A federated learning method and system for vehicle networking heterogeneous data

Abstract

The application discloses a federated learning method and system for Internet of Vehicles heterogeneous data, and relates to the technical field of intelligent transportation. The method comprises the following steps: a roadside unit sends a current global model and a global control variable to a vehicle participating in training; the vehicle performs local training combined with a client drift correction mechanism through local data and a differential privacy stochastic gradient descent algorithm, and generates a local model update; the vehicle calculates a noise multiplier for the next round of training through an individualized adaptive differential privacy strategy according to a loss value of the current round of training; the vehicle judges whether the current round of model update is effective based on selective uploading; if the update is effective, the vehicle uploads a model increment and a local control variable increment to the roadside unit, and the roadside unit updates the global model and the global control variable after aggregation. The application innovates from three aspects of privacy protection, model optimization and communication strategy, and provides a solution for constructing an efficient and safe and reliable federated learning in an Internet of Vehicles scenario.

Description

Technical Field

[0001] This application relates to the field of intelligent transportation technology, and in particular to a federated learning method and system for heterogeneous data in the Internet of Vehicles. Background Technology

[0002] With the development of vehicle-to-everything (V2X) and intelligent transportation technologies, an intelligent transportation system based on vehicle-road-cloud collaboration is gradually taking shape. The multi-source, heterogeneous data generated during vehicle operation is key to improving traffic perception and decision-making intelligence. However, centralized data processing faces the dual challenges of enormous communication overhead and user privacy breaches. Federated learning, as a distributed machine learning paradigm, allows vehicles to train models locally, sharing only model parameters, offering a new approach to solving this problem.

[0003] However, applying federated learning directly to the field of connected vehicles still faces significant challenges, as existing technologies struggle to achieve a good balance between privacy protection, model utility, and communication efficiency.

[0004] One issue is the contradiction between rigid privacy protection mechanisms and model accuracy. To defend against privacy inference attacks that may occur during model updates, existing solutions typically introduce differential privacy techniques, adding random noise to the model updates to reduce the risk of privacy leaks. However, mainstream methods employ fixed or centralized noise addition strategies, which cannot dynamically adjust the strength of privacy protection based on the characteristics of individual vehicle data and the training phase. This leads to over-protection of ordinary vehicles in the early stages of model training, severely slowing down convergence, while insufficient protection may be provided for sensitive vehicles in the later stages, making it difficult to achieve a fine-grained trade-off between privacy and utility.

[0005] Secondly, there is the client drift problem caused by highly heterogeneous data. Vehicle data has non-independent and identically distributed characteristics, causing the local model update direction to deviate significantly from the global optimum, i.e., client drift. Existing mitigation methods have limited correction capabilities or introduce complex parameter tuning, resulting in poor robustness in the dynamic environment of vehicle-to-everything (V2X) networks, leading to decreased global model accuracy and unstable convergence.

[0006] Thirdly, communication resources are consumed by ineffective or negative updates. Frequent model uploads and aggregations during multi-round federated training can impose a significant communication burden. Some vehicles, due to poor data quality, excessive privacy noise, or severe local drift, may negatively contribute to global aggregation during model updates, impacting system efficiency. To improve privacy protection capabilities...

[0007] It is evident that existing solutions often fall into a dilemma when balancing privacy, utility, and communication, sacrificing one to preserve another. Applying fixed, strong noise to meet stringent privacy requirements inevitably slows down model convergence and impairs final accuracy; while simple client filtering or compression to improve communication efficiency may amplify client drift caused by data heterogeneity, leading to the global model deviating from the optimal solution. These coupled challenges make it difficult for existing technologies to provide a systematic and efficient solution in the dynamic, heterogeneous, and resource-constrained scenario of connected vehicles. Summary of the Invention

[0008] This application provides a federated learning method and system for heterogeneous data in the Internet of Vehicles (IoV) to address the problems in the prior art.

[0009] On the one hand, embodiments of this application provide a federated learning method for heterogeneous data in the Internet of Vehicles, including:

[0010] S1, the roadside unit sends the current global model and global control variables to the vehicles participating in the training;

[0011] S2, based on the current global model and global control variables, the vehicle performs local training using local data and differential privacy stochastic gradient descent algorithm, combined with client drift correction mechanism, to generate local model updates;

[0012] S3, the vehicle calculates the noise multiplier for the next training round based on the loss value of this round of training through a personalized and adaptive differential privacy (PADP) strategy;

[0013] S4, the vehicle determines whether the model update is valid based on the selective upload rules;

[0014] S5. If the model update is effective in this round, the vehicle uploads the model increment and the local control variable increment to the roadside unit. The roadside unit aggregates the model increment and the local control variable increment and updates the global model and global control variables.

[0015] Furthermore, generating local model updates includes:

[0016] Randomly sample the local dataset to obtain training batches;

[0017] Based on the current global model, differential privacy stochastic gradient descent algorithm is executed on the training batch for federated learning to obtain gradient parameters with differential privacy protection.

[0018] By utilizing the client-side drift correction mechanism, the gradient is corrected based on the gradient parameters, global control variables, and locally stored local control variables to obtain the local model update.

[0019] Furthermore, the local model update is calculated using the following formula: Where t is the current round, and t+1 is the update round. For vehicle indexing, Update the local model. For the current global model, For hyperparameters, The gradient after adding noise. For the current local control variable, This is the current global control variable.

[0020] Furthermore, the noise multipliers used for the next round of training are calculated through a personalized adaptive differential privacy strategy, including:

[0021] Calculate the loss value of the local model update in the training batch during this training;

[0022] Based on the loss value, the noise multiplier used in the next round of training is obtained through an adaptive noise multiplier calculation formula.

[0023] Furthermore, the adaptive noise multiplier calculation formula is as follows:

[0024] Where t+1 is the training epoch for local model updates, and t+2 is the next training epoch. For vehicle indexing, For the noise multiplier in the next round of training, Noise multipliers updated for the local model The loss value updated for the local model. For hyperparameters, The maximum loss value, This is the minimum threshold for the noise multiplier.

[0025] Furthermore, the selective upload rules include:

[0026] A second random sampling is performed on the local dataset to obtain the validation batch;

[0027] Calculate the accuracy of the current global model and local model updates on the validation batch, and calculate the difference in loss between the current global model and local model updates;

[0028] Define the update formula for the local control variable:

[0029] Where t is the current round, and t+1 is the update round. For vehicle indexing, For the updated local control variables, For the current local control variable, For the current global control variable, For hyperparameters, This represents the total number of vehicles participating in the training. Update the local model. This is the current global model;

[0030] If the difference in loss values ​​is less than a preset threshold or the accuracy of the local model update is higher than the accuracy of the current global model, then the update is considered valid and proceeds according to the update formula. Update the local control variables and the noise multipliers used in the next training round separately, calculate the model increment and the local control variable increment, and upload them to the roadside unit; if the difference in loss values ​​is greater than or equal to a preset threshold and the accuracy of the local model update is lower than or equal to the accuracy of the current global model, then follow the update formula in... Update local control variables item by item, keep noise multipliers unchanged, and retain local information.

[0031] Furthermore, updating the global model and global control variables includes:

[0032] Global model update formula: Where t is the current round, and t+1 is the update round. For vehicle indexing, For the current global model, For global model updates, For the set of valid vehicles uploaded, For the kth car, For model increments;

[0033] Global control variable update formula: ,in, For updating global control variables, For the current global control variable, This is the increment of the local control variable.

[0034] On the other hand, this application also provides a federated learning system for heterogeneous data in the Internet of Vehicles, including: roadside units and vehicles.

[0035] The roadside unit is used to send the current global model and global control variables to the vehicles participating in the training, aggregate the model increment and the local control variable increment, and update the global model and global control variables.

[0036] The vehicle is used to perform local training based on the current global model and global control variables, using local data and a differential privacy stochastic gradient descent algorithm combined with a client drift correction mechanism to obtain local model updates. Based on the loss value of this round of training, a noise multiplier for the next round of training is calculated through a personalized adaptive differential privacy strategy. Based on the selective upload rule, it is determined whether the model update of this round is valid. If the model update of this round is valid, the model increment and the local control variable increment are uploaded to the roadside unit.

[0037] The federated learning method and system for heterogeneous data in the Internet of Vehicles (IoV) disclosed in this application have the following advantages:

[0038] By employing a personalized adaptive differential privacy strategy, each vehicle can dynamically adjust the noise multiplier size for the next training round based on its real-time loss value from local training. This achieves optimal allocation of the privacy budget throughout the entire training cycle, fundamentally overcoming the drawbacks of static strategies and significantly improving the final accuracy and convergence speed of the global model while maintaining equal privacy protection. A client drift correction mechanism is introduced to accurately estimate and correct local update direction deviations caused by differences in data distribution. Simultaneously, in control variable updates, valid and invalid updates are distinguished and processed differently. This dual correction mechanism ensures that, in a highly heterogeneous environment, the local updates of each vehicle remain aligned with the global objective, significantly mitigating client drift and thus improving the convergence stability and generalization performance of the global model. Furthermore, a selective upload strategy after distinguishing between valid and invalid updates drastically reduces unnecessary communication rounds and data transmission volume, improving the system's operational efficiency and scalability in dynamic vehicular network environments.

[0039] As can be seen, this application has made systematic innovations in three aspects: privacy protection mechanism, model optimization algorithm and system communication strategy, thereby achieving a breakthrough balance between privacy protection, model utility and communication efficiency, and providing a complete solution for building an efficient, safe and reliable federated learning system in the context of vehicle networking. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A flowchart illustrating the method provided in this application embodiment.

[0042] Figure 2 This is a schematic diagram of the overall system architecture provided for an embodiment of this application.

[0043] Figure 3 The accuracy evaluation graphs for the proposed method and other federated learning algorithms on the MNIST, SVHN, and GTSRB datasets under the most privacy-preserving conditions are shown.

[0044] Figure 4 The graph shows the accuracy evaluation of the proposed method and other federated learning algorithms on the MNIST dataset under different concentration parameters.

[0045] Figure 5 This is a statistical chart showing the number of times model parameters were uploaded to the MNIST dataset for the method proposed in this application.

[0046] Figure 6 Evaluation graphs show the effectiveness of the personalized adaptive differential privacy strategy validated on the MNIST and GTSRB datasets for the method proposed in this application. Detailed Implementation

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

[0048] Figure 1 This application provides a method flowchart for embodiments of the present application. Embodiments of this application provide a federated learning method for heterogeneous data in the Internet of Vehicles (IoV), including:

[0049] S1, the roadside unit sends the current global model and global control variables to the vehicles participating in the training.

[0050] For example, such as Figure 1 As shown, the roadside unit distributes the current global model and global control variables to each vehicle, and the vehicle set... , For index, i.e. For the first in the vehicle set A car. Passing by. After rounds of federated learning training, the current global model and global control variables are obtained as follows: and At the start of each training round, the roadside unit will use the current global model. and global control variables Distributed to all vehicles The vehicle then uses local data for model training.

[0051] S2, based on the current global model and global control variables, the vehicle performs local training using local data and differential privacy stochastic gradient descent algorithm, combined with a client drift correction mechanism, to generate local model updates.

[0052] In one possible implementation, generating local model updates includes:

[0053] Randomly sample the local dataset to obtain training batches;

[0054] Based on the current global model, differential privacy stochastic gradient descent algorithm is executed on the training batch for federated learning to obtain gradient parameters with differential privacy protection.

[0055] By utilizing the client-side drift correction mechanism, the gradient is corrected based on the gradient parameters, global control variables, and locally stored local control variables to obtain the local model update.

[0056] In one possible implementation, the local model update is calculated using the following formula: Where t is the current round, and t+1 is the update round. Update the local model. For hyperparameters, The gradient after adding noise. This is the current local control variable.

[0057] For example, firstly, the vehicle Your own dataset Based on size The first random sampling is performed using the sampling probability to obtain the training batch. Given a gradient clipping threshold After gradient calculation is complete, use the gradient clipping threshold. Gradient clipping is performed using the following formula: ,in, The gradient obtained from the calculation, This represents the gradient after clipping. Let i be the i-th sample in the dataset.

[0058] Then, to the clipped gradient Add standard deviation as Gaussian noise is expressed by the formula: ,in, The gradient after adding noise. The training batch size is the number of samples used to calculate the gradient in each training iteration. Let the random variable represent Gaussian noise. This is a d-dimensional identity matrix used to generate noise of the same dimension as the gradient. Subsequently, a local model update is generated using the local model update formula. .

[0059] S3, the vehicle calculates the noise multiplier for the next training round based on the loss value of the current training round through a personalized adaptive differential privacy strategy.

[0060] In one possible implementation, the noise multiplier for the next round of training is calculated using a personalized adaptive differential privacy strategy, including:

[0061] Calculate the loss value of the local model update in the training batch during this training;

[0062] Based on the loss value, the noise multiplier used in the next round of training is obtained through an adaptive noise multiplier calculation formula.

[0063] In one possible implementation, the adaptive noise multiplier calculation formula is as follows:

[0064] Where t+1 is the training epoch for local model updates, and t+2 is the next training epoch. For the noise multiplier in the next round of training, Noise multipliers updated for the local model The loss value updated for the local model. For hyperparameters, The maximum loss value, This is the minimum threshold for the noise multiplier.

[0065] For example, firstly, in the training batch Calculate the loss value Update the loss value for the current training round. With maximum loss value The difference is obtained by subtracting the values. In this application, the loss value obtained from the first round of model training for each vehicle is defined as the maximum loss value for that vehicle. By taking the maximum value of the two noise multipliers, the loss is prevented from decreasing too quickly and is kept within a controllable range. In the initial stage of model training, the loss value is relatively large, so a larger noise multiplier is used, i.e., a larger Gaussian noise is added to the gradient parameters to protect data privacy. As training progresses, the model gradually converges. At this time, the loss value decreases, and the noise multiplier also gradually decreases, i.e., a smaller noise is added to the gradient parameters to make the model converge as quickly as possible. A minimum threshold for the noise multiplier is set in the formula. This provides a maximum privacy budget cap for the vehicle. The loss value decreases as training progresses, and when the calculated noise multiplier value is less than a minimum threshold, then... .

[0066] S4, the vehicle determines whether the model update is valid based on the selective upload rules.

[0067] One possible implementation includes selective upload rules:

[0068] A second random sampling is performed on the local dataset to obtain the validation batch;

[0069] Calculate the accuracy of the current global model and local model updates on the validation batch, and calculate the difference in loss between the current global model and local model updates;

[0070] Define the update formula for the local control variable:

[0071] ,in, For the updated local control variables, For hyperparameters, The total number of vehicles participating in the training;

[0072] If the difference in loss values ​​is less than a preset threshold or the accuracy of the local model update is higher than the accuracy of the current global model, then the update is considered valid and proceeds according to the update formula. Update the local control variables and the noise multipliers used in the next training round separately, calculate the model increment and the local control variable increment, and upload them to the roadside unit; if the difference in loss value is greater than or equal to a preset threshold and the accuracy of the local model update is lower than or equal to the accuracy of the current global model, then follow the update formula in... Update local control variables item by item, keep noise multipliers unchanged, and retain local information.

[0073] For example, vehicles For dataset Perform random sampling again, with a batch size of .according to calculate and In the verification batch The loss value and accuracy. and The difference between the loss values ​​is obtained by subtracting the loss values. .like (The difference in loss values ​​is less than a preset threshold) or If the accuracy of the local model update is higher than the accuracy of the current global model, then the current learning round is considered effective. Based on the foregoing This section is updated, and the model increment is calculated. and local control variable increment Then, upload them together to the roadside unit; if and If the difference in loss values ​​is greater than or equal to a preset threshold and the accuracy of the local model update is lower than or equal to the accuracy of the current global model, the current learning round is deemed invalid, and the local control variable is... Based on the aforementioned This section will be updated; the noise multiplier in the next round will not use the aforementioned adaptive noise multiplier calculation formula, but instead... That is, it remains unchanged, and the local vehicle model is not uploaded. Then keep them locally in the vehicle.

[0074] S5. If the model update is effective in this round, the vehicle uploads the model increment and the local control variable increment to the roadside unit. The roadside unit aggregates the model increment and the local control variable increment and updates the global model and global control variables.

[0075] One possible implementation involves updating the global model and global control variables, including:

[0076] Global model update formula: ,in, For global model updates, For the set of valid vehicles uploaded;

[0077] Global control variable update formula: ,in, Update global control variables.

[0078] At this point, one round of federal learning training is complete.

[0079] like Figure 3 The figure shows the method proposed in this application and other federated learning algorithms on the MNIST, SVHN, and GTSRB datasets, demonstrating the highest level of privacy protection. Accuracy evaluation graph under certain conditions, showing the most privacy-preserving federated learning algorithms on the MNIST, SVHN, and GTSRB datasets. Comparative experiments were conducted under the condition of minimizing the privacy budget. The method proposed in this application is FedPADP, and the comparative methods include ALI-DPFL and VAN-DPFL. Under strong privacy constraints, due to the introduction of large differential privacy noise, the model performance of each comparative method decreased to varying degrees, and the convergence speed and final accuracy during training were significantly affected. In contrast, FedPADP still maintains a higher model accuracy, demonstrating its robustness and effectiveness under strict privacy constraints. Its final accuracy on the MNIST, SVHN, and GTSRB datasets reached 96.38%, 85.56%, and 78.52%, respectively, all significantly better than the comparative algorithms. Experimental results show that FedPADP can effectively mitigate the adverse effects of large noise on model training and still maintain good stability under strict privacy constraints. It can be seen that the method of this invention can balance privacy security and model utility under strong privacy protection conditions, fully demonstrating its comprehensive advantages and application value in the privacy protection scenario of federated learning.

[0080] Figure 4 This is an accuracy evaluation graph of the proposed method and other federated learning algorithms on the MNIST dataset under different concentration parameters. The three different concentration parameters are... These correspond to scenarios with strong heterogeneity, moderate heterogeneity, and weak heterogeneity, respectively. When the concentration parameter... At that time, all algorithms exhibited significant accuracy fluctuations during the initial training phase; however, FedPADP still achieved a faster performance improvement, ultimately reaching an accuracy of 96.70%, demonstrating its strong robustness and convergence ability. When set to 1, FedPADP's accuracy curve is smoother and achieves higher accuracy than both VAN-DPFL and ALI-DPFL. When the value is increased to 10, all mechanisms converge rapidly, but FedPADP still maintains the highest accuracy. Overall, this application demonstrates excellent accuracy and convergence under varying degrees of data heterogeneity. It is evident that the client drift correction mechanism introduced in this application can effectively mitigate the utility decline caused by data heterogeneity, thereby improving the overall utility of the model.

[0081] Figure 5The figure shows the statistics of model parameter upload times on the MNIST dataset using the method proposed in this application. To verify the effectiveness of the proposed selective upload rule, this application conducted a statistical analysis of the number of local model transmissions. In 30 communication rounds, compared with the FedPADP(non) mechanism without the selective upload rule, FedPADP significantly reduced the number of times vehicles uploaded model parameters to roadside units. The results show that after introducing the selective upload rule, most vehicles actually participated in significantly fewer than 30 communication rounds under the FedPADP method. Only 17 local model uploads were performed, whereas without the selective upload rule, all vehicles would have needed to participate in a full 30 rounds of communication. These results clearly demonstrate that the selective upload rule can effectively reduce communication frequency while maintaining model performance, thereby significantly reducing system communication overhead.

[0082] like Figure 6 The figure shows an evaluation graph of the effectiveness of the proposed method on the MNIST and GTSRB datasets, validating the personalized adaptive differential privacy strategy. The comparison algorithm is FedPADP (fixed), which represents adding fixed noise and not implementing the PADP policy. Figure 6 (a) and Figure 6 (b) The accuracy variations of FedPADP and FedPADP(fixed) on the MNIST and GTSRB datasets are presented respectively. It can be seen that FedPADP achieves an accuracy of 98.60% on the MNIST dataset and 88.61% on the GTSRB dataset. This is because FedPADP integrates the PADP strategy, which can adaptively adjust the noise multiplier according to the model performance changes of different vehicles and privacy requirements. Therefore, the experimental results verify the superiority of FedPADP in adaptively adjusting the noise multiplier, thus enabling more flexible and refined privacy protection in heterogeneous data scenarios.

[0083] Figure 2 This is a schematic diagram of the overall system architecture provided in an embodiment of this application. This embodiment of the application also provides a federated learning system for heterogeneous data in the Internet of Vehicles, including: roadside units and vehicles.

[0084] The roadside unit is used to send the current global model and global control variables to the vehicles participating in training, aggregate the model increments and local control variable increments uploaded by valid vehicles, and update the global model and global control variables.

[0085] The vehicle is used to perform local training based on the current global model and global control variables, using local data and a differential privacy stochastic gradient descent algorithm combined with a client drift correction mechanism, to obtain local model updates. Based on the loss value of this round of training, a personalized adaptive differential privacy strategy is used to calculate the noise multiplier for the next round of training. Based on the selective upload rule, it is determined whether the model update of this round is valid. If the model update of this round is valid, the model increment and the local control variable increment are uploaded to the roadside unit.

[0086] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0087] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A federated learning method for heterogeneous data in the Internet of Vehicles (IoV), characterized in that, include: S1, the roadside unit sends the current global model and global control variables to the vehicles participating in the training; S2, based on the current global model and the global control variables, the vehicle performs local training combined with a client drift correction mechanism using local data and a differential privacy stochastic gradient descent algorithm to generate a local model update; S3, the vehicle calculates the noise multiplier for the next training round based on the loss value of the current training round using a personalized adaptive differential privacy strategy; S4, the vehicle determines whether the current model update is valid based on the selective upload rules; S5, if the model update is effective in this round, the vehicle uploads the model increment and the local control variable increment to the roadside unit, and the roadside unit aggregates the model increment and the local control variable increment to update the global model and global control variables; The calculation of the noise multiplier for the next round of training using a personalized adaptive differential privacy strategy includes: Calculate the loss value of the local model update in the training batch during this training; Based on the loss value, the noise multiplier used in the next round of training is obtained through an adaptive noise multiplier calculation formula; The adaptive noise multiplier calculation formula is as follows: Where t+1 is the training epoch for local model updates, and t+2 is the next training epoch. For vehicle indexing, For the noise multiplier in the next round of training, Noise multipliers updated for the local model The loss value updated for the local model. For hyperparameters, The maximum loss value, This is the minimum threshold for the noise multiplier.

2. The federated learning method for heterogeneous data in the Internet of Vehicles (IoV) according to claim 1, characterized in that, The generation of local model updates includes: Randomly sample the local dataset to obtain training batches; Based on the current global model, federated learning is performed on the training batch using a differential privacy stochastic gradient descent algorithm to obtain gradient parameters with differential privacy protection. By utilizing the client-side drift correction mechanism, the gradient is corrected based on the gradient parameters, the global control variables, and the locally stored local control variables to obtain the local model update.

3. The federated learning method for heterogeneous data in the Internet of Vehicles (IoV) according to claim 1, characterized in that, The local model update is calculated using the following formula: Where t is the current round, and t+1 is the update round. For vehicle indexing, Update the local model. For the current global model, For hyperparameters, The gradient after adding noise. For the current local control variable, This is the current global control variable.

4. The federated learning method for heterogeneous data in the Internet of Vehicles according to claim 1, characterized in that, The selective upload rules include: A second random sampling is performed on the local dataset to obtain the validation batch; Calculate the accuracy of the current global model and the local model update on the validation batch, respectively, and calculate the difference in loss values ​​between the current global model and the local model update; Define the update formula for the local control variable: Where t is the current round, and t+1 is the update round. For vehicle indexing, For the updated local control variables, For the current local control variable, For the current global control variable, For hyperparameters, This represents the total number of vehicles participating in the training. Update the local model. This is the current global model; If the difference in loss values ​​is less than a preset threshold or the accuracy of the local model update is higher than the accuracy of the current global model, then the current update is deemed valid, and the update is performed according to the update formula. Update the local control variables and the noise multipliers used in the next training round separately, calculate the model increment and the local control variable increment, and upload them to the roadside unit; if the difference in the loss value is greater than or equal to a preset threshold and the accuracy of the local model update is lower than or equal to the accuracy of the current global model, then follow the update formula in... Update local control variables item by item, keep noise multipliers unchanged, and retain local information.

5. A federated learning method for heterogeneous data in the Internet of Vehicles (IoV) according to claim 1, characterized in that, The updating of the global model and global control variables includes: Global model update formula: Where t is the current round, and t+1 is the update round. For vehicle indexing, For the current global model, For global model updates, For the set of valid vehicles uploaded, For the kth car, For model increments; Global control variable update formula: ,in, For updating global control variables, For the current global control variable, This is the increment of the local control variable.

6. A federated learning system for heterogeneous data in the Internet of Vehicles, characterized in that, The system comprises a roadside unit and a vehicle, as described in any one of claims 1-5. The roadside unit is used to send the current global model and global control variables to the vehicles participating in the training, aggregate the model increment and the local control variable increment, and update the global model and global control variables. The vehicle is used to perform local training based on the current global model and global control variables, using local data and a differential privacy stochastic gradient descent algorithm combined with a client drift correction mechanism, to obtain local model updates; based on the loss value of this round of training, it calculates the noise multiplier for the next round of training through a personalized adaptive differential privacy strategy; based on selective upload rules, it determines whether the current round of model updates is valid; if the current round of model updates is valid, it uploads the model increment and the local control variable increment to the roadside unit; The calculation of the noise multiplier for the next round of training using a personalized adaptive differential privacy strategy includes: Calculate the loss value of the local model update in the training batch during this training; Based on the loss value, the noise multiplier used in the next round of training is obtained through an adaptive noise multiplier calculation formula; The adaptive noise multiplier calculation formula is as follows: Where t+1 is the training epoch for local model updates, and t+2 is the next training epoch. For vehicle indexing, For the noise multiplier in the next round of training, Noise multipliers updated for the local model The loss value updated for the local model. For hyperparameters, The maximum loss value, This is the minimum threshold for the noise multiplier.

Images