lane to ensure that vehicles can shuttle back and forth on the road. This conforms to the urban road setting.

[0039] (2) Communication loss model:

[0040] Due to the large number of high-rise buildings in the city, the communication loss needs to be considered. According to the characteristics of wireless signal propagation in the vehicle-to-everything (V2X) network, signal fading mainly includes path loss and shadow fading. The path loss can be obtained by calculating the free space path loss:

[0041] ,

[0042] where α2 is the path loss exponent in the free space model; is the reference distance, usually defaulted to 1 (unit: meter); is the straight-line distance between the transmitter and the receiver (unit: meter). The shadow fading can be obtained by calculating the penetration loss of the building:

[0043] ,

[0044] where, represents the number of walls penetrated in the signal path. Each time a wall is penetrated, it is counted as 1 "cut". Therefore, the value of directly corresponds to the total number of walls penetrated by the signal. is the fixed loss for penetrating one wall; is the penetration distance of the signal inside the building (unit: meter), that is, the propagation length of the signal inside the building; is the distance-related loss, that is, how much loss per meter. Therefore, based on the above path loss formula and shadow fading formula, the total signal fading model is:

[0045] ,

[0046] During the vehicle's journey, whether it can receive messages sent by other entities needs to be judged based on the vehicle's power and noise. First, it is necessary to calculate the vehicle's received power and signal-to-noise ratio SNR:

[0047] ,

[0048] ,

[0049] Among them, is the on-vehicle transmission power, and are the antenna gains for transmission and reception respectively, is the noise power. When both the received power of the vehicle is greater than the received sensitivity threshold and the signal-to-noise ratio is greater than the demodulation threshold , the vehicle can correctly receive the message, that is, it satisfies:

[0050] ,

[0051] .

[0052] S2. Based on the urban road vehicle networking model, design an interactive authentication module based on the Elliptic Curve Diffie-Hellman protocol and the Elliptic Curve Digital Signature Algorithm.

[0053] In the urban road model of this embodiment, vehicles achieve V2V / V2I communication by periodically broadcasting beacon messages, and the message content includes core data such as vehicle ID, real-time location, driving speed, and communication data. However, this broadcast-based open communication architecture faces various security threats. Therefore, cryptographic techniques are introduced. Cryptographic techniques are used to achieve data integrity, authenticity, confidentiality, and anonymity, so as to effectively prevent information from being modified without authorization and illegally accessed, thereby ensuring the security and trustworthiness of the communication process. In terms of the selection of cryptographic techniques, in the vehicle networking scenario with limited resources and high real-time requirements, choosing Elliptic Curve Cryptography (ECC) has obvious advantages compared with traditional public key cryptosystems (such as asymmetric cryptography RSA, etc.). ECC can use shorter keys while providing the same security strength, thereby reducing computational overhead and storage requirements, which is particularly important for high-speed moving vehicles and large-scale network environments.

[0054] The ECDH key negotiation mechanism is adopted to achieve secure and efficient key exchange between the vehicle and the control center, ensuring the confidential transmission of sensitive communication data. In this embodiment, during the vehicle's registration phase with the TA, the vehicle Negotiate and share a secret key with the control center TA through the ECDH protocol to protect sensitive information such as the vehicle's true identity. First, TA will publicly disclose the elliptic curve public base point in advance and its own public key , generate a private random number and calculate the relevant public key and submit it to TA. Then, the vehicle based on its secret , TA based on its private key respectively perform calculations to obtain the same shared key . The vehicle uses this shared key to encrypt its true ID and uses ECDSA to sign the message, and then sends it to the receiving vehicle . Since other vehicles do not participate in the above key negotiation, they cannot obtain this shared key , so they cannot decrypt and identify 's true ID and can only obtain the encrypted anonymous identifier; while TA, as the key holder, can use to decrypt and trace 's identity, achieving a balance between anonymity and regulatory traceability in V2V communication.

[0055] Adopt the ECDSA identity authentication mechanism to verify and confirm the identities of the vehicles communicating, ensuring the authenticity of the source of communication data. In this embodiment, when the vehicle sends a message m to the vehicle , the signed message is sent using the ECDSA signature scheme: , , is the elliptic curve public base point, is a security parameter randomly selected by the vehicle ; , is the hash value of the message m, is a secure hash function, is the signature private key of the vehicle (corresponding public key ). When receiving the signed message , the following verification is performed: calculate , , , ; then calculate and determine whether the equation holds. If it holds, the signature is valid.

[0056] S3. Based on the interactive authentication module, introduce a method for identifying true and false messages based on a three-factor trust model improved by transfer learning. In the analysis of the speed expectation sub-factor, design a transfer learning model TranferMegaCRN to obtain the predicted speed of the message sender.

[0057] Based on the three-factor trust model, in this embodiment, the transfer learning method is combined with the meta-graph convolutional recurrent network time series model MegaCRN to design a transfer learning model TransferMegaCRN for vehicle speed prediction, so as to better analyze the influence of the speed expectation sub-factor in the three factors.

[0058] Assume that the V2V message result follows a binomial distribution. Specifically, the interaction between each pair of vehicles is regarded as an independent trial, that is, there are two possibilities for the V2V message of each vehicle: true message or false message. The purpose of three-factor message recognition is to identify all V2V messages and distinguish the messages into trustworthy messages and untrustworthy messages.

[0059] The main implementation steps are as follows:

[0060] S301. Expected value evaluation: The receiver calculates the expected value according to the position information and speed information in the received beacon message. The position expectation is denoted as , the position distance between the receiver and the sender is denoted as , the V2V communication range is denoted as , then the position expectation is set to:

[0061] ,

[0062] The speed expectation is denoted as , which is calculated by the absolute value of the difference between the sender speed value and the receiver's predicted sender speed value :

[0063] ,

[0064] where , , is the maximum speed, and the parameter α1 is .

[0065] In this embodiment, a transfer learning model TransferMegaCRN is constructed based on MegaCRN for the predicted speed Make predictions. The network architecture of MegaCRN mainly consists of two core parts: the graph convolutional recurrent unit (GCRU) encoder-decoder and the spatio-temporal meta-graph learner. The GCRU encoder-decoder extracts the spatial dependencies (such as road network topology) and temporal dynamics (such as traffic flow changes) of spatio-temporal data during the encoding stage by fusing the graph convolutional network (GCN) and the gated recurrent unit (GRU), and generates multi-step prediction results during the decoding stage. The spatio-temporal meta-graph learner realizes adaptive graph structure learning through the meta-node library (storing learnable traffic pattern prototypes), the query-attention mechanism (dynamically matching the current state with prototypes to reconstruct node features), and the hyper-network (generating dynamic graph structures), thereby explicitly distinguishing the heterogeneity of different spatio-temporal scenarios (such as peak hours, traffic accidents). The two parts work together to improve the adaptability and prediction accuracy of the model to complex traffic environments. TransferMegaCRN adds multiple fully connected layers to the entire transfer learning model, mainly in the input and output stages, as Figure 2 shown. First, in the input stage, the original speed data x and the time feature covariate y1 pass through their respective fully connected layers for feature adjustment or preprocessing. Second, in the output stage, the output of TransferMegaCRN will first pass through a fully connected layer for further processing to adapt to specific tasks or adjust the output dimension. In addition, it should be noted that the pre-trained MegaCRN parameters are frozen to prevent the destruction of useful features already learned during the training of new tasks, while reducing the number of training parameters and the risk of overfitting. The parameter freezing training scheme is the main step of transfer learning, and can be dynamically frozen and adjusted according to the requirements of the model task in practical applications. In the transfer learning model TransferMegaCRN used in this embodiment, a modular progressive unfreezing strategy is adopted. First, freeze the parameters of all encoder modules in MegaCRN pre-trained by the traffic flow dataset METR-LA (to retain the general spatio-temporal features extracted by it), and then unfreeze the decoder and memory modules to adapt to new tasks. To ensure progressive unfreezing, during the training process, gradually unfreeze the deep layers of the encoder starting from the 5th round, continue to unfreeze the intermediate layers after the 10th round, and gradually adjust from fixing the bottom layer to dynamically releasing the high layer. This strategy can maintain the core feature extraction ability of the pre-trained model and can also achieve efficient adaptation to new tasks through staged unfreezing.

[0066] According to the evaluation of the expected value, the final expected value takes the minimum of the position expectation and the speed expectation: .

[0067] S302. Risk value assessment: This value represents the possibility of the receiver making a wrong decision and the consequences brought by the wrong decision. The risk value adopts the definition given by the National Institute of Standards and Technology (NIST) of the United States . Among them, The calculation is based on the included angle between the velocity directions of the sender and the receiver ,

[0068] 。

[0069] included angle can be calculated as follows:

[0070] ,

[0071] wherein, 、 and 、 are the velocity components of the sender and the receiver on the horizontal axis and the vertical axis, and and 。 。

[0072] The calculation of

[0073] 。

[0074] S303. Confidence value evaluation: Consider the influence of the social roles of vehicles (such as police cars, ambulances, etc.) on the message credibility. Messages broadcast by vehicles with high role authority are usually considered more credible, so a higher confidence level is given to them. Since this embodiment mainly identifies malicious vehicles on the road, all vehicle roles are treated equally and are regarded as ordinary road vehicles, so the confidence values of all vehicles are the same, all being 0.5, that is 。

[0075] According to the three-factor trust model, use the expected value , risk value , confidence value relationship to deduce a trust domain. When this condition is met, the message is determined to be credible, otherwise it is not credible.

[0076] S4. Based on the transfer learning model TranferMegaCRN, use the improved Bayesian method and the density-based clustering screening algorithm to design a vehicle trust value calculation method and construct a vehicle networking trust management model.

[0077] S401. Use the improved Bayesian method and the density-based clustering screening algorithm to design a vehicle trust value calculation method.

[0078] The messages sent each time will be used for the evaluation of the vehicle trust value after vehicle identity authentication and true / false message identification. The trust value of the vehicle includes two parts: local trust and global trust. The local trust consists of direct trust and indirect trust.

[0079] S402. Use the improved Bayesian method to calculate the direct trust of the vehicle. Based on the method of calculating local trust by the Bayesian method, first use the sliding window method and time decay factor to reduce the amount of data processed by the vehicle, and then use the normalized variance to design a non-linear risk adjustment method to make the trust value decrease rapidly when the uncertainty is high.

[0080] In this embodiment, it is assumed that vehicle interacts with vehicle (that is, vehicle provides message services to vehicle ). Preset in advance that the number of true messages sent is , receive messages from . If the proportion of proven true messages is , then this experiment is recorded as , and its likelihood function is as follows:

[0081] .

[0082] Since this experiment is similar to a coin-tossing experiment, to describe this situation, the distribution can be used as the prior distribution of , and the prior probability density is as follows:

[0083] ,

[0084] Before the interaction starts, that is, when the vehicle just enters the environment, use a uniform prior.

[0085] According to Bayes' theorem, the posterior distribution is determined by the product of the prior distribution and the likelihood function: . Thus, a new distribution is obtained, with parameters , . At this time, according to the distribution mean formula and variance formula , the posterior mean and variance of the new distribution can be obtained.. To distinguish from the calculations obtained by the following improved method, the aforementioned posterior mean and variance are respectively denoted as and .

[0086] To prevent the gradual increase in computational overhead, which may lead to exceeding the tolerance of the on-vehicle unit OBU and thus affecting the real-time performance of vehicle-to-vehicle trust assessment. In this embodiment, a sliding window and a time decay factor are used to reduce the amount of data processed by the vehicle and regulate the calculation of the trust value. The window size length is set to , then there is a queue:

[0087] ,

[0088] ,

[0089] wherein, the queues and respectively record the total interaction data and the real message data within the windows of the two vehicles. Indicates the existence of the th interaction, indicates that the th interaction is a real message. Each time a new interaction occurs, data is added to the end of the queue, and the old data that exceeds the window length is removed. That is, when a new interaction occurs for a vehicle, the interaction information stored earliest in the window is removed, new interaction information is introduced, and at the same time, the two queues are updated, and the new queues are used to perform Bayesian calculations.

[0090] Therefore, the new direct trust value can be obtained through the following calculation:

[0091] ,

[0092] ,

[0093] ,

[0094] ,

[0095] wherein, and are respectively the total number of interactions and the number of real interactions within the window; and are prior parameters, initially 1; is the time decay factor, used to control the weight of the historical trust value, initially set to 0.7, and more emphasis is placed on historical interactions.

[0096] This embodiment also uses variance to further adjust direct trust. Variance represents the uncertainty of the trust value. A larger variance indicates that there are significant fluctuations in the calculation of the trust value of the vehicle for the vehicle when calculating the trust value, and there may be an attack behavior. This embodiment combines the trust value with its variance to obtain a risk-adjusted trust score and reduce the trust assessment in the case of high uncertainty. First, the variance is normalized to establish a unified evaluation benchmark. Since the variance of the distribution is , and its maximum value is 0.25, a normalization operation is performed. Then, the normalized variance is non-linearly amplified to enhance the discrimination of small variances. The non-linear amplification function is , where is the variance amplification coefficient. Then, the exponential decay method is used to punish the clearly highly fluctuating trust to obtain the final direct trust value

[0097] ,

[0098] For the convenience of subsequent introduction, the direct trust will be denoted as subsequently.

[0099] S403. Design a density-based clustering screening algorithm to calculate indirect trust.

[0100] The source of indirect trust is neighbor vehicles. For example, there are three vehicles , , . If there are two direct trusts and , then can refer to 's view on . Therefore, this embodiment establishes a triple set, and for each pair of vehicle nodes , all possible recommenders are collected, and the corresponding triple is as . In addition, this triple needs to satisfy , , , . The purpose is to ensure that the neighbor does not include both the source vehicle and the destination vehicle ; there is communication between the source vehicle and the neighbor , and there is communication between the neighbor and the destination vehicle There is communication. However, not all nodes are reliable. If a recommending node is malicious, it may deliberately smear the behavior of the target node; while if the recommending node has a good cooperation with the target node, it may give the target node a higher score. That is to say, the evaluation of the recommending node may be affected by its relationship with the target node, so these evaluations cannot be completely relied on.

[0101] In this embodiment, in order to prevent the subjective awareness of the vehicle from affecting the indirect trust, the vehicle needs to send the direct trust to the RSU through V2I communication for processing, and a density-based clustering screening scheme is proposed to screen and eliminate a part of potential malicious recommenders.

[0102] First, extract all triples of a pair of vehicles and , and extract the feature vectors . Subsequently, randomly select vectors as the initial centroids of the clustering, and ensure that each initial centroid corresponds to a different data point to avoid duplication. For each vector, i.e., point , calculate its Euclidean distance to each centroid , and assign the data point to the cluster corresponding to the centroid with the closest distance . Then, the centroid of each cluster is updated according to the newly divided clusters. For each cluster m1, calculate the mean of all member feature vectors: , where , , is the set of data points of the th cluster, is the number of members of cluster m1, and then obtain the new clustering centroid . If a cluster has no members (such as due to unreasonable initialization), or the centroid changes, randomly select a data point as the new centroid again. After all centroids are no longer updated or the number of iterations reaches the maximum value, the clustering process ends.

[0103] Since the triples constructed during clustering are based on direct trust, the values in all triples will be affected by the direct trust. An attacker may tamper with the direct trust value between itself and a normal vehicle, resulting in an extreme trust value of 0. Often, a high-density cluster means that the features among members are highly similar, and colluding attackers will produce a dense and similar feature distribution in order to forge trust values. Therefore, the extreme value distribution may cause the feature vectors of some good vehicles and attackers to be misclassified into the same cluster. So in this embodiment, the cluster with the highest density will be selected for elimination (mark this cluster number as ), and the recommenders in other clusters will be retained. The density of the cluster is the reciprocal of the average Euclidean distance between members:

[0104] ,

[0105] Among them, is the set of data points of the th cluster; is the number of members of the th cluster. Mark the recommenders in the remaining clusters as valid for subsequent calculations. Therefore, for each pair of vehicles and , the set of valid recommenders is: . In the set of valid recommenders, different recommendation weights can be designed for each recommender and the source vehicle , that is, the trust weight of vehicle for the recommender . Quantify based on the number of historical interactions, and calculate the trust weight of vehicle for the recommender :

[0106] ,

[0107] Among them, is the number of true messages received by vehicle from vehicle ; is the number of true messages received by vehicle from all recommenders, that is, . Then, weighted aggregation is used to obtain the indirect trust of the source vehicle in the target vehicle :

[0108] .

[0109] S403. Calculate the local trust by weighted aggregating the direct trust and the indirect trust.

[0110] In this embodiment, after calculating the indirect trust, the RSU combines the direct trust and the recommendation trust to obtain the local trust. Specifically, the local trust of the source vehicle in the target vehicle is:

[0111]

[0112] Among them, is the weight factor. The weight can be dynamically adjusted according to the direct trust confidence level. The more historical interaction times, the higher. The local trust obtained by this weighted aggregation is beneficial to improving the accuracy of trust evaluation.

[0113] S404. Calculate the global trust using the VehicleRank algorithm based on PageRank.

[0114] Although the local trust can reflect the trust relationship between every two vehicles, in a complex and large-scale urban vehicle networking environment, relying solely on this point-to-point trust is difficult to meet the interconnection requirements in vehicle networking. Therefore, it is necessary to introduce global trust to uniformly manage and coordinate each trust entity, achieve security event response within the entire network, and thus ensure the comprehensive security and reliability of the entire vehicle networking system. So, after the RSU obtains the aggregated local trust, the RSU will regularly send the local trust to the TA. The TA uses the VehicleRank method to calculate the global trust and publishes the global trust. The RSU or the vehicle can make decisions autonomously based on the global trust. The specific steps of this method include:

[0115] Normalize all the local trust LTVs into probability weights to construct a trust matrix , whose element is defined as:

[0116] ,

[0117] where is the set of all vehicles that have provided information to vehicle ;

[0118] If vehicle fails to give trust to other vehicles because no other vehicle provides it with messages, then vehicle is considered a hanging node. To avoid the interruption of trust distribution caused by the hanging node vehicle , use the global trust vector in the previous round to correct the trust matrix (if the iterative calculation has not started yet, use the initially evenly distributed global trust to correct the trust matrix. For example, in an environment with 200 vehicles, initialize the global trust of each vehicle to 0.005 to obtain the initial global trust vector .), and form a new matrix:

[0119] ,

[0120] where the superscript T represents transpose, a 0 is the hanging vehicle indication vector, and when vehicle is a hanging vehicle = 1, otherwise = 0.

[0121] To ensure the convergence of the global trust calculation, introduce a welfare mechanism to transform the matrix S into a non-negative, irreducible, and aperiodic matrix W:

[0122] ,

[0123] wherein, is a scaling parameter to enhance the welfare impact and control the calculation weight, which is set to 0.45 in this embodiment; is a column vector of all 1s, indicating that the welfare is evenly distributed to all vehicles.

[0124] Based on the above matrix W, the global trust value (GTV) is iteratively updated using the power method. The global trust value (GTV) is initialized through a uniform distribution, that is , and the following iteration is performed:

[0125] ,

[0126] until the difference between adjacent iteration results is less than the threshold at which point it is considered to have converged.

[0127] Finally, the TA broadcasts the global trust value of the vehicle.

[0128] S5. Use the vehicle networking trust management model to complete the vehicle trust evaluation.

[0129] In this embodiment, based on the foregoing interactive authentication module, true / false message recognition method, and trust value calculation method, a framework of a security-enhanced vehicle networking trust management model combining transfer learning is constructed, as Figure 3 shown. Using this framework, the vehicle trust evaluation can be completed.

[0130] In this embodiment, the vehicles are divided into multiple categories:

[0131] (1) Good vehicles: Vehicles that fully conform to normal behaviors, only send true messages, and strictly maintain data integrity, and broadcast the data calculated by themselves. If an attacker shows obvious incorrect behaviors, the good vehicle will remember this attacker and reduce its direct trust in this attacker to 0.

[0132] (2) Bad vehicles: Vehicles with certain abnormal behaviors. During driving, there is a certain probability of sending true messages and a certain probability of sending false messages. Similarly, bad vehicles also strictly maintain data integrity and broadcast the data calculated by themselves.

[0133] (3) Attackers: Malicious vehicles, usually hidden among bad vehicles. During driving, they not only have a certain probability of sending true messages and a certain probability of sending false messages, but also destroy data integrity, tamper with the data calculated by themselves, and broadcast this data to interfere with the judgments of other participants in the environment.

[0134] And the effectiveness of this TSE-TMM framework is verified under the following four different scenarios.

[0135] Situation A: There are 200 vehicles in the environment, 70 of which are good vehicles and 130 are bad vehicles (without attackers). The driving behaviors of the vehicles are all normal, and there is no problem of vehicle speeding. In this situation, the vehicle trust value is completely determined by the truth or falsehood of the interaction messages and has nothing to do with the driving behavior.

[0136] Situation B: There are a total of 200 vehicles in the environment, 70 of which are good vehicles and 130 are bad vehicles (including 50 attackers), and the attackers will continuously speed. These attackers' speeding causes the good vehicles to regard them as attackers and reduce their direct trust in these vehicles to 0. At the same time, in order to prevent the trust from dropping rapidly, the attackers will cooperate with each other, initiate a tampering attack to forge their trust in other vehicles, adjust their trust value for normal vehicles to 0, and modify their trust value for other attackers to 1 to interfere with the trust system.

[0137] Situation C: There are a total of 200 vehicles in the environment, 70 of which are good vehicles and 130 are bad vehicles (including 50 attackers), and the attackers sometimes speed and sometimes drive normally. In this situation, a concept of speed factor is introduced. Each vehicle has a speed factor. For example, if the speed limit of a vehicle is 60 km / h and its speed factor is 0.8, its maximum speed is 48 km / h; if its speed factor is 1.2, its maximum speed is 72 km / h. Due to the intermittent speeding of the attackers, when the attackers speed and are observed by the good vehicles, the good vehicles will regard them as attackers and reduce their direct trust in these vehicles to 0. At the same time, the attackers will also continuously launch tampering attacks as in Situation B.

[0138] Situation D: There are a total of 200 vehicles in the environment, 70 of which are good vehicles and 130 are bad vehicles (including 50 attackers), and the attackers do not speed. Since the attackers drive normally, the good vehicles will no longer be able to intuitively detect the attackers, and the attackers will hide in the environment, but the attackers' tampering attacks will not stop.

[0139] The embodiments described above are only descriptions of the preferred embodiments of the present invention and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A security-enhanced trust evaluation method for vehicle networking combining transfer learning, characterized by the steps Including: Construct a vehicle networking model for urban roads based on the urban road conditions; Design an interactive authentication module based on the elliptic curve Diffie-Hellman protocol and the elliptic curve digital signature algorithm based on the vehicle networking model for urban roads; Based on the interactive authentication module, introduce a method for identifying true and false messages of a three-factor trust model improved by a transfer learning method, and design a transfer learning model TranferMegaCRN in the speed expectation sub-factor analysis to obtain the predicted speed of the message sender; Based on the transfer learning model TranferMegaCRN, design a vehicle trust value calculation method using an improved Bayesian method and a density-based clustering screening algorithm, and construct a vehicle networking trust management model; Based on the vehicle networking trust management model, complete vehicle trust evaluation.

2. The security-enhanced vehicle network trust evaluation method combining transfer learning according to claim 1, wherein The constructed vehicle networking model for urban roads includes four entities: vehicles equipped with on-board units OBU, roadside units RSU, trusted control centers TA, and road speed monitors.

3. The security-enhanced vehicle network trust evaluation method combining transfer learning according to claim 1, characterized in that The steps of designing the vehicle trust value calculation method include: in the direct trust value calculation, design a sliding window method with a time decay factor and a risk adjustment mechanism based on variance to optimize the calculation method based on Bayes' formula; in the indirect trust calculation, design a density-based clustering screening algorithm to extract trust feature vectors by RSU and eliminate potential malicious recommendation nodes within high-density clusters.

4. The security-enhanced vehicle network trust evaluation method combining transfer learning according to claim 1, characterized in that, The method for constructing the vehicle networking trust management model includes: Based on the vehicle trust value calculation method, construct a framework for the vehicle networking trust management model; Introduce an interactive authentication module based on the elliptic curve Diffie-Hellman protocol and the elliptic curve digital signature algorithm into the management model framework; Introduce a three-factor message recognition method improved by a transfer learning method into the management model framework to complete the construction.

5. The security-enhanced vehicle network trust evaluation method combining transfer learning according to claim 4, characterized in that, The steps of introducing an interactive authentication module based on the elliptic curve Diffie-Hellman protocol and the elliptic curve digital signature algorithm include: when a vehicle registers, negotiate a shared key with the control center TA through the elliptic curve Diffie-Hellman protocol, encrypt the real ID to achieve anonymous communication, and ensure that TA can trace the identity; during the communication process, use the elliptic curve digital signature algorithm to sign and verify all messages, and if the receiver's verification fails, mark the sender as a malicious node.

6. The security-enhanced vehicle networking trust evaluation method combining transfer learning according to claim 4, characterized in that Combine the transfer learning method with a time series model based on the meta-graph convolutional recurrent network MegaCRN to construct a transfer learning model TransferMegaCRN to predict the speed of the message sender and obtain the speed expectation sub-factor.

7. The security-enhanced vehicle networking trust evaluation method combined with transfer learning according to claim 6, characterized in that The transfer learning steps based on TransferMegaCRN include: in the input stage, the original speed data x and the time feature covariate y1 are respectively passed through their respective fully connected layers and then input into the MegaCRN model; in the MegaCRN model, the parameters of all encoder modules in the MegaCRN pre-trained via the traffic flow dataset METR-LA are frozen; a modular progressive unfreezing strategy is adopted to unfreeze the decoder and memory modules. During the unfreezing process, the deep layers of the encoder are gradually unfrozen starting from the 5th round, and the middle layers continue to be unfrozen after the 10th round, with a progressive adjustment from fixing the bottom layer to dynamically releasing the high layer; in the output stage, the output of the MegaCRN model is further processed through a fully connected layer.

8. The security-enhanced vehicle network trust evaluation method combining transfer learning according to claim 6, characterized in that, Based on the TransferMegaCRN speed prediction model, the improved true / false message recognition method considering three factors of expectation, risk, and confidence includes: the receiving vehicle calculates the expected value, risk value, and confidence value determined by the speed and position of the sending vehicle to comprehensively evaluate the message credibility, and determines the message as true when the dynamic balance condition is met, otherwise marks it as untrustworthy.

Images