Internet of vehicles trusted data sharing method and system based on deep reinforcement learning

Abstract

The application discloses a kind of based on deep reinforcement learning's vehicle networking trusted data sharing method and system thereof.The application includes trust information modeling module, trust evaluation module, trust policy updating module, threshold comparison module.The application first collects vehicle networking real-time context information, then input context information into the trust evaluation module based on evidence theory, using the comprehensive weighted average of direct trust degree and indirect trust degree to calculate global trust value.In order to improve the effectiveness of conflict evidence source synthesis, the application proposes a conflict evidence synthesis method based on Pignistic distance to ensure that false messages do not affect driving decisions.Second, in order to improve the accuracy and robustness of malicious vehicle detection, the application designs a trust policy updating model based on deep reinforcement learning, which can learn and optimize the decision strategy of the vehicle detection system through reinforcement learning algorithm.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of trust management of Internet of Vehicles, and particularly relates to a trust data sharing method and system for Internet of Vehicles based on deep reinforcement learning. BACKGROUND

[0002] Internet of Vehicles (IoV) shares vehicle perception data by using social relationships between vehicles, so as to improve traffic safety, relieve traffic congestion and provide comprehensive social services. The high mobility, dynamic topology and low neighborhood density of vehicles in IoV make the data source changeable, and collecting context data from heterogeneous vehicle-mounted devices has challenges in safety and reliability. Trust management plays an important role in context data collection, context perception and evaluation of device trustworthiness, but the existing trust decision mechanism is difficult to adapt to emergency situations and highly dynamic environments, and lacks a strategy for adaptively using context information for trust evaluation. In addition, the trust management system needs to analyze the context according to the context requirements of application programs and services, and it is challenging to filter a large amount of redundant information in a short time. SUMMARY

[0003] The application is directed to the technical problems existing in the existing IoV trust model, and relates to a trust data sharing method and system for Internet of Vehicles based on deep reinforcement learning. The application designs and deploys a context-aware trust evaluation scheme that can intelligently maintain social relationships between vehicles, develops a lightweight and efficient trust management protocol, and ensures the trustworthiness of transmission, perception data and evaluation results.

[0004] To achieve the above-mentioned purposes, the technical solutions of the application are as follows:

[0005] In a first aspect, the application provides a trust data sharing method for Internet of Vehicles based on deep reinforcement learning, which comprises the following steps:

[0006] S1, obtaining context information of each vehicle:

[0007] According to the context information of Internet of Vehicles, the context vector of each vehicle v i is extracted The context information of Internet of Vehicles includes user identity C Ui , social relationship C Ri and environmental knowledge C Ei , denoted as

[0008] S2, calculating the global trust value by using the comprehensive weighted average value of direct trust degree and indirect trust degree

[0009] S201, trust evaluation:

[0010] Using D-S evidence theory, the total uncertainty measure function UF of trust evaluation is obtained;

[0011] S202, evidence source correction based on Pignistic distance;

[0012] S203, using context information to perform trust evaluation, and obtaining a global trust value

[0013] S3, updating the trust evaluation strategy:

[0014] Using a deep reinforcement learning model to learn and update the trust evaluation strategy, and obtaining the best trust update strategy in a constantly changing network topology environment;

[0015] S4, comparing the global trust value with a predefined trust detection threshold Th, if then the vehicle v q is trustworthy, and data sharing can be performed, otherwise, warning information for v q is issued, and data sharing is refused.

[0016] In a second aspect, the application provides a trusted data sharing system for Internet of Vehicles, comprising:

[0017] A trust information modeling module, which obtains a context vector i of each vehicle v

[0018] A trust evaluation module, which obtains a global trust value

[0019] A trust strategy updating module, which uses a deep reinforcement learning model to learn and update the trust evaluation strategy, and obtains the best trust update strategy in a constantly changing network topology environment;

[0020] A threshold comparison module, which compares the global trust value with a predefined trust detection threshold Th, if then the vehicle v q is trustworthy, and data sharing can be performed, otherwise, warning information for v q is issued, and data sharing is refused.

[0021] In a third aspect, the application provides a computer readable storage medium, which stores a computer program, and when the computer program is executed in a computer, the computer executes the method.

[0022] In a fourth aspect, the present application provides a computing device comprising a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method.

[0023] The present application has the following beneficial effects:

[0024] (1) The present application proposes a trust evaluation method based on Dempster-Shafer theory, which uses the comprehensive weighted average of direct trust and indirect trust to calculate the global trust value, and provides accurate and up-to-date trust values for vehicle nodes in the communication domain. This method can improve the overall performance of the IoV system by identifying and isolating unreliable or malicious vehicle nodes, and provide a reliable data sharing environment.

[0025] (2) The trust evaluation method based on context awareness proposed by the present application can collect evidence from multiple sources, such as sensor data and shared feedback data, and can efficiently synthesize conflicting evidence sources, which helps to reduce the impact of noise, measurement errors and faults in the highly dynamic IoV environment, and provides more robust and accurate trust evaluation.

[0026] (3) The present application uses deep reinforcement learning technology to update the parameters of the trust evaluation method, which can train the agent to quickly identify and filter redundant information, reducing the computational burden. Moreover, the real-time updated training parameters according to environmental changes can improve the adaptive ability of the trust evaluation method. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is the workflow diagram of the method of the present application;

[0028] Figure 2 is the framework of the deep Q learning network algorithm;

[0029] Figure 3 is the flowchart of the trust update strategy based on the deep Q learning network;

[0030] Figure 4 shows the detection accuracy of different detection thresholds;

[0031] Figure 5 shows the detection accuracy of different detection methods. DETAILED DESCRIPTION

[0032] In order to facilitate those skilled in the art to understand the technical content of the present application, the content of the present application is further explained below in conjunction with the drawings.

[0033] The application provides a vehicular networking trusted data sharing method based on deep reinforcement learning, which specifically comprises a trust information modeling module, a trust evaluation module, a trust strategy updating module and a threshold comparison module. First, vehicular networking context information is collected, and then a global trust value is calculated by using a comprehensive weighted average value of direct trust and indirect trust based on an evidence theory and a Pignistic distance trust evaluation method. Next, a deep Q learning model is used to learn and update the trust evaluation strategy, and the model can automatically adjust the strategy and obtain a gradually optimal trust updating strategy through interaction and feedback with the environment in a constantly changing network topology environment. The method can ensure the credibility of transmission, perception data and evaluation results, and realizes safe and efficient trust management.

[0034] A vehicular networking trusted data sharing method based on deep reinforcement learning, which comprises the following steps: Figure 1

[0035] S1, acquiring context information of each vehicle:

[0036] According to the vehicular networking context information, a context vector of each vehicle v i is extracted. The vehicular networking context information comprises user identity C Ui , social relationship C Ri and environmental knowledge C Ei , and is denoted as

[0037] S2, calculating a global trust value by using a comprehensive weighted average value of direct trust and indirect trust

[0038] S201, trust evaluation:

[0039] A total uncertainty measure function UF of trust evaluation is obtained by using a D-S evidence theory.

[0040] S202, evidence source correction based on Pignistic distance

[0041] S203, trust evaluation by using context information to obtain a global trust value

[0042] S3, updating the trust evaluation strategy:

[0043] A deep reinforcement learning model is used to learn and update the trust evaluation strategy, and the best trust updating strategy is obtained in a constantly changing network topology environment.

[0044] S4, comparing the global trust value with a predefined trust detection threshold Th, if then the vehicle v​q is reliable, can share data, otherwise, a warning message will be issued for v q , refuse to share data.

[0045] In step S1, the user identity C Ui includes user ID, driving age, address and work location; for example, the identity information h1 of the evaluated person;

[0046] The social relationship C Ri includes direct social relationship and indirect social relationship, and a widely used model based on graph theory is used to represent it, that is, the average number of hops of social relationship is used to measure the closeness of social relationship between users; for example, social relationship similarity h2, common social behavior set h3;

[0047] The environmental knowledge C Ei includes message repetition rate h4 of shared messages, message loss rate h5, and message quality h6;

[0048] Step S2 specifically includes the following steps:

[0049] S201, trust evaluation:

[0050] Using Dempster shafer (D-S) evidence theory, the total uncertainty measure function UF of trust evaluation is obtained; specifically:

[0051] Dempster-Shafer (D-S) evidence theory can be used to integrate different context information sources related to the performance of IoV data sharing system, such as sensor data or data transmission accuracy and timeliness data. Each context-related information can be regarded as a source of evidence, and these evidences can be combined using Dempster's combination rule. The result of this combination is a belief function, which represents the degree of belief in different possible states of the environment, providing a more accurate representation of the environment.

[0052] S2011 D-S evidence theory includes two basic concepts of identification framework and evidence, represented by a binary tuple <Θ, M>; define the identification framework of IoV trusted data sharing Θ = {trust, distrust, uncertain}, where trust represents trusted, distrust represents untrusted, and uncertain represents uncertain, then the probability distribution function m: 2 Θ → [0, 1], where 2 Θ is the power set of Θ; if proposition B represents any subset in the identification framework, denoted as Its basic probability distribution function m(B) represents the degree of belief that proposition B is true; any piece of evidence h i(i = 1, 2, …, l), the evidence refers to a piece of context information Evidence h i The set of evidences is denoted as H = {h i | i ∈ {1, 2, …, l}}, abbreviated as evidence set.

[0053] The present application gives six types of evidence set (n = 6): the identity information of the evaluated person h1, the social relationship similarity h2, the common social behavior set h3, the message repetition rate of shared messages h4, the message loss rate h5 and the message quality h6.

[0054] S2012 Calculation of inconsistency measure function I(m):

[0055]

[0056] Where pl is the likelihood function of proposition B, describing the maximum uncertainty degree of the proposition; the proposition m(A) represents the probability distribution of proposition A;

[0057] S2013 Calculation of confusion measure function C(m):

[0058]

[0059] Where bel is the belief function of proposition B, describing the minimum uncertainty degree of the proposition;

[0060] S2014 Calculation of conflict measure function CF(m):

[0061]

[0062] Where · represents the potential of the set, and CF(m) represents the sum of the conflict degrees between proposition B and the subset proposition A;

[0063] S2015 Calculation of total uncertainty measure function UF(m):

[0064]

[0065] S202, evidence source correction based on Pignistic distance, specifically:

[0066] In real-world vehicle-to-everything (V2X) scenarios, the complexity of the environment, imperfect information collection methods, and limitations of observers often lead to inaccurate or inconsistent inferred V2X state information, resulting in multi-source information conflict problems. This invention utilizes a conflict evidence synthesis method based on Pignistic distance to correct the evidence sources, improving the effectiveness and accuracy of conflict evidence source synthesis and resolving the multi-source information conflict problem in DS evidence theory. Evidence sources include driving data provided by onboard sensors, vehicle state information provided by the vehicle control unit, communication data provided by the vehicle communication module, and traffic flow-related information provided by traffic infrastructure.

[0067] Similarity measurement between evidence sets in S2021:

[0068] Evidence set h i With evidence set h j The Pignistic distance between them, dis(m) i ,m j The metric is expressed as follows:

[0069]

[0070]

[0071]

[0072] in Let h represent the evidence h of proposition B, respectively. i ,h j The relevant Pignistic probability function, m i (B), m j (B) represents the evidence set h i ,h j The degree of confidence that proposition B is true; i, j = 1, 2, ..., l, and i ≠ j;

[0073] Then the evidence set h i With evidence set h j The similarity measure between them is sim(m) i ,m j The calculation is as follows:

[0074] sim(m i ,m j ) = 1 - dis(m i ,m j Equation (6)

[0075] S2022 Evidence Set Support Measure Calculation: Using the evidence set similarity matrix, the evidence set h is obtained. i support measure sup(m) i );

[0076]

[0077] S2023 Calculate the credibility measure of the evidence set: normalize the support measure of the evidence set, and obtain the credibility measure crd(m i ) of the evidence set h i ;

[0078]

[0079] S2024 Evidence source correction based on discount rate:

[0080]

[0081] wherein, is the corrected credibility, and the discount rate α j (j=1, 2, …, l) is calculated as follows:

[0082]

[0083] wherein, Δcrd(m j ) represents the difference value of the credibility measure of the evidence set, eu(m j ) represents the weight of the total uncertainty measure of the evidence set; UF(m j ) represents the total uncertainty measure function of the evidence set h j ;

[0084] S2025 Synthesis of the corrected evidence source based on D-S evidence rule:

[0085] The Dempster synthesis rule is used to synthesize the corrected evidence source for n-1 times; wherein the Dempster synthesis rule is as follows:

[0086]

[0087] wherein C1, C2, …, C n is a proposition under the recognition framework Θ,

[0088] S203: Use context information to perform trust evaluation, and obtain a global trust value The specific steps are as follows:

[0089] S2031 Calculate the direct trust degree:

[0090] The direct trust represents the relationship strength between two vehicles, and the direct trust degree of the vehicle v p to the evaluated vehicle v q is expressed as DT pq = <Dbel pq(trust), Dbel pq (distrust), Dbel pq (uncertain)>, where Dbel pq (trust) means v p For v q The evaluation is a trust function worthy of direct trust, Dbel pq (distrust) means v p For v q The evaluation of Dbel deems it a trust function that is not worth directly trusting, while Dbel... pq (uncertain) means v p For v q The evaluation is an uncertain direct trust function; according to equation (11), when proposition B is trust, distrust, and uncertain, the value of the direct trust function is calculated as follows:

[0091] Dbel pq (trust) = m(trust)

[0092] Dbel pq (distrust) = m(distrust)

[0093] Dbel pq (uncertain)=m(uncertain) Formula (12)

[0094] S2032 Indirect Trust Calculation:

[0095] Given source vehicle v p To evaluate the vehicle, v r For v p Recommended vehicles within the communication domain, r∈Ne(p), where Ne(p) represents v p The set of vehicles within the communication domain, v q The vehicle being evaluated, or the target vehicle;

[0096] v p For v q Indirect trust is represented by IT pq = <Ibel pq (trust), Ibel pq (distrust), Ibel pq (uncertain)>;where Ibel pq (trust) means v p For v q The evaluation is a trust function worthy of indirect trust, Ibel pq(distrust) means v p For v q The evaluation of Ibel is that it is a trust function that is not worth indirect trusting. pq (uncertain) means v p For v q The evaluation is a trust function of uncertain indirect trust; the statistical v p Other recommended vehicles within the communication domain v r Social intimacy SocFam pr Overlap with social relationships (SocCon) pr SocSim is used to express social similarity between vehicles. pr This allows for the selection of suitable recommended nodes.

[0097] SocFam pr =Inter pr / (Inter max +Inter min Equation (13)

[0098]

[0099]

[0100] Inter pr Indicates v p With v r Total number of interactions, Inter max It is v p The maximum number of interactions among all vehicle nodes within the communication domain, Inter min It is v p Minimum number of interactions among all vehicle nodes within the communication domain; Rel p With Rel r They represent v respectively p With v r Social Relationships C Ri The set of nodes;

[0101] Determine if SocSim is satisfied pr >τ, where τ represents the threshold; if it is greater than τ, then vehicle node v is considered to be... r Can be used as an evaluation vehicle v p The recommended vehicle is the one that is selected; otherwise, it is not.

[0102] Therefore, it is determined that when SocSim pr When τ > τ, vehicle node v r Can be used as source vehicle node v p The recommended node. Therefore, v p Recommended vehicle node v rTwo conditions need to be met: 1) v p has direct interaction behavior with v r , and v r also has direct interaction behavior with v q ; 2) the social similarity of v p and v r needs to meet: SocSim pr > τ.

[0103] Ibel pq (trust), Ibel pq (distrust) and Ibel pq (uncertain) are calculated as follows:

[0104]

[0105]

[0106]

[0107] wherein Dbel pr (trust) represents the trust function of the source vehicle v p evaluating the recommended vehicle v r as trustworthy, Dbel rq (trust) represents the trust function of the recommended vehicle v r evaluating the target vehicle v q as trustworthy, Dbel pr (distrust) represents the trust function of the source vehicle v p evaluating the recommended vehicle v r as untrustworthy, Dbel rq (distrust) represents the trust function of the recommended vehicle v r evaluating the target vehicle v q as untrustworthy, Dbel pr (uncertain) represents the trust function of the source vehicle v p evaluating the recommended vehicle v r as uncertain, and Dbel rq (uncertain) represents the trust function of the recommended vehicle v r evaluating the target vehicle v q as uncertain;

[0108] S2033 uses a comprehensive weighted average of direct trust and indirect trust to calculate the global trust value GT pq ;

[0109]

[0110] where a and b are the adjustment weights of direct trust and indirect trust respectively; for example, in the vehicle frequent interaction stage, direct trust can provide more valuable information, and the value of a will increase.

[0111] In summary, the global trust value is obtained, which contains three parts: the basic probability assignment function values when the proposition is trust, distrust and uncertain, which are denoted as m1, m2 and m3 respectively; the normalized global evaluation trust value obtained finally is :

[0112]

[0113] Figure 3 In step S3, the problem is described as a Markov decision model, where Episode is a round of reinforcement learning, the training stops after M rounds are performed, each Episode contains multiple time steps t, T is the maximum time step number in the Episode, and the progressive optimal strategy can be obtained through the continuous interaction and learning of the two main bodies of the agent and the environment; the agent is the executor of the trust strategy update; during the training process, the agent selects an action based on the current state and the greedy value e and interacts with the environment at each time step, the state information of the environment observed at time t is s t , and the action a t is selected according to the reinforcement learning algorithm model, the reward r t+1 and the next state s t+1 are obtained after the action is applied to the environment, and the policy is updated iteratively to converge to the optimal policy;

[0114] S301, state information s t :

[0115] The state information is formalized as a four-tuple s t ={Pos, num J , num inte , f}, where Pos is the position of the vehicle, num J is the recommended number of vehicles, num inte is the number of interactions of the vehicle, and f is the interaction frequency of the vehicle.

[0116] S302, action space a t :

[0117] The action a t is the weight a and b, which is represented as a set As={a1, a2, …, a n1}, i.e. the control action is discrete, and n1 different weights can be assigned to α or β from As; the total action space A of the reinforcement learning model is the Cartesian product of the two weight action spaces defined as follows:

[0118] A = A α × A β

[0119] α ∈ As

[0120] β ∈ As Equation (19)

[0121] where A α and A β are the action spaces of the weights α and β, respectively;

[0122] S303, reward function: the designed reward mechanism pays according to the contribution of the participants, i.e. encourages vehicles to send real trust opinion feedback;

[0123] re t = Tf t / F t Equation (20)

[0124] where Tf t is the number of messages of real feedback, and F t is the total number of feedback messages;

[0125] S304, update of the state-action function:

[0126] The state-action function is usually denoted by the symbol Q, which represents the expected value of the long-term return after performing an action in a certain state, denoted as Q value. The goal of the agent is to maximize the long-term cumulative reward by updating the model through determining the optimal trust strategy. The learning of the strategy model depends on the update of the Q table, which is used to store the values of Q. The rows and columns of the Q table represent the values of the states and actions, respectively. If the action selected by the current observation state s is a i , a i ∈ As, then the Q value function gives the expected total reward obtained by starting from state s t = s, performing action a t = a i in the first step, and thereafter performing the strategy π; the Q value function is defined as follows:

[0127] Q π (s t , a t ) = E π [re t+1 + γQ π (s t+1 , a t+1 )| s t = s, at =a i Equation (21)

[0128] The Q-value function is updated as follows:

[0129]

[0130] Where η is the learning rate, re t+1 Indicates that during the execution of action a t Then, it enters state s. t+1 Instant rewards obtained at that time; The action 'a' represents the action chosen by the agent. t+1 In the middle, Q π (s t+1 a t+1 The largest action a t+1 The corresponding Q value;

[0131] S305, Loss Function Calculation:

[0132]

[0133] Where σ and These are the function estimation Q-networks (Q-networks) during training. eval ) and the target Q-network (Q target During the training process, the AI ​​adjusts the weights and biases based on the value of the loss function to estimate the Q-value of the Q-network. eval The Q-value of the Q-network is closer to the target. target This improves the agent's action strategies; The action 'a' represents the action chosen by the agent. t+1 China makes The largest action a t+1 The corresponding Q value.

[0134] Example:

[0135] The performance of the proposed method was verified using joint simulation analysis with SUMO and OMNeT++. In the simulation, SUMO was used to generate a vehicle maneuver model with macroscopic and microscopic behaviors, and the generated tracking file was imported into the network simulator OMNeT++.

[0136] This invention introduces malicious nodes into the network to evaluate the ability of the method to handle attack behavior. It keeps the number of legitimate vehicles constant in the attack scenario and increases the presence rate of attackers in the vehicle network from 10% to 50%. This invention considers two types of attacks: external attacks - Sybil attacks and internal attacks - false message alarm attacks.

[0137] The specific implementation steps of the present invention will now be described in detail with reference to the accompanying drawings:

[0138] like Figure 1 As shown, a context-aware trust management method for vehicle-to-everything (V2X) networks based on deep reinforcement learning mainly includes a trust information modeling module, a trust evaluation module, a trust policy update module, and a threshold comparison module. The steps are as follows:

[0139] S1, Construct a trust information modeling module: Generate beacon messages at random locations in the vehicle network, including sending node ID, node location, motion status, etc., collect context information, including user identity, social relationships and environmental knowledge, and obtain context vectors.

[0140] S2, Constructing a Trust Assessment Module: This module uses a weighted average of direct and indirect trust levels. Specifically, it first utilizes orthogonal synthesis rules based on evidence theory to merge various pieces of evidence from different sources with different probability assignment functions, providing an accurate trust assessment. Then, it employs a conflict evidence synthesis method based on Pignistic distance to improve the effectiveness and accuracy of synthesizing conflict evidence sources.

[0141] S3, Construct a trust policy update module: Utilize a deep Q-learning model to learn and update the trust assessment policy, such as... Figure 2 As shown, the optimal trust update strategy is obtained in a constantly changing network topology environment.

[0142] S4, Construct a threshold comparison module: Compare the global trust value with a predefined trust detection threshold; if it is greater than the threshold, then vehicle v... q It is trustworthy and can share data; if it is less than or equal to v, it will issue a response targeting v. q The warning message indicates that data sharing is refused.

[0143] The selection of the trust detection threshold Th in S4 considers two types of malicious behavior: 1) Sybil attack (AT="Internal"), 2) spoofed alert message attack (AT="Outsider"). For example... Figure 4 As shown, when the trust threshold Th is set to 0.4, 0.5, and 0.6 respectively, the detection rate of the trust framework increases with the increase of simulation time. When Th is 0.5, the detection accuracy is higher in both Sybil attack scenarios and false alarm message attack scenarios. In the false alarm message attack scenario, the detection accuracy with a Th value of 0.5 reaches 0.95. Therefore, the selected detection threshold Th value is 0.5.

[0144] Figure 5The detection rates of different detection methods with the change of malicious vehicle rate are respectively shown in high-density scenarios and low-density scenarios. The detection rates of each detection method in the high-density scenario are generally higher than those in the low-density scenario, because the interaction behaviors between vehicles are more frequent in the high-density scenario, and the trust evaluation results are more reliable. The detection results show that the detection performance of the detection method CTVDR based on trust evaluation proposed in the application is the best, even when the malicious vehicle rate reaches 50%, the malicious vehicle node can be detected with a detection accuracy of not less than 90%, because CTVDR considers different possible improper behaviors of participating entities, and also uses a hybrid trust model to improve the performance of the network in dealing with complex attacks.

[0145] Obviously, the above embodiments are examples of the trust management method for the Internet of Vehicles, and are not limited to the embodiments. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Any modification and change made to the application still falls within the protection scope of the application.

Claims

1. A trusted data sharing method for vehicle-to-everything (V2X) networks based on deep reinforcement learning, characterized in that... The method includes the following steps: S1. Obtain the context information of each vehicle: Based on the vehicle network context information, extract the v of each vehicle i context vector The context information of the vehicle network includes user identity C Ui Social Relationships C Ri Environmental Knowledge C Ei , recorded as The social relationship C Ri This includes both directly connected and indirectly connected social relationships, represented using a broad graph theory-based model, which measures the closeness of social relationships between users by the average number of hops in the social relationship. S2. Calculate the global trust value using the weighted average of direct and indirect trust levels. S201, Trust Evaluation: Using DS evidence theory, the total uncertainty measure function UF for trust evaluation is obtained; specifically: The S2011 DS evidence theory includes two basic concepts: identification framework and evidence, represented by a tuple <Θ,M>. The identification framework for trusted data sharing in the Internet of Vehicles (IoV) is defined as Θ = {trust, distrust, uncertain}, where trust represents trustworthy, distrust represents untrustworthy, and uncertain represents uncertain. The probability assignment function is m:

2. Θ →[0,1], where 2 Θ It is the power set of Θ; if proposition B represents any subset in the recognition frame, denoted as Its basic probability assignment function m(B) represents the degree of confidence that proposition B is true; any number of pieces of evidence h are defined on the recognition frame Θ. i (i = 1, 2, ..., l), the evidence refers to a piece of contextual information. Evidence h i The set is represented as H = {h} i |i∈{1,2,…,l}}, simply referred to as the evidence set; Calculation of the S2012 inconsistency measure function I(m): Where pl is the likelihood function of proposition B, describing the maximum degree of uncertainty regarding the validity of the proposition; proposition m(A) represents the probability assignment of proposition A; Calculation of the S2013 disorder measure function C(m): Where bel is the confidence function of proposition B, which describes the minimum degree of uncertainty of the proposition's validity; Calculation of the S2014 conflict measure function CF(m): Where |·| represents the cardinality of the set, and CF(m) represents the sum of the degree of conflict between proposition B and subset proposition A; Calculation of the total uncertainty measure function UF(m) in S2015: S202, Evidence source correction based on Pignistic distance; specifically: Similarity measurement between evidence sets in S2021: Evidence set h i With evidence set h j The Pignistic distance between them, dis(m) i ,m j The metric is expressed as follows: in Let h represent the evidence h of proposition B, respectively. i ,h j The relevant Pignistic probability function, m i (B), m j (b) represents the evidence set h i ,h j The degree of confidence that proposition B is true; i, j = 1, 2, ..., l, and i ≠ j; Then the evidence set h i With evidence set h j The similarity measure between them is sim(m) i ,m j The calculation is as follows: sim(m i ,m j ) = 1 - dis(m i ,m j ) Equation (6) S2022 Evidence Set Support Measure Calculation: Using the evidence set similarity matrix, the evidence set h is obtained. i support measure sup(m) i ); Calculation of the credibility measure of the S2023 evidence set: Normalize the support measure of the evidence set to obtain the evidence set h. i The reliability measure crd(m) i ); S2024 Evidence Source Correction Based on Discount Rate: in, This is the corrected credibility, discount rate α. j The calculation of (j=1,2,…,l) is as follows: Wherein, Δcrd(m j ) represents the difference in the credibility measure of the evidence set, eu(m j UF(m) represents the weight of the total uncertainty measure of the evidence set; j ) represents the evidence set h j The total uncertainty measure function; S2025 Evidence Sources Based on DS Evidence Rules and Modifications: The modified evidence sources were synthesized n-1 times using Dempster's synthesis rules, where the Dempster synthesis rules are as follows: Where C1, C2, ..., C n It is a proposition under the recognition framework Θ. S203. Perform trust assessment using contextual information to obtain a global trust value. Specifically: S2031 Direct Trust Calculation: Direct trust is represented as the strength of the relationship between two vehicles, used to evaluate vehicle v. p For the vehicle being evaluated v q Direct trust level is represented by DT pq = <Dbel pq (trust), Dbel pq (distrust), Dbel pq (uncertain)>, where Dbel pq (trust) means v p For v q The evaluation is a trust function worthy of direct trust, Dbel pq (distrust) means v p For v q The evaluation of Dbel deems it a trust function that is not worth directly trusting, while Dbel... pq (uncertain) means v p For v q The evaluation is an uncertain direct trust function; according to equation (11), when proposition B is trust, distrust, and uncertain, the value of the direct trust function is calculated as follows: Dbel pq (trust)=m(trust) Dbel pq (distrust)=m(distrust) Dbel pq (uncertain) = m(uncertain) Equation (12) S2032 Indirect Trust Calculation: Given source vehicle v p To evaluate the vehicle, v r For v p Recommended vehicles within the communication domain, r∈Ne(p), where Ne(p) represents v p The set of vehicles within the communication domain, v q The vehicle being evaluated, or the target vehicle; v p For v q Indirect trust is represented by IT pq = <Ibel pq (trust), Ibel pq (distrust), Ibel pq (uncertain)>;where Ibel pq (trust) means v p For v q The evaluation is a trust function worthy of indirect trust, Ibel pq (distrust) means v p For v q The evaluation of Ibel is that it is a trust function that is not worth indirect trusting. pq (uncertain) means v p For v q The evaluation is a trust function of uncertain indirect trust; the statistical v p Other recommended vehicles within the communication domain v r Social intimacy SocFam pr Overlap with social relationships (SocCon) pr SocSim is used to express social similarity between vehicles. pr This allows for the selection of suitable recommended nodes. SocFam pr =Between pr / (Between max +Between min ) Equation (13) Inter pr Indicates v p With v r Total number of interactions, Inter max It is v p The maximum number of interactions among all vehicle nodes within the communication domain, Inter min It is v p Minimum number of interactions among all vehicle nodes within the communication domain; Rel p With Rel r They represent v respectively p With v r The set of nodes in a social circle; Determine if SocSim is satisfied pr >τ, where τ represents the threshold, if it is true, then vehicle node v is considered to be true. r As an evaluation vehicle v p The recommended vehicle is the one that is selected; otherwise, it is not. Ibel pq (trust), Ibel pq (distrust) and Ibel pq The calculation method for (uncertain) is as follows: Dbel pr (trust) indicates the source vehicle v p Recommended vehicle v r The evaluation is a trust function that is trustworthy, Dbel rq (trust) indicates a recommended vehicle v r For target vehicle v q The evaluation is a trust function that is trustworthy, Dbel pr (distrust) indicates the source vehicle v p Recommended vehicle v r The evaluation is a trust function that is not trustworthy, Dbel rq (ditrust) indicates a recommended vehicle v r For target vehicle v q The evaluation is a trust function that is not trustworthy, Dbel pr (uncertain) indicates the source vehicle v p Recommended vehicle v r The evaluation is an uncertain trust function, Dbel rq (ucertain) indicates recommended vehicle v r For target vehicle v q The evaluation is an uncertain trust function; S2033 uses a weighted average of direct and indirect trust levels to calculate the global trust value GT. pq ; Where α and β are the adjustment weights for direct trust and indirect trust, respectively; And because of the global trust value GT pq It consists of three parts: the basic probability assignment function values ​​for the propositions trust, distrust, and uncertainty, denoted as m1, m2, and m3 respectively; and the final normalized global evaluation trust value. for: S3. Update the trust assessment strategy: By using deep reinforcement learning models to learn and update trust evaluation strategies, the optimal trust update strategy can be obtained in a constantly changing network topology environment through interaction and feedback with the environment. S4. Set global trust value Compare with a predefined trust detection threshold Th, if Then vehicle v q It is trustworthy and can share data; otherwise, it will issue a warning against v. q The warning message indicates that data sharing is refused.

2. The method according to claim 1, characterized in that... In step S1, the user identity C Ui This includes user ID, driving experience, home address, and work location.

3. The method according to claim 2, characterized in that... In step S1, user identity C Ui This includes the identity information of the person being evaluated (h1); social relationships (C). Ri This includes social relationship similarity h2, shared social behavior set h3; environmental knowledge C Ei This includes message duplication rate (h4), message loss rate (h5), and message quality (h6) for shared messages.

4. The method according to claim 1, characterized in that... In step S3, the problem is described as a Markov decision model, where the optimal policy is obtained through continuous interaction between two agents: an agent and the environment. The agent is the executor of the trust policy update. During training, the agent continuously interacts with the environment, and at time t, the observed state information of the environment is s. t And select action a based on the reinforcement learning algorithm model. t A reward is given after performing an action on the environment. t+1 And the state s at the next moment t+1 After iterative policy updates, the policy converges to the optimal policy. S301, Status Information s t : The state information is formalized as a quadruple s t ={Pos,num J ,num inte ,f}, where Pos is the position of the vehicle, num J This is the recommended number of vehicles, num. inte is the number of vehicle interactions, and f is the vehicle interaction frequency; S302, Action a t : The action a t These are the adjusted weights α for direct trust and β for indirect trust, denoted as the set As = {a1, a2, ..., a...} n1 In other words, the control actions are discrete, and n1 different weights are assigned to α or β from As; the total action space A of the reinforcement learning model is the Cartesian product of the two weight action spaces, defined as follows: A=A α ×A β α∈As β∈As Equation (19) Where A α With A β These are the action spaces for adjusting the direct trust weight α and the indirect trust weight, respectively. S303, Reward Function: The designed reward mechanism pays based on the participant's contribution, that is, it encourages vehicles to send genuine trust feedback; re t =Tf t / F t Equation (20) Among them Tf t It represents the number of actual feedback messages, F. t It represents the total number of feedback messages; S304, Update of the state action function: The state-action function, denoted by Q, represents the expected long-term reward obtained by performing an action in a given state, denoted as the Q-value. The agent's goal is to maximize the long-term cumulative reward by determining the optimal policy model. The learning of the policy model depends on the updating of the Q-table, which stores the values ​​of Q. The rows and columns of the Q-table represent the values ​​of the state and action, respectively. If the action chosen from the observed state s at the current time step is a... i a i If ∈As, then the Q-value function gives the value of state s under the discount factor γ. t =Starting with s, the first step is to execute action a. t =a i The expected total reward obtained by executing strategy π thereafter; the Q-value function is defined as follows: Q π (s t ,a t )=E π [re t+1 +γQ π (s t+1 ,a t+1 )|s t =s,a t =a i ] Equation (21) The Q-value function is updated as follows: Where η is the learning rate, re t+1 Indicates that during the execution of action a t Then, it enters state s. t+1 Instant rewards obtained at that time; The action 'a' represents the action chosen by the agent. t+1 In the middle, Q π (s t+1 ,a t+1 The largest action a t+1 The corresponding Q value; S305, Loss Function Calculation: Where σ and These are the function estimates of the Q-values ​​of the Q-network during training. eval Q-value of the target Q-network target During the training process, the AI ​​adjusts the weights and biases based on the value of the loss function to estimate the Q-value Q of the Q-network. eval The Q-value of the Q-network is closer to the target. target This improves the agent's action strategies; The action 'a' represents the action chosen by the agent. t+1 China makes The largest action a t+1 The corresponding Q value.

5. A vehicle-to-everything (V2X) trusted data sharing system implementing the method of any one of claims 1-4, characterized in that... include: The trust information modeling module obtains the v of each vehicle based on the vehicle network context information. i context vector The trust assessment module calculates the global trust score using a weighted average of direct and indirect trust levels. The trust policy update module uses a deep reinforcement learning model to learn and update the trust evaluation policy. Through interaction and feedback with the environment, it obtains the optimal trust update policy in a constantly changing network topology environment. The threshold comparison module compares the global trust value. Compare with a predefined trust detection threshold Th, if Then vehicle v q It is trustworthy and can share data; otherwise, it will issue a warning against v. q The warning message indicates that data sharing is refused.

Images