Method and system for optimizing movable antenna in federated Internet of Vehicles based on differential privacy

Abstract

The invention discloses a method and a system for optimizing a movable antenna in a federated Internet of Vehicles based on differential privacy, and the method comprises the steps: constructing a federated Internet of Vehicles model based on an intelligent movable antenna, including a communication model, a calculation model and a privacy cost model, and building a target function with the minimization of base station compensation cost and communication energy consumption as targets; and further decomposing the target function into two sub-problems, solving a closed-form solution by adopting a Karush-Kuhn-Tucker condition, and solving and optimizing target function distribution of the two sub-problems by adopting a depth Q algorithm in reinforcement learning. The system comprises a construction module, an additive noise distribution module and an antenna position optimization module, the compensation cost of the base station in the Internet of Vehicles is modeled, the compensation cost of the base station in the Internet of Vehicles is minimized, and the communication rate between the vehicle user and the base station is improved and the communication energy consumption is reduced by optimizing the position of the movable antenna.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of Internet of Vehicles, and particularly relates to a mobile antenna optimization method and system based on differential privacy in federated Internet of Vehicles. BACKGROUND

[0002] The base station compensation cost and communication energy consumption optimization problem in Internet of Vehicles is a popular research direction. Because of the heterogeneous resources of vehicle users, such as computing resources, communication resources, etc., the mobility of vehicle users, the sensitivity of vehicle users to local data, the energy consumed by vehicle users to update local models, the energy consumed by vehicle users to add noise to the updated local models, and the energy consumed by vehicle users to upload the noisy local models, all have a certain degree of influence on the decision of the base station's compensation cost for vehicle users, and the mobile antenna deployed at the base station can improve the communication environment of vehicle users and reduce communication energy consumption by optimizing its position. Because of the numerous influencing factors, it is difficult to give an optimal method through theoretical derivation, and it is necessary to propose an optimal base station compensation cost and communication energy consumption optimization method.

[0003] According to the search, the Chinese invention patent with the publication number CN120264340A proposes an optimization method and system for Internet of Vehicles with antenna position cost based on mobile antenna, which includes: constructing an Internet of Vehicles model based on a privacy protection federated learning architecture assisted by intelligent mobile antenna, according to the total cost of vehicle users within the base station signal coverage range, the base station gives corresponding compensation, and establishes a target function of minimizing the base station compensation cost; further decompose the target function into two sub-problems, and respectively use the proposed greedy vehicle user selection algorithm and the adaptive particle swarm optimization algorithm to solve and optimize the parameters of the target function, and in each round of federated communication, the base station obtains the selected vehicle user set and schedules the position of the intelligent mobile antenna to minimize the base station compensation cost. This method models the base station compensation cost, optimizes the base station compensation cost in Internet of Vehicles, and improves the communication rate between vehicle users and the base station by scheduling the position of the intelligent mobile antenna, and reduces the communication delay.

[0004] The technical comparison between the present application and the above-mentioned patent is as follows:

[0005] The patent "optimization method and system for Internet of Vehicles with antenna position cost based on mobile antenna" uses a greedy vehicle user selection strategy. This algorithm is an improved version of the greedy algorithm, which serves the purpose of selecting vehicle users and ensures the lowest cost. Meanwhile, the particle swarm optimization algorithm is used for scheduling the position of the mobile antenna, which is a heuristic algorithm. By having each particle represent a feasible solution of the antenna position cluster, the optimal antenna position is found. The two algorithms are used to solve the target function of minimizing the cost.

[0006] This patent employs the Karush-Kuhn-Tucker conditions from convex optimization problems to solve for the objective function, which minimizes cost and communication energy consumption. This differs from the objective function of the patent "Optimization Method and System for Vehicle-to-Everything Based on Mobile Antennas and Including Antenna Location Costs." Furthermore, it utilizes the Deep Q-learning algorithm from reinforcement learning for solving and parameter optimization of intelligent mobile antenna location scheduling. This represents a combination of reinforcement learning and deep learning.

[0007] The two differ fundamentally in terms of the "objective function" and the "solution algorithm". Summary of the Invention

[0008] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a method and system for optimizing movable antennas in a federated vehicle network based on differential privacy. By optimizing the additive noise allocation of vehicle users by the base station and optimizing the position of movable antennas, the cost of base station compensation in the vehicle network is optimized, the transmission rate of vehicle users is improved, and the communication energy consumption is reduced.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A method for optimizing movable antennas in a federated vehicular network based on differential privacy, characterized by the following steps:

[0011] S1: Construct a base station compensation cost and communication energy consumption model for a federal vehicle network equipped with a smart mobile antenna based on differential privacy. Based on the compensation cost and communication energy consumption between the base station and vehicle users, establish an objective function to minimize the base station compensation cost and communication energy consumption.

[0012] S2: Decompose the target problem into two sub-problems, and solve the two sub-problems respectively using the Karush-Kuhn-Tucker condition and the Deep Q-learning algorithm in reinforcement learning to obtain the optimal additive noise and the location of the smart mobile antenna allocated by the base station to the vehicle user.

[0013] S3: Obtain the optimal additive noise allocated by the base station to the vehicle user and the antenna position after scheduling during each round of federated learning in the vehicle network. Calculate the compensation cost and communication energy consumption of the base station in the current round and output the optimized base station compensation cost and communication energy consumption.

[0014] As a preferred technical solution of the present invention: In step S1, the objective function for minimizing base station compensation cost and communication energy consumption is specifically established as follows:

[0015] The established objective function consists of two parts: compensation cost and communication energy consumption. The objective function for the optimization problem is shown below:

[0016] (1);

[0017] stC1: (2);

[0018] C2: (3);

[0019] C3: (4);

[0020] C4: (5);

[0021] in, This represents the additive noise of vehicle user n. This represents the set of locations for intelligent movable antennas; n represents the vehicle user ID. Represents the total set of vehicle users; This represents the compensation cost of the base station to vehicle users; This represents the communication energy consumption of vehicle user n; Represents the privacy budget of vehicle user n and ,in This indicates the sensitivity of the dataset in differential privacy. This represents the additive noise of vehicle user n. Indicates the total privacy budget; This indicates the position of antenna z within the intelligent movable antenna area. This indicates the position of antenna k within the intelligent movable antenna area. This represents the minimum distance between two adjacent antennas. Indicates the area of ​​the intelligent movable antenna; This represents the communication latency of vehicle user n. This represents the computational delay for vehicle user n. This represents the maximum latency that the base station can tolerate. Indicates the distance traveled by the vehicle user. Indicates the radius of the base station signal coverage area. This represents the cosine angle between the base station coverage area and the road. This indicates the distance between the vehicle user and the entrance to the base station's coverage area.

[0022] As a preferred technical solution of the present invention: Under the vehicle-to-everything (V2X) framework based on intelligent movable antennas, the communication model of vehicle users can be represented as follows:

[0023] (6);

[0024] (7);

[0025] (8);

[0026] For vehicle users' transmission rate, For vehicle users' transmission power, The base station receives the hybrid vector from the vehicle user. This refers to Gaussian white noise during channel transmission. For vehicle users' transmission latency, The amount of data uploaded by vehicle users; For vehicle users' transmission energy consumption, This refers to the vehicle's transmission power.

[0027] As a preferred technical solution of the present invention: Under the vehicle-to-everything (V2X) framework based on intelligent movable antennas, the computational model of vehicle users can be represented as follows:

[0028] (9);

[0029] For vehicle users' computational latency, The number of CPU cycles required to process a single sample for a vehicle user. For the size of the local dataset for vehicle users, The CPU frequency calculated locally for the vehicle user.

[0030] As a preferred technical solution of the present invention: Under the vehicle-to-everything (V2X) framework based on intelligent mobile antennas, the compensation cost model of the base station for vehicle user n can be expressed as:

[0031] (10);

[0032] This represents the compensation cost of the base station for vehicle user n. This represents the additive noise of vehicle user n. This represents the cost of compensation per unit of noise required by a vehicle user.

[0033] As a preferred technical solution of the present invention: Under the vehicle-to-everything (V2X) framework based on intelligent movable antennas, the mobility model of vehicle users can be represented as:

[0034] (11);

[0035] This represents the distance traveled by vehicle user n. Let n be the speed of the vehicle user. This represents the communication latency of vehicle user n. This represents the computation delay for vehicle user n.

[0036] As a preferred technical solution of the present invention: In step S2, the target problem is decomposed into two sub-problems, and the two sub-problems are solved respectively using the Karush-Kuhn-Tucker conditions and the Deep Q-learning algorithm in reinforcement learning, as follows:

[0037] First, the Karush-Kuhn-Tucker condition is used to determine the optimal additive noise allocation for vehicle users at the base station. Then, based on the determined additive noise, the Deep Q-learning algorithm in reinforcement learning is used to solve for the location of the smart mobile antenna. The specific process is as follows:

[0038] S21: The first subproblem after decomposing the objective problem is represented as:

[0039] (12);

[0040] stC1: (13)

[0041] C2: (14);

[0042] C3: (15);

[0043] in, The additive noise represents vehicle user n; n represents vehicle user; Represents the total set of vehicle users; This indicates the compensation cost of the base station to vehicle users; This represents the privacy budget for vehicle user n. Indicates the total privacy budget; This represents the communication latency of vehicle user n. This represents the computational delay for vehicle user n. This represents the maximum latency that the base station can tolerate. Indicates the distance traveled by the vehicle user. Indicates the radius of the base station signal coverage area. This represents the cosine angle between the base station coverage area and the road. This indicates the distance between the vehicle user and the entrance to the base station's coverage area;

[0044] The subproblem is solved using the Karush-Kuhn-Tucker conditions, and the specific steps are as follows:

[0045] Since the objective function and constraints of this subproblem are differentiable and satisfy the Karush-Kuhn-Tucker conditions, we can derive a specific expression for the Karush-Kuhn-Tucker conditions and construct the corresponding Lagrangian function. By solving the subproblem using the Karush-Kuhn-Tucker conditions, we can obtain the expression for additive noise. The closed-form solution can be obtained by substituting this solution into the objective function of the compensation cost to obtain the optimal compensation cost for the base station.

[0046] The closed-form solution is as follows:

[0047] (16);

[0048] S22: The second subproblem after decomposing the objective problem is represented as:

[0049] (17);

[0050] stC1: (18);

[0051] in, Represents a set of movable antennas. This indicates the position of antenna z within the antenna system. This indicates the position of antenna k within the antenna system. This represents the minimum distance between two adjacent antennas. Indicates the antenna system;

[0052] The Deep Q-learning algorithm from reinforcement learning is used to solve this subproblem. The specific steps are as follows:

[0053] First, the Markov decision process is constructed as follows:

[0054] (19);

[0055] in Representing the state space, Represents the action space. Represents the state transition probability. This indicates a reward, among which It can be represented as:

[0056] (20);

[0057] Where d represents the set of movable antennas.

[0058] It can be represented as:

[0059] (twenty one);

[0060] in This indicates that the antenna is moving eastward. This indicates that the antenna is moving south. This indicates that the antenna is moving westward. This indicates that the antenna is moving north.

[0061] Then, construct the target network:

[0062] In the Deep Q-learning algorithm, the target network is used to stabilize Q-value training by mitigating the problem of overestimating Q-values. The target network updates the target Q-value using the Bellman equation, which can be represented as follows:

[0063] (twenty two);

[0064] in Indicates the target Q value. This represents the reward at the i-th step. Indicates the discount factor. It is the action selected at the (i+1)th step. It is the original Q-value in the Q-network. This is the state at the (i+1)th step. These are the parameters of the Q network.

[0065] Next, build the experience replay pool:

[0066] Experience replay stores the agent's historical experience and randomly samples data during training. Each interaction between the agent and the environment generates sample data, represented as follows:

[0067] (twenty three);

[0068] in This represents the experience replay pool. This represents the state at the i-th step. It is the action selected in the i-th step. This represents the reward at the i-th step. This represents the state at the (i+1)th step. Indicates the termination condition.

[0069] Simultaneously, the loss function is constructed using the mean squared error and the Bellman equation, as follows:

[0070] (twenty four);

[0071] in, The parameter is loss function, Expressing expectations, This represents the experience replay pool. Indicates the target Q value. It is the original Q-value in the Q-network.

[0072] Finally, build Greedy strategy:

[0073] In the Deep Q-learning algorithm, the agent learns through... - A greedy strategy is used to balance exploration and development. The core mechanism is that the agent will randomly select a target with... The action is probabilistic, and the action with the highest current Q value is greedily selected, with a probability of 1- Specifically, it is expressed as:

[0074] (25);

[0075] (26);

[0076] in This represents the action selected in the i-th step. This represents the probability of a greedy strategy choosing an action. This represents the original Q-value in the Q-network.

[0077] right The linear decreasing strategy is specifically expressed as follows:

[0078] (27);

[0079] in This represents the state at the i-th step. value, Indicates the largest value, Represents the smallest value, This represents the value of the current step. This represents the total number of steps.

[0080] A mobile antenna optimization system for federated vehicular networks based on differential privacy, characterized by comprising:

[0081] The construction module is used to build a base station compensation cost and communication energy consumption optimization system in a federated vehicle network based on differential privacy. It establishes an objective function that minimizes base station compensation cost and communication energy consumption based on the optimal additive noise allocated by the base station to vehicle users and the location of the intelligent movable antenna. The construction module also includes a communication model, a computation model, a compensation cost model, and a mobility model for vehicle users.

[0082] Additive noise allocation module: used to obtain the optimal additive noise allocated by the base station to vehicle users, and to solve the objective function and optimize the parameters using the Karush-Kuhn-Tucker conditions to obtain the optimal additive noise for vehicle users, and use the result for antenna position optimization module;

[0083] Antenna position optimization module: This module is used to obtain the current optimized antenna position. It uses the Deep Q-learning algorithm in reinforcement learning and the results of the additive noise allocation module to solve and optimize the objective function, thereby obtaining the optimized intelligent mobile antenna position and the optimal communication energy consumption.

[0084] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0085] This invention constructs a vehicle-to-everything (V2X) model based on movable antennas. Based on the vehicle user communication model, computation model, privacy compensation cost, and vehicle user mobility, it establishes an objective function to minimize base station compensation cost and communication energy consumption. By optimally allocating additive noise to vehicle users and optimizing the location of movable antennas, it optimizes base station compensation cost in V2X, improves vehicle user transmission rate, and reduces communication energy consumption. Attached Figure Description

[0086] Figure 1 This is a flowchart of a movable antenna optimization method in a differential privacy-based federated vehicle network;

[0087] Figure 2 This is a system framework diagram for a movable antenna optimization system in a federated vehicle network based on differential privacy.

[0088] Figure 3 This is a comparison chart of the base station compensation cost between this application and the greedy allocation algorithm, the uniform allocation algorithm, and the random allocation algorithm;

[0089] Figure 4 This is a comparison diagram of the present application with fixed antenna algorithms and random antenna algorithms.

[0090] List of reference numerals in the attached diagram:

[0091] 1. Construction module; 2. Additive noise distribution module; 3. Antenna position optimization module. Detailed Implementation

[0092] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0093] like Figure 1 As shown, the movable antenna optimization method in federated vehicular networks based on differential privacy proposed in this invention minimizes the base station compensation cost of vehicular networks by optimizing the base station compensation cost and antenna position, thereby improving the transmission rate for vehicle users and reducing communication energy consumption.

[0094] Specifically, the following steps are included:

[0095] S1: Construct a base station compensation cost and communication energy consumption model for a federal vehicle network equipped with a smart mobile antenna based on differential privacy. Based on the compensation cost and communication energy consumption between the base station and vehicle users, establish an objective function to minimize the base station compensation cost and communication energy consumption.

[0096] Step S1 is as follows:

[0097] The objective function for minimizing base station compensation costs and communication energy consumption is established as follows:

[0098] The established objective function consists of two parts: compensation cost and communication energy consumption. The objective function for the optimization problem is shown below:

[0099] (1);

[0100] stC1: (2);

[0101] C2: (3);

[0102] C3: (4);

[0103] C4: (5);

[0104] in, This represents the additive noise of vehicle user n. This represents the set of locations for intelligent movable antennas; n represents the vehicle user ID. Represents the total set of vehicle users; This represents the compensation cost of the base station to vehicle users; This represents the communication energy consumption of vehicle user n; Represents the privacy budget of vehicle user n and ,in This indicates the sensitivity of the dataset in differential privacy. This represents the additive noise of vehicle user n. Indicates the total privacy budget; This indicates the position of antenna z within the intelligent movable antenna area. This indicates the position of antenna k within the intelligent movable antenna area. This represents the minimum distance between two adjacent antennas. Indicates the area of ​​the intelligent movable antenna; This represents the communication latency of vehicle user n. This represents the computational delay for vehicle user n. This represents the maximum latency that the base station can tolerate. Indicates the distance traveled by the vehicle user. Indicates the radius of the base station signal coverage area. This represents the cosine angle between the base station coverage area and the road. This indicates the distance between the vehicle user and the entrance to the base station's coverage area.

[0105] Within the vehicle-to-everything (V2X) framework based on intelligent mobile antennas, the communication model for vehicle users can be represented as follows:

[0106] (6);

[0107] (7);

[0108] (8);

[0109] For vehicle users' transmission rate, For vehicle users' transmission power, The base station receives the hybrid vector from the vehicle user. This refers to Gaussian white noise during channel transmission. For vehicle users' transmission latency, The amount of data uploaded by vehicle users; For vehicle users' transmission energy consumption, This refers to the vehicle's transmission power.

[0110] Specifically, a system is constructed that includes one base station, an edge server and a mobile antenna system deployed within the base station, and N vehicle users.

[0111] Transmission latency for vehicle users Represented as:

[0112]

[0113] in, The amount of data uploaded by vehicle users. For vehicle users, the transmission rate.

[0114] Required transmission power consumption for:

[0115] in, Transmission power for vehicle users.

[0116] Within the vehicle-to-everything (V2X) framework based on intelligent mobile antennas, the computational model for vehicle users can be represented as follows:

[0117] (9);

[0118] For vehicle users' computational latency, The number of CPU cycles required to process a single sample for a vehicle user. For the size of the local dataset for vehicle users, The CPU frequency calculated locally for the vehicle user.

[0119] Specifically, the local computing latency of vehicle users for:

[0120]

[0121] in, The number of CPU cycles required to process a single sample for a vehicle user. For the size of the local dataset for vehicle users, The CPU frequency calculated locally for the vehicle user.

[0122] In the vehicle-to-everything (V2X) framework based on intelligent mobile antennas, the compensation cost model for vehicle user n by the base station can be expressed as:

[0123] (10);

[0124] This represents the compensation cost of the base station for vehicle user n. This represents the additive noise of vehicle user n. This represents the cost of compensation per unit of noise required by a vehicle user.

[0125] As a preferred technical solution of the present invention: Under the vehicle-to-everything (V2X) framework based on intelligent movable antennas, the mobility model of vehicle users can be represented as:

[0126] (11);

[0127] This represents the distance traveled by vehicle user n. Let n be the speed of the vehicle user. This represents the communication latency of vehicle user n. This represents the computation delay for vehicle user n.

[0128] S2: Decompose the target problem into two sub-problems, and solve the two sub-problems respectively using the Karush-Kuhn-Tucker condition and the Deep Q-learning algorithm in reinforcement learning to obtain the optimal additive noise and the location of the smart mobile antenna allocated by the base station to the vehicle user.

[0129] Step S2 is as follows:

[0130] The target problem is decomposed into two subproblems, and the Karush-Kuhn-Tucker conditions and the Deep Q-learning algorithm in reinforcement learning are used to solve the two subproblems respectively, as follows:

[0131] First, the Karush-Kuhn-Tucker condition is used to determine the optimal additive noise allocation for vehicle users at the base station. Then, based on the determined additive noise, the Deep Q-learning algorithm in reinforcement learning is used to solve for the location of the smart mobile antenna. The specific process is as follows:

[0132] S21: The first subproblem after decomposing the objective problem is represented as:

[0133] (12);

[0134] stC1: (13)

[0135] C2: (14);

[0136] C3: (15);

[0137] in, The additive noise represents vehicle user n; n represents vehicle user; Represents the total set of vehicle users; This indicates the compensation cost of the base station to vehicle users; This represents the privacy budget for vehicle user n. Indicates the total privacy budget; This represents the communication latency of vehicle user n. This represents the computational delay for vehicle user n. This represents the maximum latency that the base station can tolerate. Indicates the distance traveled by the vehicle user. Indicates the radius of the base station signal coverage area. This represents the cosine angle between the base station coverage area and the road. This indicates the distance between the vehicle user and the entrance to the base station's coverage area;

[0138] The subproblem is solved using the Karush-Kuhn-Tucker conditions, and the specific steps are as follows:

[0139] Since the objective function and constraints of this subproblem are differentiable and satisfy the Karush-Kuhn-Tucker conditions, we can derive a specific expression for the Karush-Kuhn-Tucker conditions and construct the corresponding Lagrangian function. By solving the subproblem using the Karush-Kuhn-Tucker conditions, we can obtain the expression for additive noise. The closed-form solution can be obtained by substituting this solution into the objective function of the compensation cost to obtain the optimal compensation cost for the base station.

[0140] The closed-form solution is as follows:

[0141] (16);

[0142] S22: The second subproblem after decomposing the objective problem is represented as:

[0143] (17);

[0144] stC1: (18);

[0145] in, Represents a set of movable antennas. This indicates the position of antenna z within the antenna system. This indicates the position of antenna k within the antenna system. This represents the minimum distance between two adjacent antennas. Indicates the antenna system;

[0146] The Deep Q-learning algorithm from reinforcement learning is used to solve this subproblem. The specific steps are as follows:

[0147] First, the Markov decision process is constructed as follows:

[0148] (19);

[0149] in Representing the state space, Represents the action space. Represents the state transition probability. This indicates a reward, among which It can be represented as:

[0150] (20);

[0151] Where d represents the set of movable antennas.

[0152] It can be represented as:

[0153] (twenty one);

[0154] in This indicates that the antenna is moving eastward. This indicates that the antenna is moving south. This indicates that the antenna is moving westward. This indicates that the antenna is moving north.

[0155] Then, construct the target network:

[0156] In the Deep Q-learning algorithm, the target network is used to stabilize Q-value training by mitigating the problem of overestimating Q-values. The target network updates the target Q-value using the Bellman equation, which can be represented as follows:

[0157] (twenty two);

[0158] in Indicates the target Q value. This represents the reward at the i-th step. Indicates the discount factor. It is the action selected at the (i+1)th step. It is the original Q-value in the Q-network. This is the state at the (i+1)th step. These are the parameters of the Q network.

[0159] Next, build the experience replay pool:

[0160] Experience replay is a key technique in the Deep Q-learning algorithm. It stores the agent's historical experience and randomly samples data during training. Each interaction between the agent and the environment generates sample data, as shown below:

[0161] (twenty three);

[0162] in This represents the experience replay pool. This represents the state at the i-th step. It is the action selected in the i-th step. This represents the reward at the i-th step. This represents the state at the (i+1)th step. Indicates the termination condition.

[0163] Simultaneously, the loss function is constructed using the mean squared error and the Bellman equation, as follows:

[0164] (twenty four);

[0165] in, The parameter is loss function, Expressing expectations, This represents the experience replay pool. Indicates the target Q value. It is the original Q-value in the Q-network.

[0166] Finally, build Greedy strategy:

[0167] In the Deep Q-learning algorithm, the agent learns through... - A greedy strategy is used to balance exploration and development. The core mechanism is that the agent will randomly select a target with... The action is probabilistic, and the action with the highest current Q value is greedily selected, with a probability of 1- Specifically, it is expressed as:

[0168] (25);

[0169] (26);

[0170] in This represents the action selected in the i-th step. This represents the probability of a greedy strategy choosing an action. This represents the original Q-value in the Q-network.

[0171] right The linear decreasing strategy is specifically expressed as follows:

[0172] (27);

[0173] in This represents the state at the i-th step. value, Indicates the largest value, Represents the smallest value, This represents the value of the current step. This represents the total number of steps.

[0174] S3: Obtain the optimal additive noise allocated by the base station to the vehicle user and the antenna position after scheduling during each round of federated learning in the vehicle network. Calculate the compensation cost and communication energy consumption of the base station in the current round and output the optimized base station compensation cost and communication energy consumption.

[0175] like Figure 3 and Figure 4 The chart shown is a comparison of base station compensation costs and optimized communication energy consumption. Figure 3 As can be seen, the compensation cost of this invention is lower than that of other algorithms for base stations. From... Figure 4 As can be seen, this invention has lower communication energy consumption compared to other algorithms.

[0176] This invention constructs a vehicle-to-everything (V2X) model based on a movable antenna. Based on the vehicle user communication model, computation model, privacy cost, and vehicle user mobility, it establishes an objective function to minimize base station compensation cost and communication energy consumption. By optimizing the additive noise allocation of the base station to vehicle users and optimizing the location of the movable antenna, it achieves the optimization of base station compensation cost in the V2X, improves the transmission rate of vehicle users, and reduces communication energy consumption.

[0177] like Figure 2 As shown, the movable antenna optimization system in federated vehicular networks based on differential privacy proposed in this invention includes:

[0178] Module 1: This module is used to construct a base station compensation cost and communication energy consumption optimization system for a federated vehicle network based on differential privacy. It establishes an objective function that minimizes base station compensation cost and communication energy consumption based on the optimal additive noise allocated by the base station to vehicle users and the location of the intelligent movable antenna. Module 1 also includes a communication model, a computation model, a compensation cost model, and a mobility model for vehicle users.

[0179] Additive noise allocation module 2: It is used to obtain the optimal additive noise allocated by the base station to the vehicle user, and to solve the objective function and optimize the parameters using the Karush-Kuhn-Tucker condition to obtain the optimal additive noise for the vehicle user, and then use the result for antenna position optimization module 3;

[0180] Antenna position optimization module 3: It is used to obtain the current optimized antenna position. It uses the DeepQ-learning algorithm in reinforcement learning and the results of additive noise allocation module 2 to solve and optimize the objective function, so as to obtain the optimized intelligent mobile antenna position and the optimal communication energy consumption.

[0181] Module 1 establishes the objective function for minimizing base station compensation costs and communication energy consumption, including:

[0182] The objective function for base station compensation cost and communication energy consumption is established, and the optimization problem is as follows:

[0183] ;

[0184] stC1: ;

[0185] C2: ;

[0186] C3: ;

[0187] C4: ;

[0188] in, This represents the additive noise of vehicle user n. Represents the set of movable antennas; n represents the vehicle user; Represents the total set of vehicle users; This represents the compensation cost of the base station to vehicle users; This represents the communication energy consumption of vehicle user n; This represents the privacy budget for vehicle user n. Indicates the total privacy budget; This indicates the position of antenna z within the antenna system. This indicates the position of antenna k within the antenna system. This represents the minimum distance between two adjacent antennas. Indicates the antenna system; This represents the communication latency of vehicle user n. This represents the computational delay for vehicle user n. This represents the maximum latency that the base station can tolerate. Indicates the distance traveled by the vehicle user. Indicates the radius of the base station signal coverage area. This represents the cosine angle between the base station coverage area and the road. This indicates the distance between the vehicle user and the entrance to the base station's coverage area;

[0189] Construct additive noise allocation module 2, with the objective function of the subproblem being:

[0190] ;

[0191] st C1: ;

[0192] C2: ;

[0193] C3: ;

[0194] in, The additive noise represents vehicle user n; n represents vehicle user; Represents the total set of vehicle users; This indicates the compensation cost of the base station to vehicle users; This represents the privacy budget for vehicle user n. Indicates the total privacy budget; This represents the communication latency of vehicle user n. This represents the computational delay for vehicle user n. This represents the maximum latency that the base station can tolerate. Indicates the distance traveled by the vehicle user. Indicates the radius of the base station signal coverage area. This represents the cosine angle between the base station coverage area and the road. This indicates the distance between the vehicle user and the entrance to the base station's coverage area;

[0195] The objective function of the subproblem is solved using the Karush-Kuhn-Tucker conditions, resulting in an additive noise allocation model.

[0196] Construct antenna position optimization module 3, with the objective function of the subproblem as follows:

[0197] ;

[0198] st C1: ;

[0199] in, This indicates the position of antenna z within the antenna system. This indicates the position of antenna k within the antenna system. This represents the minimum distance between two adjacent antennas. Indicates the antenna system;

[0200] The objective function of this subproblem is solved using the Deep Q-learning algorithm in reinforcement learning, and the optimized antenna position model is obtained.

[0201] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A method for optimizing movable antennas in a federated vehicular network based on differential privacy, characterized in that, Includes the following steps: S1: Construct a base station compensation cost and communication energy consumption model for a federal vehicle network equipped with a smart mobile antenna based on differential privacy. Based on the compensation cost and communication energy consumption between the base station and vehicle users, establish an objective function to minimize the base station compensation cost and communication energy consumption. S2: Decompose the target problem into two sub-problems, and solve the two sub-problems respectively using the Karush-Kuhn-Tucker condition and the Deep Q-learning algorithm in reinforcement learning to obtain the optimal additive noise and the location of the smart mobile antenna allocated by the base station to the vehicle user. S3: Obtain the optimal additive noise allocated by the base station to the vehicle user and the antenna position after scheduling during each round of federated learning in the vehicle network. Calculate the compensation cost and communication energy consumption of the base station in the current round and output the optimized base station compensation cost and communication energy consumption.

2. The method for optimizing movable antennas in a federated vehicular network based on differential privacy as described in claim 1, characterized in that, In step S1, the objective function for minimizing base station compensation cost and communication energy consumption is established as follows: The established objective function consists of two parts: compensation cost and communication energy consumption. The objective function for the optimization problem is shown below: （1）； s.t.C1: (2); C2： （3）； C3： （4）； C4： （5）； in, This represents the additive noise of vehicle user n. This represents the set of locations for intelligent movable antennas; n represents the vehicle user ID. Represents the total set of vehicle users; This represents the compensation cost of the base station to vehicle users; This represents the communication energy consumption of vehicle user n; Represents the privacy budget of vehicle user n and ,in This indicates the sensitivity of the dataset in differential privacy. This represents the additive noise of vehicle user n. Indicates the total privacy budget; This indicates the position of antenna z within the intelligent movable antenna area. This indicates the position of antenna k within the intelligent movable antenna area. This represents the minimum distance between two adjacent antennas. Indicates the area of ​​the intelligent movable antenna; This represents the communication latency of vehicle user n. This represents the computational delay for vehicle user n. This represents the maximum latency that the base station can tolerate. Indicates the distance traveled by the vehicle user. Indicates the radius of the base station signal coverage area. This represents the cosine angle between the base station coverage area and the road. This indicates the distance between the vehicle user and the entrance to the base station's coverage area.

3. The method for optimizing movable antennas in a federated vehicular network based on differential privacy as described in claim 2, characterized in that, In the vehicle-to-everything (V2X) framework based on intelligent mobile antennas, the communication model for vehicle users can be represented as follows: （6）； （7）； （8）； For vehicle users' transmission rate, For vehicle users' transmission power, The base station receives the hybrid vector from the vehicle user. This refers to Gaussian white noise during channel transmission. For vehicle users' transmission latency, The amount of data uploaded by vehicle users; For vehicle users' transmission energy consumption, This refers to the vehicle's transmission power.

4. The method for optimizing movable antennas in a federated vehicular network based on differential privacy as described in claim 2, characterized in that, Within the vehicle-to-everything (V2X) framework based on intelligent mobile antennas, the computational model for vehicle users can be represented as follows: （9）； For vehicle users' computational latency, The number of CPU cycles required to process a single sample for a vehicle user. For the size of the local dataset for vehicle users, The CPU frequency calculated locally for the vehicle user.

5. The method for optimizing movable antennas in a federated vehicular network based on differential privacy according to claim 2, characterized in that, In the vehicle-to-everything (V2X) framework based on intelligent mobile antennas, the compensation cost model for vehicle user n by the base station can be expressed as: （10）； This represents the compensation cost of the base station for vehicle user n. This represents the additive noise of vehicle user n. This represents the cost of compensation per unit of noise required by a vehicle user.

6. The method for optimizing movable antennas in a federated vehicular network based on differential privacy according to claim 2, characterized in that, Within the vehicle-to-everything (V2X) framework based on intelligent mobile antennas, the mobility model of vehicle users can be represented as: （11）； This represents the distance traveled by vehicle user n. Let n be the speed of the vehicle user. This represents the communication latency of vehicle user n. This represents the computation delay for vehicle user n.

7. The method for optimizing movable antennas in a federated vehicular network based on differential privacy according to claim 1, characterized in that, In step S2, the target problem is decomposed into two sub-problems, and the Karush-Kuhn-Tucker conditions and the Deep Q-learning algorithm in reinforcement learning are used to solve the two sub-problems respectively, as follows: First, the Karush-Kuhn-Tucker condition is used to determine the optimal additive noise allocation for vehicle users at the base station. Then, based on the determined additive noise, the Deep Q-learning algorithm in reinforcement learning is used to solve for the location of the smart mobile antenna. The specific process is as follows: S21: The first subproblem after decomposing the objective problem is represented as: （12）； s.t.C1: (13) C2： （14）； C3： （15）； in, The additive noise represents vehicle user n; n represents vehicle user; Represents the total set of vehicle users; This indicates the compensation cost of the base station to vehicle users; This represents the privacy budget for vehicle user n. Indicates the total privacy budget; This represents the communication latency of vehicle user n. This represents the computational delay for vehicle user n. This represents the maximum latency that the base station can tolerate. Indicates the distance traveled by the vehicle user. Indicates the radius of the base station signal coverage area. This represents the cosine angle between the base station coverage area and the road. This indicates the distance between the vehicle user and the entrance to the base station's coverage area; The subproblem is solved using the Karush-Kuhn-Tucker conditions, and the specific steps are as follows: Since the objective function and constraints of this subproblem are differentiable and satisfy the Karush-Kuhn-Tucker conditions, we can derive a specific expression for the Karush-Kuhn-Tucker conditions and construct the corresponding Lagrangian function. By solving the subproblem using the Karush-Kuhn-Tucker conditions, we can obtain the expression for additive noise. The closed-form solution can be obtained by substituting this solution into the objective function of the compensation cost to obtain the optimal compensation cost for the base station. The closed-form solution is as follows: （16）； S22: The second subproblem after decomposing the objective problem is represented as: （17）； s.t.C1: (18); in, Represents a set of movable antennas. This indicates the position of antenna z within the antenna system. This indicates the position of antenna k within the antenna system. This represents the minimum distance between two adjacent antennas. Indicates the antenna system; The Deep Q-learning algorithm from reinforcement learning is used to solve this subproblem. The specific steps are as follows: First, the Markov decision process is constructed as follows: （19）； in Representing the state space, Represents the action space, Represents the state transition probability. This indicates a reward, among which It can be represented as: （20）； Where d represents the set of movable antennas. It can be represented as: （21）； in This indicates that the antenna is moving eastward. This indicates that the antenna is moving south. This indicates that the antenna is moving westward. This indicates that the antenna is moving north. Then, construct the target network: In the Deep Q-learning algorithm, the target network is used to stabilize Q-value training by mitigating the problem of overestimating Q-values. The target network updates the target Q-value using the Bellman equation, which can be represented as follows: （22）； in Indicates the target Q value. This represents the reward at the i-th step. Indicates the discount factor. It is the action selected at the (i+1)th step. It is the original Q-value in the Q-network. This is the state at the (i+1)th step. These are the parameters of the Q network. Next, build the experience replay pool: Experience replay stores the agent's historical experience and randomly samples data during training. Each interaction between the agent and the environment generates sample data, represented as follows: （23）； in This represents the experience replay pool. This represents the state at the i-th step. It is the action selected in the i-th step. This represents the reward at the i-th step. This represents the state at the (i+1)th step. Indicates the termination condition. Simultaneously, the loss function is constructed using the mean squared error and the Bellman equation, as follows: （24）； in, The parameter is loss function, Expressing expectations, This represents the experience replay pool. Indicates the target Q value. It is the original Q-value in the Q-network. Finally, build Greedy strategy: In the Deep Q-learning algorithm, the agent learns through... - A greedy strategy is used to balance exploration and development. The core mechanism is that the agent will randomly select a target with... The action is probabilistic, and the action with the highest current Q value is greedily selected, with a probability of 1- Specifically, it is expressed as: （25）； （26）； in This represents the action selected in the i-th step. This represents the probability of a greedy strategy choosing an action. This represents the original Q-value in the Q-network. right The linear decreasing strategy is specifically expressed as follows: （27）； in This represents the state at the i-th step. value, Indicates the largest value, Represents the smallest value, This represents the value of the current step. This represents the total number of steps.

8. The movable antenna optimization system in a federated vehicular network based on differential privacy according to any one of claims 1-7, characterized in that, include: Construction Module (1): Used to construct a base station compensation cost and communication energy consumption optimization system in a federated vehicle network based on differential privacy. It establishes an objective function that minimizes base station compensation cost and communication energy consumption based on the optimal additive noise allocated by the base station to vehicle users and the location of the intelligent mobile antenna. The construction module (1) also includes a communication model, a calculation model, a compensation cost model, and a mobility model for vehicle users. Additive noise allocation module (2): used to obtain the optimal additive noise allocated by the base station to the vehicle user, and to solve the objective function and optimize the parameters using the Karush-Kuhn-Tucker condition to obtain the optimal additive noise for the vehicle user, and to use the result for the antenna position optimization module (3). Antenna position optimization module (3): used to obtain the current optimized antenna position, and to solve and optimize the objective function using the Deep Q-learning algorithm in reinforcement learning and the results of the additive noise allocation module (2) to obtain the optimized intelligent mobile antenna position and the optimal communication energy consumption.

Images