A mobile perception federated learning-based data privacy protection method for internet of vehicles

Abstract

The present application relates to a kind of mobile perception federation learning-based data privacy protection method of Internet of Vehicles, this method realizes data privacy protection in Internet of Vehicles from two aspects, one aspect is to select part of suitable vehicles to participate in training task by considering vehicle mobility characteristics, improve model training efficiency, reduce the risk of system privacy disclosure;The other side is to enhance the privacy protection strength of the federal learning mechanism itself by optimal noise disturbance to the information uploaded by the vehicle.Compared with the prior art, the present application considers the vehicle mobility, optimizes the communication overhead and learning efficiency under provable vehicle user-level privacy constraints, effectively reduces the risk of privacy disclosure, enhances the privacy protection strength of FL, ensures that the data privacy of vehicles in Internet of Vehicles scenario is not disclosed, meets the needs of low latency, high communication efficiency and privacy protection in Internet of Vehicles scenario, and has high practicability.

Description

Technical Field

[0001] This invention relates to the field of vehicle network data security, and in particular to a method for protecting vehicle network data privacy based on mobile perception federated learning. Background Technology

[0002] Vehicle-to-everything (V2X) communication, leveraging next-generation mobile communication technologies, enables vehicles to communicate with each other and with roadside infrastructure such as pedestrians, traffic management equipment, and parking lots. V2X can collect real-time traffic flow data, effectively transmit traffic data, alleviate traffic congestion, and contribute to improving the intelligence level of transportation services, making it a crucial component of future intelligent transportation systems. To achieve intelligent driving, V2X requires the transmission of vast amounts of traffic data perceived by vehicles. This data contains a wealth of sensitive information, such as vehicle IDs, locations, speeds, and information about surrounding vehicles. While V2X enables efficient data transmission and use, it also introduces the risk of privacy breaches. Privacy protection is paramount for V2X data applications, and in recent years, research on privacy protection for intelligent vehicles has received significant attention from numerous scholars and research institutions.

[0003] Currently, privacy protection methods mainly include cryptography, anonymization mechanisms, and federated learning (FL) methods. While these methods have proven effective in protecting privacy, to achieve better privacy protection in vehicle-to-everything (V2X) intelligent application scenarios with high data transmission volumes and low latency requirements, privacy protection strategies need to balance transmission efficiency and privacy protection strength, and the strategies should not be overly complex to the point of affecting their feasibility. Among these, FL is considered a promising technology that can effectively solve privacy protection issues and provide intelligent application services for V2X.

[0004] The mechanism by which federated learning achieves privacy protection is that vehicles do not need to upload the collected or generated real-time data to the server for centralized processing. Instead, they only need to upload the parameters of the local model trained in a distributed manner based on local data. This keeps the original data within the vehicle and prevents vehicle data from being exposed on the server, thus avoiding privacy data leakage and achieving privacy protection of vehicle data in the Internet of Vehicles scenario.

[0005] In existing applications, there are very few methods for privacy protection design using federated learning in connected vehicles, and the methods for achieving privacy protection mainly rely on the characteristic that raw data in federated learning does not leave the local machine to prevent privacy leaks. Current research on data privacy protection methods for connected vehicle scenarios only considers vehicle data privacy from a data perspective, without considering the privacy leakage risks brought about by the mobility of vehicles.

[0006] Chinese patent application CN202110445916.8 discloses a federated learning model aggregation method in a vehicle-to-everything (V2X) scenario. This method requires a stable and seamless wireless communication link between vehicles and fog nodes, sending raw data to fog nodes for training, which poses a privacy risk. Furthermore, this method only considers the privacy characteristics of federated learning itself and the privacy of vehicle identity, without fully considering the challenges posed by vehicle mobility to privacy protection. On the one hand, intelligent connected vehicles in V2X scenarios are highly dynamic; some move frequently, while others occasionally disconnect from the network (e.g., unstable communication links due to mobility, multiple Road Side Unit (RSU) switching, and network unavailability at specific times / locations). On the other hand, collecting a large number of local training model parameters from intelligent vehicles in moving states results in significant communication overhead and latency, all of which affect the learning performance of federated learning, such as training efficiency and accuracy.

[0007] In summary, current methods for protecting data privacy in the Internet of Vehicles (IoV) lack consideration for vehicle mobility. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a vehicle network data privacy protection method based on mobile perception federated learning. This method considers the vehicle's mobility characteristics when selecting target vehicles for local training, thereby improving model training efficiency while reducing the risk of system privacy leakage.

[0009] The objective of this invention can be achieved through the following technical solutions:

[0010] This invention provides a data privacy protection method for vehicle-to-everything (V2X) networks based on mobile perception federated learning, applied to roadside units in V2X networks. This method protects the data privacy of each vehicle while training the model using federated learning. The method includes the following steps:

[0011] Step S1: Receive global model parameters from the cloud server;

[0012] Step S2: Send requests to each mobile vehicle within the coverage area and obtain the returned resource information and mobility information;

[0013] Step S3: Based on the resource information and mobility information, select a target vehicle that has available data and can complete the local model training process and model update upload process, and send the current global model parameters to the target vehicle;

[0014] Step S4: Obtain vehicle model parameters trained on the local dataset of the target vehicle, and update the global model parameters by weighted aggregation of the vehicle model parameters;

[0015] Step S5: Repeat steps S2-S4 until the preset first stop condition is met.

[0016] As a preferred technical solution, the obtained local model parameters are the local model parameters after adding Gaussian noise.

[0017] As a preferred technical solution, the following are also included:

[0018] Step S6: Upload the local global model parameters to the cloud server, and update the global model parameters on the cloud server through weighting and aggregation. Then repeat steps S1-S5 until the preset second stop condition is met.

[0019] Step S7: Obtain the final trained global model parameters.

[0020] As a preferred technical solution, the first stopping condition is:

[0021] The communication rounds between the roadside unit and the moving vehicle have reached the preset first round threshold.

[0022] The second stopping condition is:

[0023] The number of communication rounds between the cloud and the roadside unit has reached the preset second round threshold.

[0024] As a preferred technical solution, the process of selecting a target vehicle that has available data and can complete the local model training and model update upload processes includes the following steps:

[0025] Obtain data quality loss evaluation indicators and spatiotemporal importance indicators for each vehicle, and filter out vehicles that meet the preset conditions.

[0026] For the selected vehicles, based on the resource information and mobility information, the optimal wireless bandwidth and the optimal computing power evaluation index are obtained, and a vehicle selection optimization problem is constructed.

[0027] The vehicle selection optimization problem is solved to obtain target vehicles that can complete the local model training process and the model update and upload process.

[0028] As a preferred technical solution, a greedy algorithm is used to solve the vehicle selection optimization problem.

[0029] As a preferred technical solution, the optimal computing power evaluation index is obtained using the following formula:

[0030] f k * =min{f k max ,fk}

[0031]

[0032] Among them, f k * f is the best evaluation index of the computational capability of vehicle k. k max λ is the maximum local computing power of vehicle k, λ is the number of local iterations for vehicle k, and D is the maximum local computing power of vehicle k. k This is the number of data samples for vehicle k. v is the number of CPU cycles required for vehicle k to process one data sample. k b is the current speed of vehicle k. k It is the bandwidth of vehicle k, p k It is the average transmission power of vehicle k, g k Where N is the channel gain between vehicle k and RSU, N0 is the power spectral density of Gaussian noise, P is the coverage area of ​​RSU, and C is the channel gain between vehicle k and RSU. k s is the current position of vehicle k, and s is the size of the data transmitted by vehicle k to the RSU.

[0033] As a preferred technical solution, the process of solving the vehicle selection optimization problem includes the following steps:

[0034] Based on the resource information obtained from each vehicle, determine its dwell time within the current RSU range. By judging whether the dwell time of a vehicle within the current RSU is greater than or equal to the time delay required to complete local model training and model upload, iteratively find target vehicles whose local computing and wireless transmission time matches the dwell time.

[0035] As a preferred technical solution, the data quality loss evaluation index is used to reflect the degree of data quality degradation caused by vehicle mobility and the accuracy of the vehicle-mounted camera. The data quality loss evaluation index is obtained based on camera exposure time, focal length, the angle between the vehicle's direction of motion and the image plane, the pixel size of the charge-coupled device in the horizontal direction, and the instantaneous relative speed between the vehicle and the observed object.

[0036] The aforementioned spatiotemporal importance index is used to represent the frequency of vehicle appearance within the cloud server's coverage area.

[0037] As a preferred technical solution, the resource information includes the computing power evaluation index of the mobile vehicle, the number of local data samples of the mobile vehicle, and the number of CPU cycles required to process one data sample. The mobility information includes the position and speed of the mobile vehicle.

[0038] Compared with the prior art, the present invention has the following advantages:

[0039] (1) High model training efficiency and reduced system privacy leakage risk: This method is applied to roadside units. First, it sends requests to each mobile vehicle within the coverage area to obtain the returned resource information and movement information. Then, it combines the resource information and movement information to determine the target vehicle and sends the model parameters to the target vehicle. Finally, it obtains the model parameters after training on the vehicle locally to realize federated learning. This method selects suitable vehicles to participate in the training task based on the vehicle movement characteristics and excludes vehicles that are not suitable for training. While improving the model training efficiency, it reduces the system privacy leakage risk. It optimizes communication overhead and learning efficiency under provable vehicle user-level privacy constraints, effectively reduces the privacy leakage risk, enhances the privacy protection strength of federated learning, and ensures that the data privacy of vehicles in the Internet of Vehicles scenario is not leaked. It meets the requirements of low latency, high communication efficiency and privacy protection in the Internet of Vehicles scenario and has high practicality.

[0040] (2) After the vehicle is trained locally, noise is added and uploaded to the roadside unit to further ensure privacy and security: The vehicle trains on local data and adds noise to update the local model, which can defend against various attacks from adversaries and various advanced inferences from untrusted servers during information transmission, and prevent model data leakage. From the experimental simulation of the design, the performance index improvement can be verified, and it has a high degree of privacy protection.

[0041] (3) High deployment efficiency: This method adopts a layered aggregation strategy, which is more efficient than the traditional cloud-edge architecture, more in line with the infrastructure deployment in actual vehicle networking application scenarios, and has better privacy performance. Attached Figure Description

[0042] Figure 1 This is a flowchart of the vehicle-to-everything (V2X) data privacy protection method based on mobile-aware federated learning in Example 1;

[0043] Figure 2 A flowchart for selecting a strategy for dynamically adaptive vehicles;

[0044] Figure 3 System architecture diagram;

[0045] Figure 4 This is a performance comparison chart of Example 1 on the MNIST dataset under different privacy budgets;

[0046] Figure 5 This is a performance comparison chart of Example 1 on the CAFIR-10 dataset under different privacy budgets. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0048] Example 1

[0049] like Figure 1-3 This embodiment provides a data privacy protection method for vehicle-to-everything (V2X) networks based on mobile-aware federated learning. This method achieves data privacy protection in V2X networks from two aspects: firstly, by considering the vehicle's movement characteristics, it selects a subset of suitable vehicles to participate in the training task, thereby improving model training efficiency and reducing the risk of system privacy leakage; secondly, it enhances the privacy protection strength of the federated learning mechanism itself by performing optimal noise perturbation on the information uploaded by vehicles.

[0050] like Figure 3 The vehicle-to-everything (V2X) architecture of this embodiment includes a cloud server located at a regional center, N Roadside Units (RSUs) servers, and multiple intelligent connected vehicles traveling on the road. In the considered system architecture, the intelligent connected vehicles, RSU servers, and cloud servers are equipped with storage units, computing units, and communication units, respectively. Considering the communication environment, vehicle mobility, and the size of transmitted data, wireless transmission is specified between the vehicles and the RSU servers, while wired transmission is specified between the RSUs and the cloud servers. The system has a three-layer architecture, specifically including a cloud server located at a regional center, several Roadside Units (RSUs) located on both sides of the road, and a collection of... This indicates that a large number of intelligent vehicles k are in motion, represented by a set. The RSU server acts as an intermediary between vehicles and cloud servers, communicating with mobile vehicles within its coverage area via Vehicle-to-Infrastructure (V2I) communication to collect local models trained by vehicles using local data.

[0051] like Figure 1 The method includes the following steps:

[0052] Step S1: System initialization; specifically: the cloud server determines the specific task, specifies the completion of an intelligent application model, and initializes the global model w and global iteration rounds for that task;

[0053] Step S2: The cloud server sends a task request to the RSUs within its coverage area and broadcasts the initialized global model parameters w and the global iteration round to N RSUs.

[0054] Step S3: RSUs respond to the cloud server's request and send data to K local datasets within the communication coverage area. The moving vehicle sends a resource request; specifically: y is the input vector feature of a single data sample of vehicle k. kj These are the labels corresponding to the features.

[0055] Step S4: The vehicle k that receives the request sends its data sample number D. k bandwidth b k Computational ability f k The number of CPU cycles required to process one data sample Current position P and velocity C k The information is returned to the relevant RSU;

[0056] Step S5: Based on the resource information provided by the vehicles in Step S4, the RSUs determine the vehicles within their coverage area that meet the learning task requirements through a dynamic adaptive vehicle selection strategy, thus forming a set of vehicles participating in the task.

[0057] The following section introduces a dynamic adaptive vehicle selection method to achieve vehicle selection. From the perspectives of data availability and data feasibility, it selects a set of vehicles that can complete local model training and model upload within the current RSU communication range. Participate in training missions.

[0058] The dynamic adaptive vehicle selection method in this embodiment is based on an established optimization problem. The steps for solving the optimization problem are as follows: Figure 2 As shown:

[0059] The optimization problem can be written in the following form:

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069] in, The optimization objective is to define the vehicles k, b = (b1, ..., b2) with good data availability and feasibility. K ), f = (f1, ..., f K ), h is a binary variable, and (C1) to (C8) are constraints;

[0070] This is the local computation time for vehicle k. It is the time when vehicle k uploads the local update, L k This represents the dwell time of a vehicle at the current RSU, where B is the total bandwidth, and the set... It is a collection of vehicles that meets the requirements for data availability and resources. It is the maximum computing power of vehicle k. The acceptable maximum data quality loss for vehicle k, Γ min Acceptable spatiotemporal importance of vehicle k information, h k It is a binary variable, if h k =1 indicates that vehicle k is selected to participate in the task and becomes part of the set. Part of h k =0 indicates that vehicle k does not meet the conditions set by the cloud server.

[0071] Data quality loss refers to the degradation in data quality caused by vehicle mobility and the accuracy of onboard cameras, specifically manifested as various forms of noise, motion blur, and distortion. Noise and distortion can affect the consistency of the statistical distribution across different vehicles, while motion blur, caused by the transient velocity of vehicle k, can be expressed as: Where τ is the camera exposure time, φ is the focal length, θ is the angle between the vehicle's direction of motion and the image plane, Z is the pixel size of the charge-coupled device in the horizontal direction, and μ is the instantaneous relative velocity between the vehicle and the observed object. The spatiotemporal importance of information refers to the frequency of occurrence of vehicle k within the cloud server's coverage area, expressed as: Among them, S k It is the actual spatiotemporal matrix of vehicle k, Θ k It is the threshold spatiotemporal matrix set by the cloud server, ||·|| F It is the Frobenius norm. In this invention, for the algorithm to function properly, the quality and spatiotemporal importance of the data collected by the vehicle are required.

[0072] Determine if the spatiotemporal importance of the data information meets the condition Γ k ≥Γ min and data quality conditions Vehicle k;

[0073] The above optimization problem is transformed into the following formula:

[0074]

[0075] stConstraint(C1),(C2),(C5),

[0076]

[0077] This optimization problem is a mixed-integer nonlinear programming problem, and the optimal solution cannot be obtained directly. The problem is decoupled into subproblems for solution, and the solution process is as follows:

[0078] A. Problem decoupling

[0079] a. Decouple the optimization problem into two sub-problems: one is to obtain the optimal wireless bandwidth. k and computing resources f k The other is vehicle selection (h). k ;

[0080] b. Given the vehicle selection variable h, jointly optimize the optimal wireless bandwidth b and computational resource f, thus transforming the optimization problem into,

[0081]

[0082] stConstraint(C1),(C2),(C5),

[0083]

[0084] c. Since the problem is still non-convex, it is decoupled into two sub-problems: optimization of wireless bandwidth b and computational resource f, which are solved separately.

[0085] B. Optimal computing power

[0086] a. Given wireless bandwidth b, this problem is a linear programming problem with respect to computing power f;

[0087] b. According to constraint (C1), the vehicle's computational capability f can be obtained by the following formula.

[0088]

[0089] Where λ is the number of local iterations for vehicle k, and D k This is the number of data samples for vehicle k. v is the number of CPU cycles required for vehicle k to process one data sample. k b is the current speed of vehicle k. k It is the bandwidth of vehicle k, p k It is the average transmission power of vehicle k, g kWhere N is the channel gain between vehicle k and RSU, N0 is the power spectral density of Gaussian noise, P is the coverage area of ​​RSU, and C is the channel gain between vehicle k and RSU. k s is the current position of vehicle k, and s is the size of the data transmitted from vehicle k to the RSU;

[0090] c. The optimal computing power can then be calculated.

[0091]

[0092] C. Optimal wireless bandwidth

[0093] a. With the acquired computational power f, the optimization problem can be rewritten as follows:

[0094]

[0095] stConstraint(C1),(C2)and(C'6).

[0096] b. This optimization problem is convex, satisfies the Karush-Kuhn-Tucker conditions, and its Lagrangian function can be expressed by the following formula.

[0097]

[0098] Where γ is the Lagrange multiplier associated with constraint (C2), Regarding variable b k The first derivative is obtained through calculation.

[0099]

[0100] c. The optimal computing power can then be calculated using the following formula.

[0101]

[0102] Among them, b k (γ) is The solution,

[0103]

[0104] in, It is the maximum transmission power of vehicle k; γ satisfies

[0105] D. Vehicle Selection

[0106] a. Lagging vehicles will affect the efficiency of the entire task, preventing them from participating in the model training task; specifically, based on the obtained wireless bandwidth b and computing resources f, the optimization problem can be rewritten as follows:

[0107]

[0108]

[0109] b. A greedy algorithm is used to solve the problem. Here, the main purpose is to determine the dwell time of vehicle k within the current RSU range based on the resource information provided by vehicle k. By judging the relationship between the dwell time of vehicle k within the current RSU and the time delay required to complete local model training and model upload, the algorithm iteratively finds vehicles whose local computing and wireless transmission time meet the dwell time requirements.

[0110] c. Based on vehicle mobility characteristics, vehicle selection is a dynamic process. When a new vehicle enters the coverage area of ​​the current RSU server, the dynamic adaptive vehicle selection strategy will be implemented automatically.

[0111] In this step, RSUs determine suitable vehicles within their coverage area to participate in the training task. Specifically, this involves comprehensively considering multiple factors such as the availability and feasibility of mobile vehicle data in the vehicle-to-everything (V2X) scenario, and designing a dynamic adaptive resource scheduling strategy. By constraining the dwell time of mobile vehicles within the RSUs' coverage area, the optimal vehicle with available data that can complete the local model training and model update upload processes is selected to participate in the learning task. This achieves the target accuracy in as few communication rounds as possible, improves learning efficiency, and reduces the risk of privacy leakage.

[0112] Step S6: RSUs will distribute the initial global model parameters w received from the cloud server to the vehicles k selected in step S5, i.e., the set. Vehicles in the middle;

[0113] Step S7: Vehicle k is based on the local dataset Obtain local model updates with initial model parameters w.

[0114] In this embodiment, the local model parameters are obtained using the following formula.

[0115]

[0116] in, It is the loss function for the (j-th)th sample of vehicle k. It is all D of vehicle k k The loss function for each sample; finding the loss function that makes the loss function optimal through iteration. To reach the minimum

[0117] Step S8: Update the local model obtained for vehicle k. Add an appropriate amount of Gaussian noise to obtain the model after noise perturbation. In this step, vehicle k adds an appropriate amount of Gaussian noise to the local model update to obtain the noise-perturbed model, ensuring differential privacy protection for the parameters of the locally updated model; specifically, a theoretically optimal noise addition algorithm is given to perturb the local model update with additive Gaussian noise.

[0118] Step S9: The vehicle will adjust the local model parameters for noise disturbance. Uploaded to RSUs;

[0119] Step S10: RSUs use a weighted method to aggregate the received perturbation-adjusted model updates.

[0120] In this embodiment, it is assumed that there are n within each RSUi coverage area. i For each selected vehicle, the aggregation method used is coefficient-weighted aggregation. The updated model w after aggregation is then calculated using the following formula: i ,

[0121]

[0122] in, This is the number of data samples for vehicle k in RSUi. y is the input vector feature of a single data sample of vehicle k. kj These are the labels corresponding to the features.

[0123] Step S11: RSUs use the aggregated model w i Replace the initial global model parameter w and return to step S3; if the number of communication rounds between RSUi and vehicle k is equal to the preset threshold for the number of communication rounds between vehicle and RSUs, then RSUi will upload the model update obtained after this aggregation to the cloud server for global model aggregation update.

[0124] In this embodiment, the cloud service performs global model aggregation using the following formula to obtain the global model update w for the t-th communication round. t ,

[0125]

[0126] Step S12: The cloud server aggregates the model updates from RSUs using a coefficient weighting method to obtain the global model update; return to step S2; if the number of communication rounds between the cloud server and RSUs reaches the preset number of global iterations, the currently obtained global model parameters are determined as the final model.

[0127] In this embodiment, there are 3 RSUs within the cloud server coverage area, each with a coverage diameter of 2km, containing 30 vehicles traveling at different speeds. The lower and upper limits of vehicle speeds in the three lanes traveling in the same direction are (17,22), (22,27), and (27,33) m / s, respectively. The local model training iterations for the vehicles are 10, the communication round threshold between RSUs and vehicles is 5, and the communication round threshold between RSUs and the cloud server is 45. The privacy budgets to ensure differential privacy are ε = 3, ε = 5, and ε = 7, respectively. For the classification and recognition task in intelligent driving, a convolutional neural network model is used on two federated learning image classification datasets. The vehicle data is partitioned using a non-independent and identically distributed method. A comparative experiment between the method of this invention and the Fed-SPA method is given. Figure 4 and Figure 5 This is the simulation result under this condition.

[0128] The above embodiments overcome the impact of vehicle mobility on privacy leakage in federated learning-based data privacy protection methods in vehicle-to-everything (V2X) scenarios. They achieve improved learning efficiency, reduced privacy leakage risks, and enhanced privacy protection strength of the federated learning mechanism under verifiable user-level privacy constraints. Furthermore, the aforementioned intelligent connected vehicle data privacy protection method based on mobile perception federated learning not only provides real-time services for intelligent driving services but also defends against attacks from malicious vehicles in specific scenarios.

[0129] This invention takes into account vehicle mobility and optimizes communication overhead and learning efficiency under provable vehicle user-level privacy constraints, effectively reducing the risk of privacy leakage, enhancing the privacy protection strength of FL, and ensuring that vehicle data privacy is not leaked in the vehicle-to-everything (V2X) scenario. It meets the requirements of low latency, high communication efficiency and privacy protection in the V2X scenario and has high practicality.

[0130] By adopting a layered aggregation strategy, this model is more efficient than the traditional cloud-edge architecture, better suited to the infrastructure deployment in actual vehicle-to-everything (V2X) application scenarios, and offers better privacy protection.

[0131] Based on the federated learning mechanism, this study addresses the specific challenges to privacy protection posed by vehicle mobility in connected vehicle scenarios. It comprehensively considers important factors related to mobility and heterogeneity, such as vehicle location, speed, limited communication resources, and data heterogeneity. Through a comprehensive resource scheduling design, it achieves dynamic and adaptive optimal vehicle selection in mobile scenarios, improves the training efficiency of the federated learning algorithm (FL), reduces the number of communication rounds between vehicles and RSUs, reduces the risk of privacy leakage, and achieves better privacy protection.

[0132] Simultaneously, the vehicle trains on local data and updates the local model by adding noise, enabling it to defend against various attacks from adversaries and advanced inference from untrusted servers during information transmission, thus preventing model data leakage. Experimental simulations demonstrate that the performance improvements are verifiable. Therefore, this invention offers high levels of privacy protection.

[0133] This method optimizes communication overhead and learning efficiency under provable user-level privacy constraints, thereby reducing the risk of privacy leakage and enhancing privacy protection. Specifically, it designs a novel mobile-aware vehicle scheduling strategy to select suitable vehicles for the learning task, improving training efficiency and reducing the risk of privacy leakage. Furthermore, it enhances the privacy protection of the federated learning mechanism by providing user-level differential privacy guarantees for the selected participating vehicles.

[0134] Example 2

[0135] This embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for executing the vehicle-to-everything (V2X) data privacy protection method based on mobile perception federated learning as described in Embodiment 1.

[0136] Example 3

[0137] This embodiment provides a computer-readable storage medium including one or more programs executable by one or more processors of an electronic device, the one or more programs including instructions for performing the mobile-aware federated learning-based vehicle network data privacy protection method as described in Embodiment 1.

[0138] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for protecting data privacy in vehicle-to-everything (V2X) networks based on mobile-aware federated learning, characterized in that, A roadside unit applied in vehicle-to-everything (V2X) communication, used to protect the data privacy of each vehicle while performing federated learning training on a model, the method includes the following steps: Step S1: Receive global model parameters from the cloud server; Step S2: Send requests to each mobile vehicle within the coverage area and obtain the returned resource information and mobility information; Step S3: Based on the resource information and mobility information, select a target vehicle that has available data and can complete the local model training process and model update upload process, and send the current global model parameters to the target vehicle; Step S4: Obtain vehicle model parameters trained on the local dataset of the target vehicle, and update the global model parameters by weighted aggregation of the vehicle model parameters; Step S5: Repeat steps S2-S4 until the preset first stop condition is met; The process of selecting a target vehicle with available data and capable of completing both local model training and model update upload processes includes the following steps: Obtain data quality loss evaluation indicators and spatiotemporal importance indicators for each vehicle, and filter out vehicles that meet the preset conditions. For the selected vehicles, based on the resource information and mobility information, the optimal wireless bandwidth and the optimal computing power evaluation index are obtained, and a vehicle selection optimization problem is constructed. Solve the vehicle selection optimization problem to obtain target vehicles that can complete the local model training process and the model update and upload process. The optimal computing power evaluation index is obtained using the following formula: in, It is a vehicle The best evaluation index for computing power, It is a vehicle Maximum local computing power It is a vehicle The number of local iterations, It is a vehicle Number of data samples It is a vehicle The number of CPU cycles required to process one data sample. It is a vehicle The current speed, It is a vehicle bandwidth, It is a vehicle Average transmission power, It is a vehicle Channel gain between RSU and RSU It is the power spectral density of Gaussian noise. This refers to the coverage area of ​​the RSU. It is a vehicle Current location It is a vehicle The size of the data transmitted to the RSU.

2. The method for protecting vehicle network data privacy based on mobile perception federated learning according to claim 1, characterized in that, Also includes: Step S6: Upload the local global model parameters to the cloud server, and update the global model parameters on the cloud server through weighted aggregation. Then repeat steps S1-S5 until the preset second stop condition is met. Step S7: Obtain the final trained global model parameters.

3. The method for protecting vehicle network data privacy based on mobile perception federated learning according to claim 2, characterized in that, The first stopping condition is: The communication rounds between the roadside unit and the moving vehicle have reached the preset first round threshold. The second stopping condition is: The number of communication rounds between the cloud and the roadside unit has reached the preset second round threshold.

4. The method for protecting vehicle network data privacy based on mobile perception federated learning according to claim 1, characterized in that, The vehicle selection optimization problem is solved using a greedy algorithm.

5. A method for protecting vehicle network data privacy based on mobile perception federated learning according to claim 1, characterized in that, The process of solving the vehicle selection optimization problem includes the following steps: Based on the resource information obtained from each vehicle, determine its dwell time within the current RSU range. By judging whether the dwell time of a vehicle within the current RSU is greater than or equal to the time delay required to complete local model training and model upload, iteratively find target vehicles whose local computing and wireless transmission time matches the dwell time.

6. The method for protecting vehicle network data privacy based on mobile perception federated learning according to claim 1, characterized in that, The data quality loss evaluation index is used to reflect the degree of data quality degradation caused by vehicle mobility and the accuracy of the vehicle camera. The data quality loss evaluation index is obtained based on camera exposure time, focal length, the angle between the vehicle's direction of motion and the image plane, the pixel size of the charge-coupled device in the horizontal direction, and the instantaneous relative speed between the vehicle and the observed object. The aforementioned spatiotemporal importance index is used to represent the frequency of vehicle appearance within the cloud server's coverage area.

7. A method for protecting vehicle network data privacy based on mobile-aware federated learning according to claim 1, characterized in that, The resource information includes the computing power evaluation index of the mobile vehicle, the number of local data samples of the mobile vehicle, and the number of CPU cycles required to process one data sample. The mobility information includes the position and speed of the mobile vehicle.

Images