A Trusted Task Offloading Method Based on Autonomous Collaboration of Vehicle-Edge-Cloud

Abstract

The present invention discloses a trusted task offloading method based on vehicle-edge-cloud autonomous collaboration. This method can be applied to dynamic heterogeneous vehicle networks to solve inefficient problems such as task accumulation, resource preemption, and poor collaboration caused by false tasks and malicious nodes. First, the present invention provides a trust evaluation model for offloading tasks and collaborating nodes, which accurately identifies and filters false tasks and malicious nodes by analyzing data sensitivity, resource preemption, historical offloading success rate, and collaboration feedback evaluation. Then, the present invention uses the discrete particle swarm algorithm to iteratively obtain vehicle-edge-cloud collaborative offloading decisions, redesigns particle coding, fitness functions, and correction mechanisms, and introduces random quantities, greed, and mapping functions to enhance the diversity of offloading decisions. Compared with traditional task offloading methods, the present invention exhibits better energy consumption and latency performance in scenarios with compact and evenly distributed tasks, effectively improving offloading efficiency and security.

Description

Technical Field

[0001] The present invention belongs to the field of vehicle networking covered by the new generation of information technology, and relates to a trusted task offloading method based on autonomous collaboration of vehicle-edge-cloud. Background Art

[0002] Vehicle networking uses roadside sensing, edge computing, and cloud fusion technologies to achieve information exchange and sharing among people, vehicles, and roads, and provides collaborative perception, decision-making, and control of complex traffic environments. With the development of vehicle networking technology, emerging applications such as real-time navigation, cooperative lane change, and collision warning have been installed in the vehicle environment, posing higher requirements for the computing, storage, and response speed of vehicles. For example, in services such as real-time navigation and collision warning, it is necessary to recommend multiple driving routes in real time based on dynamic road conditions and vehicle positions, and the task calculation density is relatively high; in the collision warning service, in order to solve problems such as conflict avoidance and lane change, the calculation delay is generally required to be within dozens of milliseconds; in addition, in intelligent assisted driving scenarios such as unmanned driving, task calculations usually also have multiple characteristics such as being calculation-intensive, latency-sensitive, and resource-consuming. Therefore, in the face of a large number of calculation-intensive and latency-sensitive calculation tasks generated by application programs, traditional traffic intelligent terminals such as vehicles are difficult to meet the system task processing requirements due to limited computing power and serious battery consumption. Especially when the vehicle density is relatively high, there may be an imbalance between supply and demand where the available resources are limited but the required resources are too high.

[0003] Task offloading is a key technology to improve the task processing efficiency of vehicle networking. It offloads tasks that are difficult for vehicles to process to the edge or cloud for execution, and uses the powerful computing performance of the cloud and the latency advantage of the edge being "local and nearby" to make up for the deficiencies of vehicles in terms of computing, storage, and energy efficiency. In a vehicle-edge-cloud collaborative vehicle networking system, there are multiple intelligent agent nodes such as vehicle terminals, edge servers, and cloud servers, which have heterogeneous computing and storage capabilities. By using autonomous collaboration and offloading of multiple intelligent agents in vehicle-edge-cloud, resource fusion and computing collaboration can be achieved, effectively improving the task processing efficiency and traffic service level.

[0004] However, although a large number of task offloading methods have been proposed in the current academic and industrial circles, most of these methods aim to reduce latency and energy consumption, and rarely analyze the security and trust issues that may be faced during the task offloading process. In fact, due to the open, collaborative, and dynamic characteristics of the vehicle networking environment, and the large number of vehicles and service nodes involved, the security and reliability issues faced are more complex than those of ordinary networks. First, due to the existence of malicious request nodes such as free riders and selfishness, they preempt and consume the resources of edge and cloud service nodes, resulting in the death of service nodes due to excessive resource consumption, causing problems such as piled-up offloading tasks and high execution failure rates. Second, the essence of task offloading is the outsourcing of computing. If it is offloaded to malicious service nodes, it will seriously threaten privacy and data security. Especially in the computing mode of vehicle-edge-cloud collaboration, due to interest conflicts among service nodes, poor collaboration and resource preemption occur frequently. Finally, the vehicle networking has characteristics such as resource heterogeneity, dynamic access, and vehicle mobility, and the offloading process is affected by channel quality, propagation loss, server availability, etc., making it difficult to transplant and apply the existing few task offloading methods to real vehicle networking scenarios.

[0005] To address the above problems, the present invention provides a trusted task offloading method based on autonomous collaboration of vehicle-edge-cloud for the requirements of trusted and efficient task offloading. First, a trust mechanism is used to identify and filter false tasks and malicious nodes, then the discrete particle swarm algorithm is used to iteratively obtain the offloading decision, and finally, through trust perception and mutual benefit enhancement of collaborative offloading, the offloading efficiency and security are improved. Summary of the Invention

[0006] The present invention provides a trusted task offloading method based on autonomous collaboration of vehicle-edge-cloud, which can be applied to dynamic heterogeneous vehicle networking to solve inefficient problems such as task accumulation, resource preemption, and poor collaboration caused by false tasks and malicious nodes.

[0007] The technical solution adopted by the present invention is as follows: First, by analyzing data sensitivity, resource preemption, historical offloading success rate, and collaborative feedback evaluation, false tasks and malicious nodes are accurately identified and filtered; then, the discrete particle swarm algorithm is used to iteratively obtain the vehicle-edge-cloud collaborative offloading decision, and the particle coding, fitness function, and correction mechanism are redesigned, and a random quantity, greed, and mapping function are introduced to enhance the diversity of offloading decisions, mainly including the following steps:

[0008] Step 1: Evaluate the authenticity of the to-be-processed tasks generated in the current period. One is based on the data sensitivity of the task itself, and the other is based on the task request frequency of the task generation source node. For any task T i , a trust evaluation function is constructed based on its task sensitivity and source node resource preemption to conduct authenticity evaluation, which is specifically expressed as:

[0009]

[0010] Among them, α and β are weight adjustment factors, j is the subscript of the source node N that generates task T i of j The subscript of z is used to traverse K tasks, and the subscript of x is used to traverse all nodes. represents the total number of nodes; and are the amount of sensitive data and the total amount of data involved in task T i K is the total number of tasks during the generation period of this task. is the number of task processing requests initiated by the task source node in the k-th observation period. Then, based on the authenticity obtained from task evaluation classification processing is performed. If the authenticity is lower than the set threshold it is determined that the task has an attack risk and is removed from the task set If the authenticity is higher than the threshold it is retained in the set and enters the subsequent offloading decision stage.

[0011] Step 2: Evaluate the trustworthiness of the vehicles and edge nodes participating in task calculation. One is based on the historical task processing success rate of the nodes, and the other is based on the feedback evaluation of the collaborative interaction nodes. For any node N i , analyze its historical task success rate and interactive collaboration reputation, introduce the weight adjustment factor ω, and construct its trust evaluation function in the k-th observation period:

[0012]

[0013] Among them, and are the number of successful and failed tasks processed by node N i in the k-th observation period, is the feedback evaluation provided by the j-th interactant about node N i , R j is the reliability of the j-th interactant, and A is the total number of interactive nodes providing feedback evaluations. Then, considering the behavioral trends and time decay characteristics of the nodes in multiple observation periods, construct a comprehensive evaluation and filtering model. Divide each period into M trust observation periods. Considering time decay, the comprehensive trustworthiness of node N i in the k-th observation period can be expressed as:

[0014]

[0015] Among them,

[0016] Among them, is the decay function of the node trust memory retention with respect to time t, η 1 = 1.84, η 2 = 1.25, t represents the difference between the time when the corresponding trust relationship is generated and the current time, in minutes, t j represents the difference between the j-th observation period and the current time, t k represents the difference between the k-th observation period and the current time. Then, based on the node trust degree, classification processing is performed. If the node credibility is lower than the trust threshold then it is determined that the node may be a malicious node and it is deleted from the node set ; if the node credibility is higher than the threshold then it is allowed to enter the subsequent offloading decision stage and serve as a candidate node for task offloading and computing.

[0017] Step 3: Taking real tasks and trusted nodes as objects, improve and redesign the steps of population initialization, fitness evaluation, speed and position update, and effectiveness correction of the discrete particle swarm algorithm, and iteratively obtain the optimal task offloading decision. First, determine the particle position and velocity encoding. The fixed encoding length is n, where n is the number of tasks. Each value of the position encoding is null or n i , represents the task processing node. Each value of the particle velocity encoding is 0 or 1. 0 means to adjust the corresponding position encoding segment in the next iteration, and 1 means not to adjust the position segment in the next iteration. Then, use the local priority-random adjustment algorithm to generate the initial particle population; then, based on the optimization goal of minimizing the weighted sum of energy consumption and delay, construct the fitness function for particle evaluation, which is specifically expressed as:

[0018]

[0019] Among them, is the processing decision of task T i with respect to node n j . If then it means to allocate task T i to n j , and at this time the task processing cost is Among them, is the energy consumption generated by allocating task T i to node n j for processing, is the energy consumption generated by allocating task T i to node n jHandle the resulting delay, where λ is the weight adjustment factor for energy consumption and delay, and find the individual optimal particle and the global optimal particle based on the evaluated particle fitness; then, introduce random quantities, mapping functions, etc. to update the velocity encoding of the particle in the next iteration, and update the position encoding in the next iteration based on the velocity vector and in combination with the individual and optimal particles; finally, correct the particle effectiveness against the node resource constraint and the task delay constraint; repeat the iteration until a convergent offloading decision is obtained or the maximum number of iterations is reached. After obtaining the optimal offloading decision, execute the task offloading service, and update the node trust evaluation model in reverse based on the task completion quality, node cooperation situation, etc.

[0020] The method of the present invention has the following beneficial effects: First, by evaluating the authenticity and trustworthiness of offloading tasks and cooperative nodes, false tasks and malicious nodes can be identified and filtered before offloading, reducing the decision space and saving offloading costs; then, by using the discrete particle swarm algorithm to obtain the offloading decision and introducing random quantities, greed, and mapping functions, the global optimization ability of the offloading decision can be improved, and the efficiency of task allocation and offloading can be enhanced. Generally speaking, the present invention shows good energy consumption and delay performance in both the vehicle-to-everything (V2X) scenarios with compact task distribution and uniform distribution, and simultaneously improves the offloading efficiency and security. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic diagram of the network architecture applied to the present invention;

[0022] Figure 2 It is the overall framework diagram of the implementation of the method of the present invention;

[0023] Figure 3 It is a schematic diagram of the particle encoding involved when implementing the offloading decision of the present invention;

[0024] Figure 4 It is a schematic diagram of the particle velocity and position update involved when implementing the offloading decision of the present invention;

[0025] Figure 5 It is a schematic diagram of the comparison of the node identification rates between the method of the present invention and the comparative method;

[0026] Figure 6 It is a schematic diagram of the comparison of the task identification rates between the method of the present invention and the comparative method;

[0027] Figure 7 It is a schematic diagram of the comparison of the energy consumption between the method of the present invention and the comparative method;

[0028] Figure 8 It is a schematic diagram of the comparison of the delays between the method of the present invention and the comparative method;

[0029] Figure 9 It is a schematic diagram of the comparison of the task benefits between the method of the present invention and the comparative method;

[0030] Figure 10 This is a schematic diagram comparing the task completion rates of the method of the present invention and the comparative method. Detailed implementation manners

[0031] To facilitate understanding of the essence of the present invention, the technical parameters mainly involved in the present invention will be defined and described first.

[0032] Table 1 Symbol definitions and descriptions of main technical parameters

[0033]

[0034]

[0035] Next, the network scenarios applicable to the present invention will be described.

[0036] Figure 1 The following shows a schematic diagram of the network architecture applied in the present invention. The entire system consists of m vehicle nodes, n edge nodes, and 1 cloud node. At the same time, multiple roadside units and base stations are also deployed to assist in communication and data processing. Denote the vehicle set as Denote the edge server set as Denote the set of all task computing nodes n i , including vehicle, edge, and cloud nodes. The task offloading involved in the present invention adopts a binary full offloading mode. Each task can only be executed locally on the vehicle or offloaded to the edge server or the cloud as an independent whole. Denote the task set as For each task T i , its characteristics can be represented by the tuple , where T ID is the task number, is the identity identifier of the source node that generates the task, c i is the CPU cycles required for task computing, d i is the input data volume, t i is the maximum tolerable delay, represents the task processing decision, represents that the task T i is assigned to the computing node n j , then represents not assigning the task to n j .

[0037] The optimization objective of the method of the present invention is to minimize the weighted sum of the average task processing energy consumption and delay. For the task T i , assuming it is processed by the node n j , the energy consumption generated during the processing is The delay is The objective optimization function is expressed as follows:

[0038]

[0039] Wherein, is the number of tasks, is the number of nodes, is the number of tasks successfully executed, is the task processing decision, λ is a weight parameter, λ ∈ [0, 1], which is used to achieve the dynamic trade-off between task latency and system energy consumption. C1 - C4 are constraint conditions, where C1 and C2 indicate that a task can be executed by at most one node completely and independently, C3 indicates that the resources required for task execution cannot be greater than the remaining available resources of the computing node, and C4 indicates that the execution time of task T i must be within its tolerance delay.

[0040] Next, the specific implementation process of the present invention will be described.

[0041] Figure 2 The overall framework diagram of the implementation of the method of the present invention is shown, and the complete implementation steps are as follows

[0042] Step 1: Evaluate the authenticity of the tasks to be processed generated in the current cycle. One is based on the data sensitivity of the tasks themselves, and the other is based on the offloading request frequency of the task generation source node. If the task authenticity is lower than the set threshold then it is determined that the task has an attack risk and it is deleted from the task set and the remaining real tasks enter the offloading decision stage;

[0043] Step 2: Evaluate the trustworthiness of the vehicles and edge nodes participating in task calculation. One is based on the historical task processing success rate of the nodes, and the other is based on the feedback evaluation of the collaborative interaction nodes. The comprehensive credibility of the nodes is obtained by synthesizing the behavior trends and time decay characteristics of the nodes in multiple observation periods. If the node credibility is lower than the set threshold then it is determined that the node may be a malicious node and it is deleted from the node set and the remaining trusted nodes enter the offloading decision stage;

[0044] Step 3: Take real tasks and trusted nodes as objects, improve and redesign the steps of population initialization, fitness evaluation, speed and position update, and effectiveness correction of the discrete particle swarm algorithm, iteratively obtain the optimal offloading decision of the tasks, execute the offloading service, and update the node trust evaluation model in reverse based on the task completion quality, node collaboration situation, etc.

[0045] It can be seen that the implementation of the method of the present invention mainly includes three key steps: authenticity evaluation of offloading tasks, trust evaluation of collaborative nodes, and offloading decision based on discrete particle swarm. The specific implementation schemes of these three steps are introduced in detail below:

[0046] ① Evaluate the authenticity of the tasks to be processed generated in the current period. To prevent the current task from excessively requesting relevant data from non-self nodes and causing data theft, privacy leakage and other behaviors, the present invention first analyzes the data sensitivity of the task itself. For any task T i , assuming that the amount of sensitive data and the total amount of data it involves are and respectively, then the data sensitivity characteristic of this task can be expressed as:

[0047]

[0048] where α is a weight adjustment factor, 0 ≤ α ≤ 1, the first summation term is the proportion of the task's own sensitive data, the second summation term is the proportion of sensitive data in the total tasks, and K is the total number of tasks generated in the period where this task is located.

[0049] In addition, the present invention also analyzes the resource preemption of the source node that generates the task, mainly based on the task processing request frequency of the source node within the current time window. Divide a period into multiple trust observation periods. Assuming that the source node corresponding to task T i is In the k-th observation period, the number of task processing requests initiated by the source node is Then the resource preemption of the source node that generates the task can be expressed as:

[0050]

[0051] Then, comprehensively considering the task sensitivity and the resource preemption of the source node, introduce the weight adjustment factor β, and establish the following expression for evaluating the authenticity of the task:

[0052]

[0053] Finally, classify the task according to the task authenticity . If the authenticity is lower than the set threshold , then it is determined that the task has an attack risk and it is deleted from the task set . If the authenticity is higher than the threshold , then it is retained in the set and enters the subsequent offloading decision stage.

[0054] ②Perform trustworthiness evaluation on the vehicles and edge nodes participating in task calculation. To ensure the comprehensiveness and objectivity of trust evaluation, the present invention takes the historical behavior of nodes as the basis for trust evaluation and obtains it directly or indirectly in two ways. One is based on the success rate of previous task processing of nodes, and the other is based on the feedback evaluation of collaborative interaction nodes. Therefore, node N i The trust characteristics in the k-th observation period can be expressed as

[0055] For the feature Assume that in the k-th observation period, node N i The number of successful and failed tasks processed are respectively and Then there are:

[0056]

[0057] For the feature It is provided by other nodes that have had direct interaction with the node to be evaluated. Since this method depends on the reliability of the provider, therefore, it is necessary to ensure that the node providing the trust evidence is itself trustworthy. Assume that in the k-th observation period, there are A nodes providing feedback evaluations i about node N The trust degree of the j-th node is R j , then there are:

[0058]

[0059] Finally, by integrating the task completion quality and interaction and cooperation reputation of the node, and introducing the weight adjustment factor ω, the basic trust degree of the node in the k-th observation period can be expressed as:

[0060]

[0061] The present invention takes into account that node trust is a dynamically changing attribute that decays over time, and the behavior closer to the present can better reflect the current behavior trend of the node. Therefore, based on the Ebbinghaus human brain memory forgetting theory, a decay function of the node trust memory retention amount with respect to time t is constructed Specifically, it is expressed as follows:

[0062]

[0063] Among them, The trust information retention rate that decays over time is expressed as a percentage. It starts from the current time, t represents the difference between the time when the corresponding trust relationship is generated and the current time, in minutes, and in experiments, η 1 = 1.84, η 2 = 1.25.

[0064] Divide each period into M trust observation periods. Considering time decay, comprehensively evaluate the trust of nodes at the M observation period nodes to obtain node N i The comprehensive trust degree in the k-th observation period is as follows:

[0065]

[0066] Finally, perform classification processing based on the node trust degree. If the node credibility is lower than the trust threshold Delete it from the node set ; if the node credibility is higher than the threshold Then let it enter the subsequent offloading decision stage and serve as a candidate node for task offloading and computing.

[0067] ③ Use the discrete particle swarm algorithm to iteratively obtain the vehicle-edge-cloud collaborative offloading decision. After completing the offloading task and collaborative node evaluation and filtering, obtain the real task set and the trusted node set And use the discrete particle swarm algorithm for task allocation and offloading decision. The present invention maps the discrete problem of offloading decision to the continuous particle motion space. The specific steps are as follows: First, particle coding and initialization; then, particle fitness evaluation; then, particle velocity and position update; finally, correct the particle validity. Repeat the second to fourth steps until convergence or the iteration number is reached.

[0068] Figure 3 The particle coding schematic diagram involved in the present invention is shown. In the present invention, each particle represents an allocation scheme, including two parameters: position and velocity. Assuming there are a total of n tasks, encode the position parameters in sequence according to the task encoding, which can be expressed as where the position segment represents the allocation scheme of task T 7 . A valid position encoding should meet the following conditions: (1) The encoding length is fixed at n, where n is the number of tasks; (2) Each bit takes a value of null or n i , represents the task processing node, including vehicle, edge, and cloud nodes. The particle velocity can be expressed as A valid velocity encoding should meet the following conditions: (1) The encoding length is fixed at n; (2) Each bit takes a value of 0 or 1, represents adjusting the position encoding segment corresponding to T 7 in the next iteration; if then keep the allocation status of task T 7 unchanged.

[0069] Then, use the local priority-random adjustment algorithm for particle initialization. Create a candidate set of tasks to be allocated Candidate set of nodes Traversal set of nodes The initialization process is as follows: (1) Randomly select a task T from the set , and by default, its processing node is the source node vehicle v that generates T i , check whether the C1 - C4 constraint conditions are met. If it is a valid allocation that satisfies the constraints, encode the corresponding position segment of T i as the vehicle number; if not, randomly find a node n that satisfies the constraints from the node set SR , and encode the corresponding position segment as the node n i number; finally, if no node that meets the requirements can be found after traversing the entire , encode it as null; (2) Update the remaining resource parameters of the processing node of T. If it is found that the node has no available resources, delete it from the set k , and set the next - round node traversal set k ; (3) Randomly generate 0 or 1 as the initial speed encoding of T , and delete T from i ; (4) Randomly select a task again from , and repeat the above process until all tasks are traversed; (5) Sort in sequence according to the task numbers to obtain a legal task - allocation particle. (3) Randomly generate 0 or 1 as the initial speed encoding of T i , and delete T from i ; (4) Randomly select a task again from ; (5) Randomly select a task again from , and repeat the above process until all tasks are traversed; (5) Sort in sequence according to the task numbers to obtain a legal task - allocation particle.

[0070] Next, evaluate the particle fitness. The present invention constructs a fitness function based on the goal of minimizing the weighted sum of energy consumption and delay, and designs the fitness evaluation function as follows:

[0071]

[0072] Among them, is the processing decision of task T i with respect to node n j . If , it means that task T i is allocated to n j . Calculate, and at this time, the task - processing cost is λ is the adjustment weight factor of energy consumption and delay. When the task - processing cost is higher, the fitness is smaller, indicating that the particle is inferior. Then, find the individual - optimal particle and the global - optimal particle to guide the particle to approach the optimal particle. Assume that the particle swarm includes μ particles, has experienced I rounds of update iterations, and the individual - optimal particle and the global - optimal particle are defined as follows:

[0073] Denotes the individual optimal particle, for each particle , it refers to the position value with the maximum fitness of itself in this I-th round;

[0074] Denotes the global optimal particle, which refers to the position value of the particle with the maximum fitness among all particles in this I-th round.

[0075] Then, update the particle velocity and position. Denote the velocity of the particle as For any velocity segment v i , if v i =0, then keep the position segment of task T i unchanged in the next iteration, otherwise, the allocation scheme of T needs to be adjusted. Assume that the position of the particle i in the k-th iteration is denoted as The individual optimal particle is The global optimal particle is The global optimal particle is Then its velocity in the (k + 1)-th iteration is as follows:

[0076]

[0077] Among them, represents the exclusive OR operation, r 1 and r 2 are two random numbers between 0 and 1, c 1 and c 2 represent the acceleration, which is used to adjust the maximum learning step, usually set to 1.494. The velocity update of the particle mainly considers three aspects. One is the inertia of the particle's own velocity The second is the influence of the particle's own historical experience When c 1 is 0, it means not considering its own experience, and at this time, the group diversity will be lost; the third is the influence of the historical experience of other particles When c 2 is 0, it means not considering the experience of others, and at this time, due to the lack of information sharing, the convergence will slow down.

[0078] To ensure is a binary vector to represent the 0 and 1 states of the particle velocity, let Y = {y 1 , y 2 ,... y n}, For each task T i 's velocity segment The present invention defines its calculation as:

[0079]

[0080] Among them, is a random number between 0 and 1, which can balance the convergence speed, avoid particle prematurity, and prevent falling into local optimality. At the same time, the present invention introduces the function which is symmetric about 0.5 and can map to the numerical interval of 0-1 without being restricted by the domain of definition, so as to realize binary value.

[0081] After obtaining the particle velocity of the next iteration, the present invention updates the particle position according to For any position segment in the particle its position in the (k + 1)-th iteration is as follows:

[0082]

[0083] In the above formula, if we adjust the position segment of task T i in the particle, and make it adjusted to the position segment with the highest fitness.

[0084] Figure 4 FIG. is a schematic diagram of the update of the particle velocity and position involved in the present invention. The particle encoding and fitness in the figure are randomly set and are only used for illustrative explanation. First, obtain the particle as well as the individual optimal particle and the global optimal particle, and randomly generate a decimal between 0 and 1 for, and update to obtain the velocity. The particle velocity will present different results with different, which helps to improve the particle diversity. Then, according to the updated particle velocity, adjust the particle position. Assume that the fitness of the individual optimal particle and the global optimal particle of the particle in the figure are {0.1, 0.2, 0.1, 0, 0.1, 0.2, 0.1}, {0.1, 0.2, 0.1, 0.3, 0.4, 0.2, 0.1}, {0.1, 0.1, 0.1, 0.3, 0.4, 0.4, 0.1} respectively. For the position segment to be adjusted, the method of the present invention selects the encoding scheme with the highest fitness and updates to obtain Figure 5 the position encoding shown, and the fitness of the adjusted particle is {0.1, 0.2, 0.1, 0.3, 0.4, 0.4, 0.1}, indicating that the present invention can retain the segment with high fitness to enter the next round, thus guiding the particle to continuously approach the optimal value.

[0085] Finally, particle validity correction is performed. Since particle segments are all adjusted based on previous legal particles, constraints C1, C2, and C4 are all satisfied. The main task is to solve the node resource allocation conflict problem of C3. The specific correction process is as follows: (1) Traverse each node n in the set k in turn. If it is found that the resources required for the task calculation of n k exceed its available resources, then retain the effective task subset with the maximum fitness from the original task allocation, and adjust the position encoding corresponding to the remaining tasks assigned to n k to null. (2) Repeat the above process until the task resource allocation of all nodes no longer conflicts. (3) Reuse the remaining resources. Traverse the unassigned tasks in turn to see if a node with remaining resources and meeting the constraint conditions can be found, and then assign the task to this node. Repeat the above process until no further allocation is possible.

[0086] After obtaining the offloading decision, upload the task-related data to the corresponding computing node, execute the task offloading service, and return the result to the vehicle node after the calculation is completed. Through vehicle-edge cloud collaborative computing, the task processing efficiency is improved.

[0087] The beneficial effects of the method of the present invention are introduced below, and the experimental comparison data of the method of the present invention and the comparative method are provided.

[0088] The experimental scenario of the present invention is set as follows: An experimental scenario consisting of 200 intelligent vehicles, 50 edge servers, and 1 cloud server is built, and they are distributed within 25 kilometers centered at the longitude and latitude (116.41667, 39.91667). In order to more realistically simulate the communication and cooperation process between vehicles and between vehicles and servers, the present invention uses the dataset "parking places of Beijing" to construct a compact distribution scenario of vehicle nodes and tasks, uses the dataset "T-drive" to construct a compact distribution scenario of edge servers, and manually creates a uniform distribution scenario to avoid affecting the conclusion due to scenario characteristics. Other experimental parameters of the present invention are as follows: (1) Among the 200 vehicles, 10% are malicious nodes, which will initiate resource preemption behaviors irregularly. The available computing resources of each vehicle are 1 - 2 GHz, the task processing rate is 1 GHz / s, the transmission power is 0.2 W, and the computing power is 0.5 W; (2) Among the 50 edge servers, 10% are malicious nodes, which will initiate task discarding behaviors irregularly. The available computing resources of each edge server are 3 - 5 GHz, the task processing rate of the edge server is 2.6 GHz / s, and the task processing rate of the cloud server is 4.2 GHz / s; (3) Ordinary vehicle nodes randomly generate 1 - 3 tasks in each cycle, and resource preemption type nodes generate 3 - 6 tasks, among which the randomly generated false tasks are 5% - 10%. The CPU cycles required for each task are 0.5 - 1 GHz, the data volume is 0.5 - 1.5 MB, the maximum tolerable delay is 0.5 - 2 s, and the task revenue is 2 - 5 currency units. (4) In the trust evaluation of tasks and nodes, the information of the last 5 cycles of the nodes is always retained for comprehensive trust evaluation. (5) When making offloading decisions, the number of particles in each round of iteration is set to 10, and the maximum number of iterations is 50.

[0089] To objectively evaluate the effectiveness of the proposed method, the present invention selects three comparison methods. The first is the comparison method RG, which aims to minimize the task offloading cost and makes task offloading decisions based on random and greedy ideas. Under the condition of meeting the task allocation constraints, it selects the node with the minimum task execution cost among ε candidate nodes as the task computing node each time. The second is the comparison method TENG, which first identifies and filters the node trust, and then conducts task offloading through non-cooperative games. The third is the comparison method MOPS, which iteratively obtains the optimal offloading decision among multiple tasks and multiple nodes based on the traditional particle swarm optimization method.

[0090] Figure 5 and Figure 6Schematic diagram of the comparison of node and task identification rates between the method of the present invention and the comparative method. Combining the trust evaluation values of the task in two scenarios of compact and uniform distribution, it can be seen that compared with the comparative method TENG, the method of the present invention improves the edge node identification accuracy by 1.69% - 22.59%, improves the vehicle node identification rate in the uniform distribution scenario by 8.49% - 21.28%, and improves the task identification rate by 26.69% - 29.24%.

[0091] Figure 7 Schematic diagram of the comparison of energy consumption between the method of the present invention and the comparative method. Since the comparative method four MOPS algorithm is relatively complex, the energy consumption is significantly higher than that of other methods. In the previous few rounds, the energy consumption of the method of the present invention is not much different from that of other methods. As the network runs, the energy consumption of the method of the present invention gradually decreases. This is because the method of the present invention uses a trust evaluation mechanism to evaluate the node trust, filters out malicious nodes and false tasks, and avoids additional energy consumption and waste. Compared with the comparative method RG, the method of the present invention reduces the energy consumption by 1.30% - 3.84%. Compared with the comparative method MOPS, the method of the present invention reduces the energy consumption by about 15%.

[0092] Figure 8 Schematic diagram of the comparison of latency between the method of the present invention and the comparative method. The method of the present invention fully considers the latency characteristic constraints of the task when formulating the offloading decision, and introduces random quantities, optimization functions, etc. in the process of particle swarm optimization to solve the offloading strategy, avoiding the particles falling into local optima. The experimental results also prove that the final decision-making performance of the method of the present invention is relatively optimized. Generally speaking, the method of the present invention has obvious latency advantages. Compared with the three comparative methods, the average task latency is reduced by 6 - 21 ms.

[0093] Figure 9 Schematic diagram of the comparison of task benefits between the method of the present invention and the comparative method. The benefit here refers to the profit generated by the successfully executed tasks. Since the method of the present invention filters out false tasks and malicious nodes based on task and node trust evaluation, the task execution success rate is higher, so the task benefits generated are also significantly higher than those of other methods. Generally speaking, in scenario 1 where tasks and nodes are compactly distributed, the method of the present invention improves the benefit by 3.68% - 21.07%; in scenario 2 with uniform distribution, the method of the present invention improves the task benefit by 10.03% - 19.12%.

[0094] Figure 10 Schematic diagram of the comparison of task completion rates between the method of the present invention and the comparative method. The task completion rate refers to the proportion of successfully executed tasks in the total number of tasks. Since the method of the present invention randomly generates 5% - 10% malicious tasks, it is very high that the method of the present invention can achieve a task success rate of 89%. Compared with the comparative method, the task success rate is increased by about 5%.

Claims

1. A trusted task offloading method based on vehicle-side cloud autonomous collaboration, characterized in that: By analyzing data sensitivity, resource preemption, historical unloading success rate, and collaborative feedback evaluation, false tasks and malicious nodes can be accurately identified, and discrete particle swarm algorithm can be used to iteratively obtain vehicle-side cloud collaborative unloading decisions, effectively improving unloading efficiency and security. The main steps include: Step 1: Evaluate the authenticity of pending tasks generated in the current cycle, based on the data sensitivity of the task itself and the frequency of unloading requests from the source node where the task is generated. If the authenticity of the task is lower than the set threshold, Then the task is judged to have an attack risk and removed from the task set The remaining real tasks enter the offloading decision stage; Step 2: Conduct trust evaluation on vehicles and edge nodes involved in task calculations. The first step is based on the node's historical task processing success rate, and the second step is based on the feedback evaluation of collaborative interaction nodes. The node's comprehensive credibility is obtained by integrating the node's behavioral trends and time decay characteristics over multiple observation periods. If the node's credibility is lower than the set threshold, Then the node is determined to be a malicious node and removed from the node set The remaining trusted nodes enter the uninstallation decision phase; Step 3: Taking real tasks and trusted nodes as the objects, improve and redesign the population initialization, fitness evaluation, speed and position update, and validity correction steps of the discrete particle swarm algorithm, iteratively obtain the best task offloading decision, execute the offloading service, and reversely update the node trust evaluation model based on the task completion quality and node collaboration.

2. The trusted task offloading method based on vehicle-side cloud autonomous collaboration according to claim 1 is characterized in that: The specific operation of step 1 is: for any task T i , based on its task sensitivity and source node resource preemption, a trust evaluation function is constructed, and then the task authenticity evaluation is performed based on the trust evaluation function, which is specifically expressed as: Among them, α and β are weight adjustment factors, j is the number of generated tasks T i The source node N j The subscript of z is used to traverse K tasks, and x is used to traverse all nodes. Indicates the total number of nodes; and It is task T i The amount of sensitive data and total data involved, K is the total number of tasks in the task generation cycle; is the task source node N in the kth observation period j The number of task processing requests initiated; then, based on the task trust Perform classification processing. If the trust level is lower than the threshold Then remove it from the task set If the trust level is higher than the threshold, Keep it in the collection and make subsequent uninstall decisions.

3. The trusted task offloading method based on vehicle-side cloud autonomous collaboration according to claim 1 is characterized in that: The specific operation of step 2 is: for any node N i , analyze its historical task success rate and interactive collaboration reputation, introduce the weight adjustment factor ω, and construct its trust evaluation function in the kth observation period, which is specifically expressed as: in, and is node N i The number of successful and failed tasks processed during the k-th observation period, is the information provided by the jth interactor about node N i Feedback evaluation, R j is the reliability of the jth interactor, A is the total number of interacting nodes that provide feedback evaluation; each cycle is divided into M trust observation periods, and considering time decay, the trust evaluation of nodes in M ​​observation periods is integrated to obtain the node N i The comprehensive trust in the kth observation period is: in, is the decay function of the node trust memory retention with respect to time t, η1 = 1.84, η2 = 1.25, t represents the difference between the corresponding trust relationship and the current time, in minutes; then, classification is performed based on the node trust. If the node credibility is lower than the trust threshold Remove it from the node collection Delete; if the node credibility is higher than the threshold It is then allowed to enter the subsequent offloading decision phase and serve as a candidate node for task offloading and calculation.

4. The trusted task offloading method based on vehicle-side cloud autonomous collaboration according to claim 1 is characterized in that: The specific operation of step three is as follows: First, determine the particle position and velocity code, fix the code length to n, n is the number of tasks, and the position code value is null or n. i , represents the task processing node, and each bit of the particle speed encoding takes a value of 0 or 1. 0 means adjusting the corresponding position encoding segment in the next iteration, and 1 means not adjusting the position segment in the next iteration. Then, the local priority-random adjustment algorithm is used to generate the initial particle population. Next, the fitness function of particle evaluation is constructed based on the weighted and optimized objective of minimizing energy consumption and delay, which is specifically expressed as: in, It is task T i About Node n j If the processing decision This means that task T i Assign to n j , at this time the task processing cost is λ is the weight adjustment factor of energy consumption and delay, and the individual optimal particle and the global optimal particle are found based on the evaluated particle fitness; then, the random quantity and mapping function are introduced to update the velocity coding of the particle in the next iteration, and the position coding in the next iteration is updated based on the velocity coding and combined with the individual optimal particle and the global optimal particle; finally, the particle validity is corrected according to the node resource constraints and task delay constraints; the iteration is repeated until a converged offloading decision is obtained or the maximum number of iterations is reached.

Images