A method, system, device and medium for offloading safe computing of Internet of Vehicles assisted by intelligent reflective surface

Abstract

A method, system, device and medium for offloading safe computing of Internet of Vehicles assisted by intelligent reflective surfaces, the method comprising: constructing a RIS-assisted MEC vehicle network communication scenario; constructing a RIS-assisted safe communication scenario; constructing an optimization objective function of a RIS-assisted MEC vehicle network scenario; constructing a deep reinforcement learning algorithm model; constructing a deep reinforcement learning training model, setting the state, action and reward of the training model, and performing model training on the optimization target; a RIS-assisted MEC vehicle network decision model to obtain a solution for offloading safe computing of Internet of Vehicles; the system, device and medium are used to implement a method for offloading safe computing of Internet of Vehicles assisted by intelligent reflective surfaces; the present invention solves the problem of task offloading delay and safety in dynamic Internet of Vehicles scenarios by jointly designing a RIS phase shift matrix and allocating MEC computing resources in real time, meets the communication link security, improves the overall service quality of MEC, and ensures the service quality and safety performance of Internet of Vehicles.

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication technology, and specifically relates to a method, system, device and medium for vehicle network security computing offloading assisted by an intelligent reflective surface. Background Art

[0002] With the continuous innovation of 5G mobile communication technology, the emerging vehicle-to-everything (V2X) technology is becoming more and more mature, where V represents vehicle and X represents any object that interacts with the vehicle, which can be a vehicle, a person, a traffic facility, or a network. The widespread application of the vehicle-to-everything (V2X) technology has driven a large amount of data demand and delay-sensitive services, which require a large amount of computing resources to process. However, traditional cloud computing is not suitable for the emerging V2X technology because the distance between the target user and the server is long, which increases the computing latency. In order to solve the shortcomings of cloud computing, mobile edge computing (MEC) has become important as a new computing paradigm. MEC can be well combined with the vehicle-to-everything (V2X) technology, using the rich computing resources at the edge of the network to free resource-limited vehicle users from heavy computing tasks. By deploying MEC servers in the vehicle-to-everything (V2X) technology, multiple vehicles can simultaneously offload their tasks to the MEC servers and obtain high-speed computing services, reducing the processing latency of tasks and improving the user experience. However, due to the severe channel fading in crowded urban environments, the task offloading rate may be low, which prolongs the offloading delay. In addition, due to the broadcast characteristics of wireless signals, wireless links are vulnerable to security threats such as eavesdropping. Therefore, it is crucial to improve the service quality and data security of MEC vehicle networks from the perspective of secure communication.

[0003] Currently, smart reflective surfaces (RIS) are considered as a promising technology to improve the quality and coverage of wireless transmission. By designing the elements of smart reflective surfaces, signal reflections are designed to enhance the power of the desired signal while mitigating multi-user interference. Previous studies have shown that physical layer security (PLS) can be an effective alternative or complementary solution to protect the security of complex wireless networks by exploiting the inherent randomness in wireless channels. However, many PLS technologies will be severely degraded when the eavesdropper is closer to the base station (BS) than the legitimate user, or when the legitimate user and the eavesdropper have correlated channels. In response to these severe challenges, RIS combined with PLS brings hope for designing a robust secure transmission mechanism because it can flexibly reconstruct the channel environment in real time. Therefore, the technology of jointly studying RIS and MEC to achieve security services is proposed. However, the solution of RIS and MEC joint research has high complexity, and it is impossible to use mathematical methods to infer a low-complexity optimal solution. Deep reinforcement learning, as a powerful state estimation and function approximation tool, can adapt to various dynamic networks and solve complex optimization problems. Based on this, it is proposed to use deep reinforcement learning algorithms to optimize RIS and MEC resource allocation to achieve optimal security services.

[0004] In the literature [Y.Liu, W.Wang, H.-H.Chen, F.Lyu, L.Wang, W.Meng, and X.Shen, "Physical Layer Security Assisted Computation Offloading in Intelligently Connected Vehicle Networks," IEEE Transactions on Wireless Communications, vol.20, no.6, pp.3555–3570, 2021.], the authors proposed a secure computing offloading scheme in vehicle networks, focusing on optimizing the security MEC service delay of the target vehicle, in which artificial noise is added to resist potential eavesdroppers and achieve secure communication in the vehicle network. However, this scheme optimizes the security mobile edge computing service delay problem of the target vehicle in a static vehicle network scenario and cannot be applied to a dynamic vehicle network scenario with heavy computing tasks.

[0005] In the literature [Y.Ju, Y.Chen, Z.Cao, H.Wang, L.Liu, Q.Pei, and N.Kumar, "Learning Based and Physical-layer Assisted Secure Computation Offloading in Vehicular Spectrum Sharing Networks," in IEEE INFOCOM 2022-IEEE Conference on Computer Communications Workshops (INFOCOMWKSHPS), 2022.], the authors proposed a solution based on deep reinforcement learning to implement secure MEC services in a dynamic Internet of Vehicles scenario. However, this solution implements security services through physical layer security technology, which has limitations and does not explore the potential benefits of smart reflective surfaces.

[0006] In summary, the prior art still has the following shortcomings:

[0007] (1) The existing technology is to optimize the security mobile edge computing service delay problem of the target vehicle in a static vehicle network scenario, which is not suitable for dynamic mobile edge computing vehicle networks with heavy computing tasks.

[0008] (2) In the existing technology, the base station will allocate MEC computing resources to the target vehicle only when all target vehicles have completed the task transmission to the base station equipped with a mobile edge computing server, which greatly increases the delay of the Internet of Vehicles service.

[0009] (3) In the dynamic Internet of Vehicles scenario, the existing technology does not consider the potential benefits of smart reflective surfaces when studying the delay problem of mobile edge computing security services.

[0010] In a real-time changing channel, how to choose a suitable deep reinforcement learning algorithm to cope with the high-dimensional state space; how to use deep reinforcement learning to optimize RIS and MEC are the key issues that RIS-assisted MEC security service technology needs to solve. Summary of the invention

[0011] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a method, system, device and medium for vehicle network security computing offloading assisted by an intelligent reflective surface, optimize MEC services based on a communication scheme based on a deep deterministic policy gradient algorithm, minimize the maximum MEC service time by jointly designing a RIS phase shift matrix and allocating MEC computing resources in real time, so as to achieve optimal MEC security services, solve the task offloading delay and security issues in dynamic vehicle network scenarios, and improve the overall service quality of MEC on the premise of meeting the security of the communication link, so that the service quality and safety performance of the vehicle network are guaranteed.

[0012] In order to achieve the above object, the technical solution adopted by the present invention is:

[0013] A method for offloading secure computing in an Internet of Vehicles (IoV) assisted by an intelligent reflective surface comprises the following steps:

[0014] Step 1: Construct a RIS-assisted MEC vehicle network communication scenario and add an eavesdropper model;

[0015] Step 2: Build a RIS-assisted secure communication scenario;

[0016] Step 3: Model the optimization target of the RIS-assisted MEC vehicle network scenario constructed in step 1 and construct the objective function for model solving;

[0017] Step 4: Construct a deep reinforcement learning algorithm model based on the optimization goal proposed in step 3;

[0018] Step 5: Construct a deep reinforcement learning training model based on the deep reinforcement learning algorithm model proposed in step 4. Combined with the communication scenarios and objective functions in steps 1, 2, and 3, set the state, action, and reward of the training model, and perform model training on the optimization target of the RIS-assisted MEC vehicle network communication scenario.

[0019] Step 6: Based on the training model in step 5, the RIS-assisted MEC vehicle network decision model is obtained to obtain the optimal solution to the optimization problem, that is, the vehicle network security computing offloading solution is obtained.

[0020] The specific method of step 1 is:

[0021] The BS simultaneously establishes multiple communication links with vehicle users in different orthogonal sub-bands. The target vehicle with limited resources can offload its computing tasks to the BS equipped with MEC server, thereby obtaining MEC computing resources. The target vehicle that obtains computing services is expressed as:

[0022]

[0023] Among them, User M represents the Mth target vehicle user;

[0024] Vehicles that are not being served are considered potential eavesdroppers and can be represented as:

[0025] E={Eve1,Eve2,…,Eve E}

[0026] Among them, Eve E represents the Eth potential eavesdropper.

[0027] The specific method of step 2 is:

[0028] Step 2.1: Assume that the reflection coefficient of the nth element of RIS is expressed as: Among them, φ n ∈[0,2π), the RIS reflection coefficient matrix is ​​defined as:

[0029] θ=diag([θ1,θ2,...,θ N ])

[0030] In the absence of in-band interference, the receive beamforming is designed by the maximum ratio combining technique, which can be expressed as:

[0031]

[0032] Among them, f M represents the beamforming vector of the Mth V2I link;

[0033] Step 2.2: Communication channel modeling;

[0034] In the MEC vehicle network, the channel includes: the mth V2I link The link between the mth target vehicle and the RIS The link from the mth target vehicle to the eth potential eavesdropper The link between RIS and the e-th potential eavesdropper RIS to BS link The channel from RIS to BS follows the Rician distribution, which can be expressed as:

[0035]

[0036] Among them, κ i,b is the Rician factor, ρ is the path loss at the reference distance d0 = 1m, d i,b is the distance between RIS and BS, α i,b is the path loss index of the RIS to BS link, non-LOS component Each element of follows a complex Gaussian distribution with zero mean and unit variance, and similarly h m,e ,h m,b ,h m,i ,h i,e Following the Rician distribution, due to the crowded urban environment and the blocking effect between vehicles, κ m,b and κ m,e All are zero;

[0037] Step 2.3: Modeling of signal reception process;

[0038] The mth V2I link received signal at the BS can be expressed as:

[0039]

[0040] Among them, P m is the transmission power of the mth target vehicle, s m represents the unit energy signal sample associated with the computational task, and the noise vector n m It can be expressed as:

[0041] n m =[n1,...n K ] T

[0042] in,

[0043] The uplink signal-to-interference-plus-noise ratio (SINR) of the mth V2I link at the BS is given by:

[0044]

[0045] Similarly, the eavesdropping signal of the mth V2I link at the eth eavesdropping vehicle is expressed as:

[0046]

[0047] in,

[0048] The SINR of the mth V2I link at the eth eavesdropping vehicle can be expressed as:

[0049]

[0050] Therefore, the capacity of the mth V2I link and the eavesdropping capacity from the e-th eavesdropping vehicle to the mth V2I link can be expressed as:

[0051] C m =log(1+η m )

[0052] C e,m =log(1+η e,m )

[0053] In the MEC vehicle network, once the user completes the offloading process, the BS flexibly allocates the computing resources of the MEC server according to the size of the task. Each CPU cycle of the MEC server can process a certain number of data bits, assuming that the total computing power is ζ bit / s.

[0054] The specific method of step 3 is:

[0055] Step 3.1: Security process modeling;

[0056] Any unserved vehicle can eavesdrop on any V2I link. In order to protect mission data from being eavesdropped, the redundancy used to protect confidential information can be expressed as:

[0057] max{0,R b -R S}

[0058] Among them, R b is the code rate, R S target confidentiality rates for confidential information;

[0059] If the capacity of the eavesdropper is C e Greater than R b -R S , a confidentiality interrupt will be sent, using capacity C b Approximate R b , so the confidentiality transmission rate of the mth V2I link can be expressed as:

[0060] R S,m =[0,(C m -maxC e,m )] + ,e∈ε

[0061] Where [x] +=max{0,x};

[0062] The MEC service time (offloading and computation time) of the mth V2I link can be expressed as:

[0063]

[0064] Among them, S m is the task size, m is the allocated computing resource;

[0065] Step 3.2: Optimization target modeling;

[0066] The optimization goal is to design the RIS reflection coefficient matrix Θ and MEC resource allocation for different computing tasks. To minimize the service time, the former will affect the transmission time, while the latter will determine the calculation time. Considering that the entire MEC service cycle is determined by the maximum service time of all V2I links, the above goal is transformed into the following minimum-maximum problem:

[0067]

[0068]

[0069]

[0070] Among them, constraint C1 means that the sum of computing resources allocated to different target vehicles is a fixed value, and constraint C2 means that the modulus of the RIS reflection coefficient is constrained to be unit modulus.

[0071] The specific method of step 4 is:

[0072] DDPG is a model-free, off-policy Actor-Critic architecture algorithm. The Actor network is used to predict actions, and the Critic network is used to evaluate the future benefits of taking the action in the current state. Both the Actor network and the Critic network are composed of two deep neural network DNN networks: a training network and a target network. The training and target network parameters of the Actor network are θ and θ, respectively. a and θ a′ , the training and target network parameters of the Critic network are θ c and θ c′ ;

[0073] At time slot t, the Actor training network will S t As input, and output action a t , the Critic training network will S t and a tTakes as input and outputs the state-action function value state-action value Q π (S t ,a t ∣θ c ), which can be expressed as:

[0074] Q π (S t ,a t ∣θ c )=E π [R t ∣S t ,a t ,π]

[0075] Where E[·] represents the expected function, π represents the strategy of the Actor training network, and when enough quadruplets (S t ,a t ,r t ,S t+1 ), the model optimizer randomly draws N from the experience replay pool d The samples are used to update the training network of Actor and Critic. The k-th tuple y k The target state-action function value Q′ can be expressed as:

[0076] y k =r k +γQ′ π′ (S k+1 ,π′(S k+1 ∣θ a′ )|θ c′ )

[0077] Among them, π′ represents the strategy of the Actor target network;

[0078] The Critic training network uses the mean square error MSE function to update the network, which can be expressed as follows:

[0079]

[0080]

[0081] The Actor training network uses a deterministic policy gradient function to update the network, which can be expressed as:

[0082]

[0083]

[0084] The updates of the Actor and Critic target networks are as follows:

[0085] ″

[0086] θ c′ =τ c θ c +(1-τ c )θ c′

[0087] θ a′ =τ a θ a +(1-τ a )θ a′

[0088] Among them, τ c and τ a is the soft update coefficient, which satisfies τ c ,τ a ∈[0,1].

[0089] The specific method of step 5 is:

[0090] Step 5.1: State space setup;

[0091] At time slot t, the state of the mth V2I link Including confidentiality rate Remaining unloading tasks Remaining computing tasks Amount of MEC resources occupied Global channel state information It can be expressed as:

[0092]

[0093] In summary, the state of the mth V2I link is expressed as:

[0094]

[0095] At time slot t, the total environment of M V2I links can be expressed as:

[0096]

[0097] Step 5.2: Action space setting;

[0098] Based on the current state S t , BS will design the RIS phase shift matrix and MEC resource allocation, in each time slot t, the action space can be expressed as:

[0099] a t ={Θ t ,ζ t}

[0100] in, It is the allocation of computing resources;

[0101] Step 5.3: Reward function setting;

[0102] At time slot t, corresponding to the current action a t The reward can be expressed as:

[0103]

[0104] in, represents the secure MEC service time of the mth V2I link at time slot t, t m,1 is the current consumed time, t m,2 It is the remaining time estimated based on the current action, which includes the remaining transmission time and the remaining calculation time. There are three cases for estimating the remaining time:

[0105] (1) All target vehicles are in the process of task offloading. The remaining transmission time of each target vehicle is based on the current action, and the remaining computation time of each target vehicle is calculated by distributing the computation resources evenly to all target vehicles, that is, ζ min ;

[0106] (2) Some target vehicles are in the process of task offloading, and other target vehicles are in the process of task calculation. For the target vehicles in the process of task offloading, the remaining transmission time of each user in the process of task offloading is calculated based on the current action, and the remaining transmission time of each user in the process of task offloading is calculated based on the computing resources. The strategy estimates the remaining computation time, where ζ min It is the minimum computing resource of the target vehicle during the task calculation process. For the target vehicle during the task calculation process, it only needs to estimate the remaining computing time based on the current action;

[0107] (3) All target vehicles are in the task calculation process, and the remaining calculation time of all target vehicles is estimated based on the current action;

[0108] In order to improve the confidentiality transmission rate, the penalty factor is expressed as:

[0109]

[0110] If the current action can meet the confidentiality rate requirement of the mth link Then ν m = 0, otherwise ν m =v * , ν * It is a parameter that can be set manually, which is a negative number;

[0111] Based on the setting of the reward function, the DDPG algorithm will continuously learn the action strategy to reduce the maximum safe MEC service time within the given constraints. The total cumulative reward can be expressed as:

[0112]

[0113] Where γ is the discount factor.

[0114] The specific method of step 6 is:

[0115] Step 6.1: Initialization;

[0116] Randomly initialize the parameters θ of the Actor and Critic training networks a ,θ c , the parameters θ of the Actor target network a′ Initialized to θ a , the parameters θ of the Critic target network c′ Initialized to θ c , clear the experience replay pool D;

[0117] Step 6.2: Training phase;

[0118] Randomly initialize the positions of the target vehicle and the eavesdropped vehicle, and initialize the task volume of the target vehicle's requested service;

[0119] At each time slot t, BS interacts with the dynamic environment and obtains the state S t Based on the current state, BS obtains action a from the Actor network of the Mth V2I link t , thereby setting the reflection coefficient matrix and MEC resource allocation for the target vehicle;

[0120] BS obtains the state S of the next time slot t+1 from the changing environment t+1 , and calculate the t The reward r obtained from the environment t ;

[0121] The states, actions, and rewards in the above process are stored as tuples (S t ,a t ,r t ,S t+1 ), and stores the tuple in the experience replay pool D, and obtains the state-action function Qπ(S t ,a t ∣θ c );

[0122] When there are enough tuples in the experience replay pool, N tuples are taken from it. dThe parameters of the Critic and Actor networks are updated with samples of different sizes. When the task loads of all target vehicles have been calculated, the model training ends and the above process is repeated until the model training converges.

[0123] Step 6.3: Decision-making phase

[0124] The decision model that has been trained and converged is used in a random dynamic vehicle network scenario. The optimal RIS reflection coefficient matrix and MEC resource allocation are determined in each time slot, so that the maximum MEC service time in the whole process is minimized, and finally the optimal solution of the optimization objective is obtained.

[0125] The present invention also provides a system for implementing the intelligent reflective surface-assisted vehicle network security computing offloading method, comprising:

[0126] RIS-assisted MEC vehicle network communication module: used to realize the construction of RIS-assisted MEC vehicle network communication scenarios. In this module, the base station establishes multiple communication links with dynamic vehicles;

[0127] RIS-assisted safety communication module: used to realize the construction of RIS-assisted safety communication scenarios. In this module, RIS technology provides security for dynamic vehicle communication;

[0128] Secure computing service optimization target module: used to realize the construction of optimization targets for RIS-assisted MEC vehicle network scenarios;

[0129] Deep reinforcement learning algorithm selection module: used to build a deep reinforcement learning algorithm model based on optimization objectives;

[0130] Deep reinforcement learning model training module: used to build a deep reinforcement learning training model. In this module, the model training is performed on the optimization target of the RIS-assisted MEC vehicle network scenario.

[0131] Deep reinforcement learning decision model module: used to implement the RIS-assisted MEC vehicle network decision model. In this module, the optimal RIS coefficient matrix and MEC resource allocation in the dynamic Internet of Vehicles scenario are obtained.

[0132] The present invention also provides a vehicle network security computing offloading device assisted by an intelligent reflective surface, comprising:

[0133] Memory for storing computer programs;

[0134] A processor is used to implement the intelligent reflective surface-assisted vehicle network security computing offloading method when executing the computer program.

[0135] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it is capable of performing computational offloading on a vehicle network safety assisted by an intelligent reflective surface.

[0136] Compared with the prior art, the present invention has the following beneficial effects:

[0137] 1. Currently, there is no research on using deep reinforcement learning algorithms to optimize the reflection coefficient matrix of intelligent reflective surfaces and the allocation of mobile edge computing resources in dynamic scenarios; the solution provided by the present invention can decide multiple continuous optimal actions in a high-dimensional continuous state space, reduce vehicle network service delays, and provide protection for communication security.

[0138] 2. The present invention regards the base station as an intelligent entity, which can make decisions based on the ever-changing state of the surroundings and has high adaptability to the highly dynamic Internet of Vehicles scenarios. At the same time, as long as a target vehicle completes the task unloading, the base station will allocate computing resources to it, which makes effective use of idle MEC resources.

[0139] 3. At present, security issues in the Internet of Vehicles scenario are all solved based on physical layer security technology, which has limitations. The intelligent reflective surface technology provided by the present invention is combined with physical layer security technology to realize a solution for security services, which solves the problem that the physical layer security technology cannot resist the eavesdropping user when the eavesdropping user is closer to the base station than the target user and the eavesdropping user and the target user have related channels.

[0140] 4. The RIS-assisted MEC vehicle network security communication scenario provided in steps 1 and 2 of the present invention can be associated with the actual dynamic Internet of Vehicles security communication scenario, providing a solution to the security service problem in the actual scenario, and has the advantage of strong applicability.

[0141] 5. The deep reinforcement learning algorithm provided in step 4 of the present invention can solve complex high-dimensional continuous state space problems and output continuous action values ​​according to the continuous state space. It has the advantages of adapting to dynamic scenes and solving non-convex problems.

[0142] In summary, compared with the prior art, the present invention has the advantages of using deep reinforcement learning algorithms to jointly optimize intelligent reflective surfaces and mobile edge computing in dynamic scenarios to achieve security services and reduce service delays. BRIEF DESCRIPTION OF THE DRAWINGS

[0143] Figure 1 It is a flow chart of the present invention.

[0144] Figure 2 This is a schematic diagram of a smart reflective surface-assisted mobile edge computing scenario provided by an embodiment of the present invention.

[0145] Figure 3 It is a deep reinforcement learning training model architecture diagram provided by an embodiment of the present invention.

[0146] Figure 4 This is a simulation result diagram of the average MEC service time, MEC successful service probability, and average MEC service confidentiality interruption probability, provided by an example of the present invention, by comparing the DDPG algorithm with other algorithms under different eavesdropping levels.

[0147] Figure 5 It is a simulation result diagram of the average MEC service time and the MEC successful service probability provided by an example of the present invention, in which the DDPG algorithm is compared with other algorithms and analyzed when the target vehicle is in different mission scopes. DETAILED DESCRIPTION

[0148] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0149] The present invention provides a method, system, device and medium for offloading safe computing of Internet of Vehicles assisted by intelligent reflective surface. The scheme first models the RIS-assisted MEC vehicle network scenario, in which the base station simultaneously establishes multiple communication links with vehicle users in different sub-bands to realize high-speed data rate transmission service. In the MEC scenario, the resource-constrained target vehicle offloads its computing task to the base station (BS) equipped with MEC server through the vehicle-to-base station (V2I) link. The BS flexibly allocates MEC resources for different task requests, and then feeds back the results to the target user. The RIS-assisted safe communication is modeled, and the communication channel obeys the Rician distribution. All vehicles are equipped with a single omnidirectional antenna, and the BS has a uniform linear array of K antennas. The intelligent reflective surface is a diagonal matrix with N reflective elements. Since there is no in-band interference, the BS uses the maximum ratio combining (MRC) method to design beamforming for each V2I link. Secondly, in order to realize the safety service in the MEC scenario, an optimization problem of minimizing the maximum MEC service time by jointly designing the RIS reflection coefficient matrix and MEC resource allocation is proposed. This optimization problem is non-convex and is a long-term decision-making process with high dynamics. Therefore, a deep reinforcement learning algorithm is used to solve it and achieve the optimal MEC service. By designing the state, action, and reward of the deep reinforcement learning algorithm, the location information of the dynamic vehicle and the task volume and other parameters are used as the basis for the agent's decision-making. Finally, the optimal RIS reflection coefficient matrix and MEC resource allocation are trained to achieve safe and low-latency MEC services.

[0150] like Figure 1 As shown, a flow chart of the intelligent reflective surface-assisted Internet of Vehicles safety computing offloading solution based on deep reinforcement learning.

[0151] A method for offloading safe computing of Internet of Vehicles assisted by intelligent reflective surface, the steps of which are:

[0152] Step 1: Construct a RIS-assisted MEC vehicle network communication scenario to provide services for vehicles that send computing service requests, and add an eavesdropper model for subsequent modeling and analysis; further, the specific method of step 1 is:

[0153] like Figure 2 As shown in the figure, it is a scenario of intelligent reflective surface assisted mobile edge computing. The BS simultaneously establishes multiple communication links with vehicle users in different orthogonal sub-bands. Resource-constrained vehicles can offload their computing tasks to the BS equipped with MEC servers. The BS flexibly allocates MEC resources for different task requests and then feeds back the results to the vehicle users. In the present invention, it is assumed that the feedback delay time is negligible relative to the time required to meet the computing task. Due to limited resources at the BS, only vehicles that send computing service requests can be provided. The target vehicle that obtains the computing service is represented as:

[0154]

[0155] Among them, User M represents the Mth target vehicle user.

[0156] Vehicles that are not being served are considered potential eavesdroppers and can be represented as:

[0157] ε={Eve1,Eve2,…,Eve E}

[0158] Among them, Eve E represents the Eth potential eavesdropper.

[0159] Step 2: Construct a RIS-assisted secure communication scenario to lay the foundation for the communication channel used in the subsequent use of the present invention;

[0160] Further, the specific method of step 2 is:

[0161] Step 2.1: Assume that the reflection coefficient of the nth element of RIS is expressed as: where φ n ∈[0,2π), the RIS reflection coefficient matrix is ​​defined as:

[0162] θ=diag([θ1,θ2,...,θ N ])

[0163] Since there is no in-band interference, the receive beamforming is designed by the maximum ratio combining technique, which can be expressed as:

[0164]

[0165] Among them, f M represents the beamforming vector of the Mth V2I link.

[0166] Step 2.2: Communication channel modeling;

[0167] In the MEC vehicle network, the channel includes: the mth V2I link The link between the mth target vehicle and the RIS The link from the mth target vehicle to the eth potential eavesdropper The link between RIS and the e-th potential eavesdropper RIS to BS link The channel from RIS to BS follows the Rician distribution, which can be expressed as:

[0168]

[0169] Among them, κ i,b is the Rician factor, ρ is the path loss at the reference distance d0 = 1m, d i,b is the distance between RIS and BS, α i,b is the path loss index of the RIS to BS link. Non-LOS component Each element of follows a complex Gaussian distribution with zero mean and unit variance. Similarly, h m,e ,h m,b ,h m,i ,h i,e follows a Rician distribution. Due to the crowded urban environment and the blocking effect between vehicles, κ m,b and κ m,e All are zero.

[0170] Step 2.3: Modeling of signal reception process;

[0171] The mth V2I link received signal at the BS can be expressed as:

[0172]

[0173] Among them, P m is the transmission power of the mth target vehicle, s m represents the unit energy signal sample associated with the computational task, and the noise vector n m It can be expressed as:

[0174] n m =[n1,...n K ] T

[0175] in,

[0176] The uplink signal-to-interference-plus-noise ratio (SINR) of the mth V2I link at the BS is given by:

[0177]

[0178] Similarly, the eavesdropping signal of the mth V2I link at the eth eavesdropping vehicle is expressed as:

[0179]

[0180] in,

[0181] The SINR of the mth V2I link at the eth eavesdropping vehicle can be expressed as:

[0182]

[0183] Therefore, the capacity of the mth V2I link and the eavesdropping capacity from the e-th eavesdropping vehicle to the mth V2I link can be expressed as:

[0184] C m =log(1+η m )

[0185] C e,m =log(1+η e,m )

[0186] In MEC vehicle networks, once a user completes the offloading process, the BS flexibly allocates the computing resources of the MEC server according to the size of the task. Each CPU cycle of the MEC server can process a certain number of data bits, assuming the total computing power is ζ bit / s. In order to provide stable services, the BS aims to minimize the time of the entire MEC service while ensuring the safety of task offloading for all users.

[0187] Step 3: Model the optimization target of the RIS-assisted MEC vehicle network scenario constructed in step 1, and construct the objective function when solving the model, so as to lay the foundation for the subsequent use of deep reinforcement learning for model solving in the present invention;

[0188] Further, the specific method of step 3 is:

[0189] Step 3.1: Security process modeling;

[0190] The present invention considers the worst security threat scenario, in which any unserved vehicle can eavesdrop on any V2I link. In order to protect the mission data from being eavesdropped, the transmitter encodes the data and then needs to determine two code rates before transmission, namely, the code rate R and the code rate R. band the target confidentiality rate R of confidential information S . Therefore, the redundancy used to protect confidential information can be expressed as:

[0191] max{0,R b -R S}

[0192] Among them, R b is the code rate, R S Target confidentiality rate for confidential information.

[0193] If the capacity of the eavesdropper is C e Greater than R b -R S , a confidentiality interrupt will be sent. In the present invention, we use capacity C b Approximate R b Therefore, the confidentiality transmission rate of the mth V2I link can be expressed as:

[0194] R S,m =[0,(C m -maxC e,m )] + ,e∈ε

[0195] Where [x] + =max{0,x}.

[0196] The MEC service time (offloading and computation time) of the mth V2I link can be expressed as:

[0197]

[0198] Among them, S m is the task size, m is the allocated computing resource.

[0199] Step 3.2: Optimize target modeling;

[0200] The optimization goal of the present invention is to design the RIS reflection coefficient matrix θ and MEC resource allocation for different computing tasks. to minimize the service time. The former will affect the transmission time, while the latter will determine the computation time. Considering that the entire MEC service cycle is determined by the maximum service time of all V2I links, we transform the above objectives into the following min-max problem:

[0201]

[0202]

[0203]

[0204] Among them, constraint C1 means that the sum of computing resources allocated to different target vehicles is a fixed value, and constraint C2 means that the modulus of the RIS reflection coefficient is constrained to be unit modulus.

[0205] Step 4: Construct a deep reinforcement learning algorithm model based on the optimization goal proposed in step 3, lay a theoretical foundation for the practical problems to be solved, and reduce the difficulty of solving the optimization problem;

[0206] Further, the specific method of step 4 is:

[0207] The joint design of the RIS reflection coefficient matrix and MEC resource allocation of the entire MEC service can be modeled as a Markov decision process (MDP). The process consists of multiple time periods and their specific actions, each of which affects future benefits. The optimization problem of the present invention is non-convex and is a long-term decision-making problem with high dynamics, which is difficult to express with explicit mathematical expressions. Therefore, the present invention adopts a deep reinforcement learning (DRL) algorithm of deep deterministic policy gradient (DDPG). The algorithm can train appropriate parameters according to the continuous state space, so as to design the desired RIS coefficient matrix and MEC resource allocation to minimize the service time.

[0208] like Figure 3 As shown in Figure 1, DDPG is a model-free, off-policy Actor-Critic architecture algorithm. The Actor network is used to predict actions, and the Critic network is used to evaluate the future benefits of taking the action in the current state. Both the Actor network and the Critic network consist of two deep neural network (DNN) networks: a training network and a target network. The training and target network parameters of the Actor network are θ and θ, respectively. a and θ a′ , the training and target network parameters of the Critic network are θ c and θ c′ . DDPG deep reinforcement learning training model architecture.

[0209] At time slot t, the Actor training network will S t As input, and output action a t , the Critic training network will S t and a t Takes as input and outputs the state-action function value Q π (S t ,a t ∣θ c ), which can be expressed as:

[0210] Q π(S t ,a t ∣θ c )=E π [R t ∣S t ,a t ,π]

[0211] Where E[·] represents the expected function and π represents the strategy of the Actor training network. When enough quadruplets (S t ,a t ,r t ,S t+1 ), the model optimizer randomly draws N from the experience replay pool d The kth tuple y k The target state-action function value Q′ can be expressed as:

[0212] y k =r k +γQ′ π′ (S k+1 ,π′(S k+1 ∣θ a′ )|θ c′ )

[0213] Among them, π′ represents the strategy of the Actor target network.

[0214] The Critic training network uses the mean square error (MSE) function to update the network, which can be expressed as follows:

[0215]

[0216]

[0217] The Actor training network uses a deterministic policy gradient function to update the network, which can be expressed as:

[0218]

[0219]

[0220] The updates of the Actor and Critic target networks are as follows:

[0221] ″

[0222] θ c′ =τ c θ c +(1-τ c )θ c′

[0223] θ a′ =τ a θ a +(1-τ a )θ a′

[0224] Among them, τ c and τ a is the soft update coefficient, which satisfies τ c ,τ a ∈[0,1];

[0225] Step 5: Construct a deep reinforcement learning training model based on the deep reinforcement learning algorithm model proposed in step 4. Combined with the communication scenarios and objective functions in steps 1, 2, and 3, set the state, action, and reward of the training model, and perform model training on the optimization target of the RIS-assisted MEC vehicle network communication scenario, laying the foundation for the subsequent decision-making model.

[0226] Further, the specific method of step 5 is:

[0227] Step 5.1: State space setup;

[0228] At time slot t, the state of the mth V2I link Including confidentiality rate Remaining unloading tasks Remaining computing tasks Amount of MEC resources occupied Global channel state information It can be expressed as:

[0229]

[0230] In summary, the state of the mth V2I link is expressed as:

[0231]

[0232] At time slot t, the total environment of M V2I links can be expressed as:

[0233]

[0234] Step 5.2: Action space setting;

[0235] Based on the current state S t , BS will design the RIS phase shift matrix and MEC resource allocation, in each time slot t, the action space can be expressed as:

[0236] a t ={Θ t ,ζ t}

[0237] in, It is the allocation of computing resources;

[0238] Step 5.3: Reward function setup

[0239] At time slot t, corresponding to the current action a t The reward can be expressed as:

[0240]

[0241] in, represents the secure MEC service time of the mth V2I link at time slot t, t m,1 is the current consumed time, t m,2 It is the remaining time estimated based on the current action, which includes the remaining transmission time and the remaining calculation time. There are three cases for estimating the remaining time:

[0242] (1) All target vehicles are in the process of task offloading. The remaining transmission time of each target vehicle is based on the current action, and the remaining computation time of each target vehicle is calculated by evenly distributing the computation resources to all target vehicles, that is, ζ min .

[0243] (2) Some target vehicles are in the process of task offloading, and other target vehicles are in the process of task calculation. For target vehicles in the process of task offloading, the remaining transmission time of each user in the process of task offloading is calculated based on the current action, and the remaining transmission time of each user in the process of task offloading is calculated based on the computing resources. The strategy estimates the remaining computation time, where ζ min It is the minimum computing resource of the target vehicle during the task calculation process. For the target vehicle during the task calculation process, it only needs to estimate the remaining computing time based on the current action.

[0244] (3) All target vehicles are in the task calculation process. Estimate the remaining calculation time of all target vehicles based on the current action.

[0245] In order to improve the confidentiality transmission rate, the penalty factor is expressed as:

[0246]

[0247] If the current action can meet the confidentiality rate requirement of the mth link Then ν m = 0, otherwise ν m =v * , ν * It is a parameter that can be set manually and is a negative number.

[0248] Based on the setting of the reward function, the DDPG algorithm will continuously learn the action strategy to reduce the maximum safe MEC service time within the given constraints. The total cumulative reward can be expressed as:

[0249]

[0250] Among them, γ is the discount factor;

[0251] Step 6: According to the training model in step 5, the RIS-assisted MEC vehicle network decision model is obtained to obtain the optimal solution to the optimization problem, that is, the vehicle network security computing offloading solution is obtained;

[0252] Further, the specific method of step 6 is:

[0253] Step 6.1: Initialization;

[0254] Randomly initialize the parameters θ of the Actor and Critic training networks a ,θ c , the parameters θ of the Actor target network a′ Initialized to θ a , the parameters θ of the Critic target network c′ Initialized to θ c . Clear the experience replay pool D;

[0255] Step 6.2: Training phase;

[0256] Randomly initialize the positions of the target vehicle and the eavesdropped vehicle, and initialize the task volume of the target vehicle's requested service;

[0257] At each time slot t, BS interacts with the dynamic environment and obtains the state S t Based on the current state, BS obtains action a from the Actor network of the Mth V2I link t , thereby setting the reflection coefficient matrix and MEC resource allocation for the target vehicle;

[0258] BS obtains the state S of the next time slot t+1 from the changing environment t+1 , and calculate the t The reward r obtained from the environment t ;

[0259] The states, actions, and rewards in the above process are stored as tuples (S t ,a t ,r t ,S t+1 ), and stores the tuple in the experience replay pool D, and obtains the state-action function Qπ(S t ,a t∣θ c );

[0260] When there are enough tuples in the experience replay pool, N tuples are taken from it. d The parameters of the Critic and Actor networks are updated with samples of different sizes. When the task loads of all target vehicles have been calculated, the model training is completed. The above process is repeated until the model training converges;

[0261] Step 6.3: Decision-making stage;

[0262] The decision model that has been trained and converged is used in a random dynamic vehicle network scenario. The optimal RIS reflection coefficient matrix and MEC resource allocation are determined in each time slot, so that the maximum MEC service time in the whole process is minimized, and finally the optimal solution of the optimization objective is obtained.

[0263] like Figure 4 As shown in the figure, it shows the simulation results of the average MEC service time, MEC successful service probability, and average MEC service confidentiality interruption probability of the DDPG algorithm compared with other algorithms under different eavesdropping levels. It can be seen that the DDPG-based method significantly reduces the average maximum MEC service time, while improving the MEC success probability, achieving secure MEC services, and reducing service delays.

[0264] like Figure 5 As shown in the figure, it shows the simulation results of the average MEC service time and the MEC successful service probability of the DDPG algorithm compared with other algorithms when the target vehicle is in different mission ranges. From the comparative simulation analysis diagram, it can be seen that the DDPG-based deep reinforcement learning algorithm can solve high-dimensional non-convex problems well, can successfully learn effective strategies in complex and dynamic communication scenarios, and obtain the optimal RIS reflection coefficient and MEC resource allocation.

[0265] The present invention also provides a system for implementing the intelligent reflective surface-assisted vehicle network security computing offloading method, comprising:

[0266] RIS-assisted MEC vehicle network communication module: used to implement the construction of the RIS-assisted MEC vehicle network communication scenario in step 1. In this module, the base station establishes multiple communication links with the dynamic vehicle;

[0267] RIS-assisted safety communication module: used to implement the construction of the RIS-assisted safety communication scenario in step 2. In this module, RIS technology provides security for dynamic vehicle communication;

[0268] Secure computing service optimization target module: used to realize the construction of optimization targets for the RIS-assisted MEC vehicle network scenario in step 3;

[0269] Deep reinforcement learning algorithm selection module: used to implement the construction of a deep reinforcement learning algorithm model based on the optimization objective in step 4;

[0270] Deep reinforcement learning model training module: used to implement the construction of the deep reinforcement learning training model in step 5. In this module, model training is performed on the optimization target of the RIS-assisted MEC vehicle network scenario;

[0271] Deep reinforcement learning decision model module: used to implement the RIS-assisted MEC vehicle network decision model in step 6. In this module, the optimal RIS coefficient matrix and MEC resource allocation in the dynamic vehicle network scenario are obtained.

[0272] The present invention also provides a vehicle network security computing offloading device assisted by an intelligent reflective surface, comprising:

[0273] Memory for storing computer programs;

[0274] A processor is used to implement the intelligent reflective surface-assisted vehicle network security computing offloading method when executing the computer program.

[0275] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it is capable of performing computational offloading on a vehicle network safety assisted by an intelligent reflective surface.

Claims

1. A method for offloading safe calculations of Internet of Vehicles assisted by intelligent reflective surfaces, characterized in that: The following steps are involved: Step 1: Construct a RIS-assisted MEC vehicle network communication scenario and add an eavesdropper model; The BS simultaneously establishes multiple communication links with vehicle users in different orthogonal sub-bands. The resource-constrained target vehicle offloads its computing tasks to the BS equipped with the MEC server, thereby obtaining MEC computing resources. The target vehicle that obtains computing services is represented as: Among them, User M represents the Mth target vehicle user; Vehicles that are not being served are considered potential eavesdroppers, represented by: ε={Eve1,Eve2,…,Eve E } Among them, Eve E represents the Eth potential eavesdropper; Step 2: Build a RIS-assisted secure communication scenario; Step 2.1: Assume that the reflection coefficient of the nth element of RIS is expressed as: Among them, φ n ∈[0,2π), the RIS reflection coefficient matrix is ​​defined as: Θ=diag([θ1,θ2,...,θ N ]) In the absence of in-band interference, the receive beamforming is designed by the maximum ratio combining technique, which is expressed as: Among them, f M represents the beamforming vector of the Mth V2I link; Step 2.2: Communication channel modeling; In the MEC vehicle network, the channel includes: the mth V2I link The link between the mth target vehicle and the RIS The link from the mth target vehicle to the eth potential eavesdropper The link between RIS and the e-th potential eavesdropper RIS to BS link The channel from RIS to BS follows the Rician distribution, which can be expressed as: Among them, κ i,b is the Rician factor, ρ is the path loss at the reference distance d0 = 1m, d i,b is the distance between RIS and BS, α i,b is the path loss index of the RIS to BS link, non-LOS component Each element of follows a complex Gaussian distribution with zero mean and unit variance, and similarly h m,e ,h m,b ,h m,i ,h i,e Following the Rician distribution, due to the crowded urban environment and the blocking effect between vehicles, κ m,b and κ m,e All are zero; Step 2.3: Modeling of signal reception process; The received signal of the mth V2I link at the BS is expressed as: Among them, P m is the transmission power of the mth target vehicle, s m represents the unit energy signal sample associated with the computational task, and the noise vector n m It is expressed as: n m =[n1,...n K ] T in, The uplink signal-to-interference-plus-noise ratio (SINR) of the mth V2I link at the BS is given by: Similarly, the eavesdropping signal of the mth V2I link at the eth eavesdropping vehicle is expressed as: in, The SINR of the mth V2I link at the eth eavesdropping vehicle is expressed as: Therefore, the capacity of the mth V2I link and the eavesdropping capacity from the e-th eavesdropping vehicle to the mth V2I link are expressed as: C m =log(1+η m ) C e,m =log(1+η e,m ) In the MEC vehicle network, once the user completes the offloading process, the BS flexibly allocates the computing resources of the MEC server according to the size of the task. Each CPU cycle of the MEC server processes a certain number of data bits. Assume that the total computing power is ζ bit / s; Step 3: Model the optimization target of the RIS-assisted MEC vehicle network scenario constructed in step 1 and construct the objective function for model solving; Step 3.1: Security process modeling; Any unserved vehicle eavesdrops on any V2I link. To protect mission data from being eavesdropped, the redundancy used to protect confidential information is expressed as: max{0,R b -R S } Among them, R b is the code rate, R S target confidentiality rates for confidential information; If the capacity of the eavesdropper is C e Greater than R b -R S , a confidentiality interrupt will be sent, using capacity C b Approximate R b , so the confidentiality transmission rate of the mth V2I link is expressed as: R S,m =[0,(C m -max C e,m )] + ,e∈ε Where [x] + =max{0,x}; The MEC service time of the mth V2I link, that is, the sum of the task offloading time and the task calculation time, is expressed as: Among them, S m is the task size, m is the allocated computing resource; Step 3.2: Optimization target modeling; The optimization goal is to minimize the service time by designing the RIS reflection coefficient matrix Θ and MEC resource allocation for different computing tasks. The former will affect the transmission time, while the latter will determine the computing time. Considering that the entire MEC service cycle is determined by the maximum service time of all V2I links, the above goal is transformed into the following minimum-maximum problem: Among them, constraint C1 means that the sum of computing resources allocated to different target vehicles is a fixed value, and constraint C2 means that the modulus of the RIS reflection coefficient is constrained to be unit modulus; Step 4: Construct a deep reinforcement learning algorithm model based on the optimization goal proposed in step 3; Step 5: Construct a deep reinforcement learning training model based on the deep reinforcement learning algorithm model proposed in step 4. Combined with the communication scenarios and objective functions in steps 1, 2, and 3, set the state, action, and reward of the training model, and perform model training on the optimization target of the RIS-assisted MEC vehicle network communication scenario. Step 5.1: State space setup; At time slot t, the state of the mth V2I link Including confidentiality rate Remaining unloading tasks Remaining computing tasks Amount of MEC resources occupied Global channel state information It is expressed as: In summary, the state of the mth V2I link is expressed as: At time slot t, the total environment of M V2I links is expressed as: Step 5.2: Action space setting; Based on the current state S t , BS will design the RIS phase shift matrix and MEC resource allocation, in each time slot t, the action space is expressed as: a t ={Θ t ,g t } in, It is the allocation of computing resources; Step 5.3: Reward function setting; At time slot t, corresponding to the current action a t The reward is expressed as: in, represents the secure MEC service time of the mth V2I link at time slot t, t m,1 is the current consumed time, t m,2 It is the remaining time estimated based on the current action, which includes the remaining transmission time and the remaining calculation time. There are three cases for estimating the remaining time: (1) All target vehicles are in the process of task offloading. The remaining transmission time of each target vehicle is based on the current action, and the remaining computation time of each target vehicle is calculated by distributing the computation resources evenly to all target vehicles, that is, ζ min ; (2) Some target vehicles are in the process of task offloading, and other target vehicles are in the process of task calculation. For the target vehicles in the process of task offloading, the remaining transmission time of each user in the process of task offloading is calculated based on the current action, and the remaining transmission time of each user in the process of task offloading is calculated based on the computing resources. The strategy estimates the remaining computation time, where ζ min It is the minimum computing resource of the target vehicle during the task calculation process. For the target vehicle during the task calculation process, it only needs to estimate the remaining computing time based on the current action; (3) All target vehicles are in the task calculation process, and the remaining calculation time of all target vehicles is estimated based on the current action; In order to improve the confidentiality transmission rate, the penalty factor is expressed as: If the current action can meet the confidentiality rate requirement of the mth link Then ν m = 0, otherwise ν m =v * , ν * It is an artificially set parameter, which is a negative number; Based on the setting of the reward function, the DDPG algorithm will continuously learn the action strategy to reduce the maximum safe MEC service time within the given constraints. The total cumulative reward is expressed as: Where γ is the discount factor; Step 6: Based on the training model in step 5, the RIS-assisted MEC vehicle network decision model is obtained to obtain the optimal solution to the optimization problem, that is, the vehicle network security computing offloading solution is obtained.

2. The method for offloading secure computing in connected vehicles assisted by intelligent reflective surfaces according to claim 1, characterized in that: The specific method of step 4 is: DDPG is a model-free, off-policy Actor-Critic architecture algorithm. The Actor network is used to predict actions, and the Critic network is used to evaluate the future benefits of taking the action in the current state. Both the Actor network and the Critic network are composed of two deep neural network DNN networks: a training network and a target network. The training and target network parameters of the Actor network are θ and θ, respectively. a and θ a′ , the training and target network parameters of the Critic network are θ c and θ c′ ; At time slot t, the Actor training network will S t As input, and output action a t , the Critic training network will S t and a t Takes as input and outputs state-action function value state-actionvalue Q π (S t ,a t ∣θ c ), which is expressed as: Q π (WITH t ,and t ∣θ c )=Eπ[R t ∣S t ,and t ,π] Where E[·] represents the expected function, π represents the strategy of the Actor training network, and when enough quadruplets (S t ,a t ,r t ,S t+1 ), the model optimizer randomly draws N from the experience replay pool d The samples are used to update the training network of Actor and Critic. The k-th tuple y k The target state-action function value Q′ is expressed as: y k =r k +γQ′ π′ (S k+1 ,π′(S k+1 ∣θ a′ )∣θ c′ ) Among them, π′ represents the strategy of the Actor target network; The Critic training network uses the mean square error MSE function to update the network, which is expressed as follows: The Actor training network uses a deterministic policy gradient function to update the network, expressed as: The updates of the Actor and Critic target networks are as follows: i c’ =t c i c +(1-t c )i c’ i a’ =t a i a +(1-t a )i a’ Among them, τ c and τ a is the soft update coefficient, which satisfies τ c ,τ a ∈[0,1].

3. The method for offloading secure computing in connected vehicles assisted by intelligent reflective surfaces according to claim 1, characterized in that: The specific method of step 6 is: Step 6.1: Initialization; Randomly initialize the parameters θ of the Actor and Critic training networks a ,θ c , the parameters θ of the Actor target network a′ Initialized to θ a , the parameters θ of the Critic target network c′ Initialized to θ c , clear the experience replay pool D; Step 6.2: Training phase; Randomly initialize the positions of the target vehicle and the eavesdropped vehicle, and initialize the task volume of the target vehicle's requested service; At each time slot t, BS interacts with the dynamic environment and obtains the state S t Based on the current state, BS obtains action a from the Actor network of the Mth V2I link t , thereby setting the reflection coefficient matrix and MEC resource allocation for the target vehicle; BS obtains the state S of the next time slot t+1 from the changing environment t+1 , and calculate the t The reward r obtained from the environment t ; The states, actions, and rewards in the above process are stored as tuples (S t ,a t ,r t ,S t+1 ), and stores the tuple in the experience replay pool D, and obtains the state-action function Qπ(S t ,a t ∣θ c ); When there are enough tuples in the experience replay pool, N tuples are taken from it. d The parameters of the Critic and Actor networks are updated with samples of different sizes. When the task loads of all target vehicles have been calculated, the model training ends and the above process is repeated until the model training converges. Step 6.3: Decision-making stage; The decision model that has been trained and converged is used in a random dynamic vehicle network scenario. The optimal RIS reflection coefficient matrix and MEC resource allocation are determined in each time slot, so that the maximum MEC service time in the whole process is minimized, and finally the optimal solution of the optimization objective is obtained.

4. A system for implementing the intelligent reflective surface-assisted vehicle networking security computing offloading method according to any one of claims 1 to 3, characterized in that: include: RIS-assisted MEC vehicle network communication module: used to realize the construction of RIS-assisted MEC vehicle network communication scenarios. In this module, the base station establishes multiple communication links with dynamic vehicles; RIS-assisted safety communication module: used to realize the construction of RIS-assisted safety communication scenarios. In this module, RIS technology provides security for dynamic vehicle communication; Secure computing service optimization target module: used to realize the construction of optimization targets for RIS-assisted MEC vehicle network scenarios; Deep reinforcement learning algorithm selection module: used to build a deep reinforcement learning algorithm model based on optimization objectives; Deep reinforcement learning model training module: used to build a deep reinforcement learning training model. In this module, the model training is performed on the optimization target of the RIS-assisted MEC vehicle network scenario. Deep reinforcement learning decision model module: used to implement the RIS-assisted MEC vehicle network decision model. In this module, the optimal RIS coefficient matrix and MEC resource allocation in the dynamic Internet of Vehicles scenario are obtained.

5. A vehicle network security computing offloading device assisted by an intelligent reflective surface, characterized in that: include: Memory for storing computer programs; A processor is used to implement a method for unloading safe computing in an Internet of Vehicles assisted by an intelligent reflective surface as described in any one of claims 1 to 3 when executing the computer program.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it can implement the intelligent reflective surface-assisted vehicle network security computing offloading method described in any one of claims 1 to 3.

Images