Federated forgetting method for internet of vehicles based on precise knowledge compensation

Abstract

The application discloses a kind of federated forgetting methods of Internet of Vehicles based on accurate knowledge compensation, entire Internet of Vehicles service area is divided into several regions according to the physical coverage range of road side unit, each vehicle continuously collects local data set in the driving process and continuously uploads state data to central server through road side unit, the state data records the area switching of vehicle, federated training and forgetting, in each federated training round, central server determines federated training vehicle set to carry out federated training, when there is vehicle to send forgetting request, central server carries out forgetting learning, introduce vehicle calibration and position perception compensation calibration based on knowledge distillation in forgetting learning, obtain new global model after forgetting learning.The application can realize efficient, accurate and verifiable federated forgetting under the premise of meeting privacy compliance requirements, and effectively alleviate the performance degradation of the region caused by vehicle data removal.

Description

Technical Field

[0001] This invention belongs to the field of vehicle networking technology, and more specifically, relates to a vehicle networking federated forgetting method based on precise knowledge compensation. Background Technology

[0002] As a core component of intelligent transportation systems, the Internet of Vehicles (IoV) provides crucial support for environmental perception, intelligent decision-making, and collaborative control through collaborative communication between vehicles, roadside units, the cloud, and other traffic participants. With the rapid development of IoV technology, vehicles, as edge nodes, continuously generate a large amount of spatiotemporal data on driving behavior, environmental perception, and user status. This data is not only privacy-sensitive but also exhibits strong spatiotemporal heterogeneity—due to differences in road structure, traffic flow characteristics, and driving patterns, the data distribution in different regions often shows significant regional specificity.

[0003] Federated learning, as a distributed machine learning paradigm, can collaboratively train a global model across multiple clients while protecting local data privacy, and has been widely researched and applied in the Internet of Vehicles (IoV). By allowing vehicles to train models locally and only upload model updates, federated learning effectively avoids the privacy risks associated with centralized transmission of raw data. However, the high dynamism of vehicles in the IoV (frequent vehicle entry / exit from the network) and the non-independent, identically distributed nature of data present new challenges to federated learning.

[0004] In practical applications of connected vehicles, users may exercise their "right to be forgotten" (such as the right stipulated by GDPR) at any time, or the system needs to isolate malicious vehicles in real time. This requires actively and explicitly removing the historical data contributions of specific clients from the global model—a process known as federated forgetting. Existing federated forgetting methods mainly face the following technical challenges: First, while simply using complete retraining can thoroughly achieve forgetting, its computational and communication costs are extremely high, making it difficult to meet the real-time requirements of connected vehicle scenarios. Second, existing approximate forgetting methods based on parameter backtracking, knowledge distillation, or Bayesian inference are mostly based on the assumption of independent and identically distributed data, failing to fully consider the challenges brought by the inherent strong spatiotemporal heterogeneity of connected vehicle data and the high dynamism of devices. Third, after removing vehicle data from specific areas or behavioral patterns, the model's performance in the target area often suffers severe degradation due to the lack of distribution-aware compensation strategies. Finally, the frequent movement of vehicles leading to changes in data ownership also makes it difficult for traditional forgetting methods to accurately trace and eliminate the historical contributions of target vehicles.

[0005] Therefore, how to achieve efficient, accurate, and regionally adaptable federal forgetting while meeting privacy compliance requirements has become a key technical problem that urgently needs to be solved in the field of connected vehicles. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a federated forgetting method for vehicle networks based on precise knowledge compensation. This method introduces vehicle calibration and location awareness compensation calibration into the forgetting learning process, achieving efficient, accurate, and verifiable federated forgetting while meeting privacy compliance requirements, and effectively mitigating regional performance degradation caused by vehicle data removal.

[0007] To achieve the above-mentioned objectives, the present invention provides a federated forgetting method for vehicle networks based on precise knowledge compensation, comprising the following steps:

[0008] S1: Divide the entire vehicle-to-everything (V2X) service area into several zones based on the physical coverage of the roadside units;

[0009] S2: Each vehicle Continuously collect local datasets during driving. , It indicates the number of vehicles and continuously uploads status data to the central server through the roadside units. ,in, Indicates vehicle The historical area that last participated in federal training, Indicates vehicle The current region to which it belongs; This represents a forgetting counter used to record vehicle information. The number of forget requests that have been successfully executed; Indicates vehicle In historical areas The identifier for participation in federal training, with a default value of 0, is used when the vehicle... In historical areas Participation in federal training would lead to ; This represents the training counter, used to record vehicle data. In historical areas The number of times China participated in federal training;

[0010] S3: In each federal training round The central server determines the set of federally trained vehicles from all vehicles, and then processes the global model. The global model is distributed to each vehicle in the federated training vehicle set via roadside units. After receiving the global model, each vehicle trains on its local data and updates its local model via roadside units. Uploaded to a central server, the central server performs global aggregation to obtain a global model. And save the global model parameters and the local model update amount for each vehicle according to the preset round interval;

[0011] S4: When a vehicle To perform data forgetting, a forgetting request is sent to the central server. The central server then performs forgetting learning according to the following process:

[0012] S4.1: The central server filters out forgotten vehicles from the saved historical update information. The global model set that participated ,in Indicates historical cycle Update the obtained global model, , Indicates forgetting the vehicle The number of global models participated in;

[0013] S4.2: Forget the cycle The latest global model will be used as the teacher model. ;

[0014] S4.3: The central server never participated in any previous rounds. Selected from other vehicles in federal training One vehicle was used as a calibration vehicle, and then the teacher model was used. Send to each calibration vehicle , Each calibration vehicle Use the local model as the student model Teacher model For knowledge supervision, sample local datasets from local data. Perform knowledge distillation calibration training on the local model to obtain the local model update. And upload it to the central server;

[0015] S4.4: The central server calibrates each vehicle. Local model update volume Generate historical parameter update calibration values :

[0016] ,

[0017] in, Indicates historical cycle The historical global model update volume. This indicates the process of finding the norm;

[0018] The central server updates the calibration values ​​for all historical parameters. Aggregate to obtain the global model update amount Then update the current model to be forgotten to obtain the initial forgetting learning model. :

[0019] ;

[0020] S4.5: Retrieve forgotten vehicles from stored historical parameters In the cycle of history The local model gradient generated in Gradient reversal yields the forgetting learning model. :

[0021] ,

[0022] in, Indicates the reversal coefficient. ;

[0023] S4.6: The central server selects and forgets vehicles from other vehicles. The vehicles have similar spatiotemporal characteristics and the forgetting count does not exceed a preset threshold. One vehicle was used as a compensation vehicle, and then compensation calibration was performed. The specific method was as follows:

[0024] The central server will model the teacher and forgetting learning models Send to compensation vehicle , Each compensation vehicle Use the local model as the student model Teacher model For knowledge supervision, sample local datasets from local data. Perform knowledge distillation training on the local model to obtain the local model update. And upload it to the central server; the central server... Local model update volume for each compensation vehicle Aggregate the data to obtain the global model update amount after compensation training. This leads to the final forgetting learning model. :

[0025] ;

[0026] S4.7: Determine if If yes, proceed to step S4.8; otherwise, proceed to step S4.9.

[0027] S4.8: Order Teacher Model Return to step S4.3;

[0028] S4.9: Transfer historical rounds The final forgetting learning model obtained As a forgotten vehicle A new global model after forgetting learning.

[0029] This invention is based on a federated forgetting method for vehicle-to-everything (V2X) based on precise knowledge compensation. The entire V2X service area is divided into several regions according to the physical coverage of roadside units. Each vehicle continuously collects local datasets during its journey and continuously uploads status data to the central server through roadside units. The status data records the vehicle's region switching, federated training, and forgetting status. In each federated training round, the central server determines the set of vehicles for federated training. When a vehicle sends a forgetting request, the central server performs forgetting learning. In forgetting learning, vehicle calibration based on knowledge distillation and location-aware compensation calibration are introduced to obtain a new global model after forgetting learning.

[0030] The present invention has the following beneficial effects:

[0031] 1) This invention integrates historical parameter backtracking calibration and gradient reversal mechanism, accurately traces and offsets the historical contribution of the target vehicle without accessing the original data, and distributes the impact to multiple rounds of calibration through progressive gradient reversal, avoiding parameter mutation and model collapse, significantly accelerating the model reconstruction process, and improving the forgetting efficiency by about 3 times compared to complete retraining.

[0032] 2) This invention introduces a location-aware, precise knowledge compensation mechanism. Based on the spatiotemporal clustering results, online vehicles with similar behavior patterns are selected from the original region of the target vehicle for lightweight knowledge distillation training. This specifically compensates for the loss of regional knowledge caused by vehicle removal. Only a few training rounds are needed to restore the model's performance in the target region, effectively alleviating the regional performance degradation caused by forgetting operations.

[0033] 3) This invention achieves robust management of vehicle dynamic topology through vehicle status data design, supports contribution tracing and targeted processing of forgotten requests during area switching, and introduces a forgotten counting mechanism to prevent resource exhaustion attacks by malicious users, thereby enhancing the fairness and stability of the system in a dynamic vehicle network environment.

[0034] 4) When determining the set of federated training vehicles, this invention can proactively address the strong spatiotemporal heterogeneity of vehicle network data during the training phase through a spatiotemporally aware dynamic priority scheduling strategy. It prioritizes the selection of vehicles with high data quality and regional representativeness to participate in training, effectively improving the convergence speed and model performance of federated learning in heterogeneous data environments, and laying a high-quality foundation model and regional distribution information for the subsequent forgetting phase. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating a specific implementation of the federated forgetting method for vehicle networking based on spatiotemporal perception scheduling and precise knowledge compensation, as described in this invention.

[0036] Figure 2This is a schematic diagram of the state transition rules in this embodiment;

[0037] Figure 3 This is a flowchart of the spatiotemporal awareness scheduling in this embodiment;

[0038] Figure 4 This is a flowchart of data forgetting learning in this invention;

[0039] Figure 5 This is a comparison chart of the accuracy of the federated learning algorithm of the present invention under different Non-IID data partitions in this embodiment;

[0040] Figure 6 This is a comparison chart of the forgetting accuracy of the present invention and the comparison method under different tasks in this embodiment;

[0041] Figure 7 This is a performance comparison chart of the present invention under continuous forgetting in this embodiment;

[0042] Figure 8 This is a comparison chart of the forgetting time of the present invention and the comparison method under different tasks in this embodiment;

[0043] Figure 9 This is a supplementary training convergence speed diagram of the present invention and the comparison method under different tasks in this embodiment;

[0044] Figure 10 This is a comparison chart of MIA attack metrics of the present invention and the comparison method under different tasks in this embodiment. Detailed Implementation

[0045] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0046] Example

[0047] Figure 1 This is a flowchart illustrating a specific implementation of the federated forgetting method for vehicle networking based on precise knowledge compensation, as described in this invention. Figure 1 As shown, the vehicular network federated forgetting method based on precise knowledge compensation of the present invention includes the following steps:

[0048] S101: Service Area Division:

[0049] The entire vehicle-to-everything (V2X) service area is divided into several regions based on the physical coverage of Roadside Units (RSUs). Each region contains multiple vehicle clients, and the distribution of vehicle data will exhibit region-specific characteristics due to differences in geographical location and driving behavior patterns. For example, Region 1 is a city center area characterized by high vehicle density; Region 2 is a residential and school district characterized by high traffic volume during peak hours; Region 3 is a highway characterized by high vehicle speeds; and Region 4 is an industrial and logistics area characterized by a large number of large vehicles. In this invention, different regions collaboratively train the global model through federated learning to jointly improve the model's generalization ability, while maintaining the independence of each region's data distribution. In practical applications, region clustering can be performed based on data volume, i.e., the resulting regions are clustered... The amount of data in each region tends to be consistent within the error range.

[0050] S102: Vehicle data reporting:

[0051] Each vehicle Continuously collect local datasets during driving. , It indicates the number of vehicles and continuously uploads status data to the central server through the roadside units. ,in:

[0052] Indicates vehicle The last historical area involved in federal training, namely the vehicle The region to which the last valid data contribution originated. When a vehicle initiates a forgetting request, this field uniquely identifies the target model whose contribution needs to be removed.

[0053] Indicates vehicle The current region to which it belongs determines the data upload target for its participation in federal training.

[0054] This represents a forgetting counter used to record vehicle information. The number of successfully executed forget requests is used to assess behavioral loyalty and system resource usage.

[0055] Indicates vehicle In historical areas The identifier for participation in federal training, with a default value of 0, is used when the vehicle... In historical areas Participation in federal training would lead to .

[0056] This represents the training counter, used to record vehicle data. In historical areas The number of times China participated in federal training.

[0057] In practical applications, each vehicle The following state transition rules can be used to maintain state data. To adapt to dynamic environments. Figure 2 This is a schematic diagram of the state transition rules in this embodiment. For example... Figure 2 As shown, the state transition rules in this embodiment are as follows:

[0058] When the vehicle When a roadside unit registers for the first time, the historical area is initialized. and the current area All of these are areas where the roadside unit is located, and then the forgetting counter is initialized based on the current participation in federated training. Logo ,frequency And will be updated based on subsequent participation in federal training.

[0059] When the vehicle From the current area Switch to new region When, if its identifier Then let If its identifier Then let In other words, if a vehicle only switches regions and no memory is forgotten, its status information is selectively updated based on whether it has participated in federated training. If it has participated, the current region is updated to the new region while keeping the historical region unchanged; if it has not participated, both the historical region and the current region are updated to the new region.

[0060] When the vehicle If a forget request is initiated and executed successfully, then the state data is reset. In other words, when a vehicle successfully performs data forgetting, the "historical area" is reset to match the "current area," and the identifier is changed. A value of 0 indicates that past contributions have been cleared. If participation in subsequent federal training is required, new contribution history will begin accumulating from the current region. Meanwhile, the forgetting counter... Add 1, mark A record of 0 indicates that the current client did not contribute anything; this operation is recorded. This is the identifier. The purpose is to ensure that if a client has participated in training in a previous region, and its data changes after moving to a new region, it may respond to a request from the central server and participate in training again. Its data would then represent the new region, thus controlling that a vehicle can only contribute data to one region. This prevents the same vehicle from contributing differently in different regions, ensuring it only participates in contributing to one region. When receiving a training invitation again, it first determines... The number of times the activity is updated is 0. Increasing by 1 indicates that another round of training has been performed in the current region. When the value is 1, it checks whether the current region is consistent with the historical region. If they are consistent, it participates; otherwise, it does not. Simultaneously, the number of updates is determined based on whether participation is involved. The value of . The number of times successful forgetting occurs. It becomes 0. If the vehicle If a region is switched during the forgetting process, the topology update will be delayed until the end of the current forgetting round; that is, the forgetting counter will be updated first. Logo ,frequency Update the historical area again. and the current area This ensures consistency in contribution tracing.

[0061] S103: Federal Training

[0062] In each federal training round The central server determines the set of federally trained vehicles from all vehicles, and then processes the global model. The global model is distributed to each vehicle in the federated training vehicle set via roadside units. After receiving the global model, each vehicle trains on its local data and updates its local model via roadside units. Uploaded to a central server, the central server performs global aggregation to obtain a global model. The global model parameters and the local model update amount for each vehicle are saved according to a preset round interval (every two rounds in this embodiment).

[0063] In this embodiment, when performing global model aggregation, sample size weighted aggregation is adopted. That is, based on the proportion of the number of local training samples of each vehicle in this calibration training to the total number of samples, corresponding aggregation weights are assigned, and weighted federated average aggregation is performed.

[0064] To improve the effectiveness of federated training, this embodiment uses spatiotemporal awareness scheduling to determine the set of federated training vehicles, thereby dynamically prioritizing the vehicles during the training process. Figure 3 This is a flowchart of the spatiotemporal awareness scheduling in this embodiment. For example... Figure 3 As shown, the specific steps of spatiotemporal awareness scheduling in this embodiment include:

[0065] S301: Vehicle Clustering:

[0066] For each vehicle, its location area is determined based on its real-time location information. Since vehicles within a region have high consistency in terms of road structure and traffic flow, the region can be used as a spatial feature. Then, the temporal characteristics of the vehicle's behavior are obtained. In this embodiment, the temporal characteristics of the vehicle's behavior include historical trajectory characteristics, speed characteristics, and peak-hour activity levels. These characteristics can characterize the temporal features of the vehicle.

[0067] The vehicle's regional and behavioral temporal features are encoded to obtain the vehicle's spatiotemporal features. Based on these features, vehicles are clustered to obtain... There are 3 clusters, each representing a subgroup of vehicles with similar data distributions.

[0068] S302: Calculate vehicle-cluster center similarity:

[0069] Record each vehicle In the current round The spatiotemporal characteristics of the vehicle are For each cluster, the spatiotemporal features of all vehicles within it are averaged to obtain the cluster center features. , Then, the vehicle-cluster center similarity is calculated. ,in, This indicates the calculation of feature similarity. Indicates vehicle Cluster center characteristics of the cluster to which it belongs.

[0070] S303: Calculate the priority of vehicles in the current round:

[0071] The vehicle is calculated using the following formula. priority :

[0072] ,

[0073] in, Indicates the vehicles in the previous round priority, This represents the historical priority decay coefficient. . Indicates vehicle Local data is concentrated in the current round. The amount of data that has not yet participated in federal training. This indicates the preset balance weights. This is used to balance the impact of data size and vehicle data representativeness on priority. This indicates frequent participation in penalty weighting. Indicates the current round vehicle The dynamic update factor is calculated using the following formula:

[0074] ,

[0075] in, Indicates the current round Vehicles within the previously reserved time period The number of rounds a vehicle participates in federated training is determined based on the state data. If a vehicle has been participating frequently recently or has a small amount of data remaining, its priority is appropriately reduced to avoid some clients becoming overly involved while others are marginalized.

[0076] As can be seen, the vehicle priority in this embodiment takes into account both the historical continuity of priority and can be dynamically adjusted according to the latest status of the vehicle.

[0077] S304: Select Vehicle:

[0078] Sort all vehicles in descending order of priority, then iterate through each vehicle and check its forgetting counter. If the number of vehicles exceeds a preset threshold (3 in this embodiment), the vehicle is skipped; otherwise, it is added to the federated training vehicle set. This process continues until the number of vehicles in the federated training vehicle set reaches the preset value.

[0079] As can be seen from the above process, this embodiment explicitly introduces spatiotemporal clustering and dynamic priority scheduling during federated training, giving differentiated importance to vehicle data from different regions and behavioral patterns. This not only improves the convergence efficiency and model quality of federated training, but also lays the foundation for rapid screening and knowledge compensation of similar vehicles in the subsequent forgetting stage.

[0080] S104: Data Forgetting Learning

[0081] When a vehicle Data forgetting is required. The roadside unit sends a forgetting request to the central server, and the central server performs data forgetting learning on the global model.

[0082] Figure 4 This is a flowchart of data forgetting learning in this invention. For example... Figure 4 As shown, the specific steps of data forgetting learning in this invention include:

[0083] S401: Selecting a forgetting model:

[0084] The central server filters out forgotten vehicles from the saved historical update information. The global model set that participated ,in Indicates historical cycle Update the obtained global model, , Indicates forgetting the vehicle The number of global models that have participated in.

[0085] S402: Initialize forgotten parameters:

[0086] Forgetting the cycle The latest global model will be used as the teacher model. .

[0087] S403: Local Model Calibration

[0088] The central server has never participated in any historical rounds. Selected from other vehicles in federal training One vehicle was used as a calibration vehicle, and then the teacher model was used. Send to each calibration vehicle , Each calibration vehicle Use the local model as the student model Teacher model For knowledge supervision, sample local datasets from local data. Perform knowledge distillation calibration training on the local model to obtain the local model update. And upload it to the central server.

[0089] Generally, the number of calibration training epochs can be fewer than the number of training epochs for the local model in conventional federated learning. Knowledge distillation is a commonly used training method. Specifically, in this invention, for the same input, the teacher model... With student model The softened probability distributions are output separately. Then, the KL divergence is used to measure the distribution difference loss between the teacher model and the student model. The loss is then weighted and summed with the cross-entropy loss of the student model to obtain the joint loss. The local model parameters are then optimized using gradient descent based on this joint loss.

[0090] S404: Global Model Calibration

[0091] In step S403, a local model update is generated for each vehicle participating in the calibration. This update reflects the removal of forgotten vehicles. The direction in which the model parameters should be adjusted after the impact. The central server adjusts the parameters for each calibration vehicle. Local model update volume Generate historical parameter update calibration values :

[0092] ,

[0093] in, Indicates historical cycle The historical global model update volume. This indicates the norm calculation. The above formula replaces the direction of historical global model updates with historical parameter update calibration values. The direction of normalization.

[0094] To eliminate the influence of the target vehicle client on the direction of historical updates, the contributions of the remaining vehicle clients are preserved. The central server updates the calibration amount for all historical parameters. Aggregate to obtain the global model update amount Utilizing global model update volume The server updates the current model to be forgotten, thus obtaining the initial forgetting learning model. :

[0095] .

[0096] This process achieved an initial correction to the direction of historical parameters, significantly mitigating the impact of forgotten vehicles. Its historical impact.

[0097] S405: Gradient Inversion

[0098] Obtain the initially calibrated forgetting learning model Retrieve forgotten vehicles from stored historical parameters In the cycle of history The local model gradient generated in By introducing a negative gradient, the forgotten vehicle can be directly offset. The gradient-reversed forgetting learning model is derived by considering the update contributions to the global model during historical training iterations. :

[0099] ,

[0100] in, Indicates the reversal coefficient. .

[0101] By performing multiple rounds of gradient reversal, we can ensure that the historical contributions of forgotten vehicles are completely offset, and the final model almost completely removes the influence of the target vehicle, while significantly reducing the reconstruction time cost.

[0102] S406: Position Awareness Compensation Calibration:

[0103] The forgetting learning model obtained after calibration and gradient inversion This has significantly reduced the impact of forgotten vehicles. However, due to the removal of their contribution and the fact that the model has less data to train, its ability to represent the original region or related patterns of forgotten vehicles may be reduced compared to the global model before forgetting learning.

[0104] Therefore, after the forgetting operation is completed, in order to accurately compensate for the knowledge loss in specific areas caused by the departure of the vehicle, avoid model performance degradation (especially in the area corresponding to the original cluster of the target vehicle), and minimize the cost of supplementary training, the key to compensatory training lies in selecting suitable participating clients. The central server selects vehicles from other vehicles that are similar to the forgotten vehicle. Vehicles with similar spatiotemporal characteristics (similar geographical location, similar driving patterns) and forgot counts not exceeding a preset threshold. One vehicle was used as a compensation vehicle for compensation calibration.

[0105] The core objective of this mechanism is not to find a replacement for the target vehicle, but to maintain the statistical integrity of the data distribution in the region—by selecting clients with similar spatiotemporal contexts, it compensates for the knowledge gaps caused by removal at the distribution level. If the target vehicle was originally a statistical outlier in the region, its removal has a negligible impact on overall performance. In this case, the compensation mechanism will naturally guide the model to fit the mainstream distribution more robustly, which aligns with the original intention of the right to forget and demonstrates robustness to extreme cases. In addition, the "one vehicle, one region, one contribution" design significantly reduces the probability of extreme uniqueness.

[0106] The specific method for compensation calibration is as follows: the central server will use the teacher model and forgetting learning models Send to compensation vehicle , Each compensated vehicle Use the local model as the student model Teacher model For knowledge supervision, sample local datasets from local data. Perform knowledge distillation training on the local model to obtain the local model update. And upload it to the central server. The central server... Local model update volume for each compensation vehicle Aggregate the data to obtain the global model update amount after compensation training. This leads to the final forgetting learning model. :

[0107] .

[0108] Similarly, compensation training employs only mini-batch local iterations to reduce additional computational and energy consumption on in-vehicle equipment. This progressive compensation mechanism enables the model to gradually recover performance under continuous forgetting scenarios and avoids unstable updates caused by a one-time strong calibration. Through this location-aware, data distribution-matching targeted fine-tuning, this invention can effectively compensate for the loss of knowledge in specific areas caused by vehicle departure at minimal cost, allowing the forgetting model to maintain high performance on target area tasks.

[0109] S407: Determine if If yes, proceed to step S408; otherwise, proceed to step S409.

[0110] S408: Order Teacher Model Return to step S403.

[0111] S409: Determine the new global model:

[0112] Historical cycles The final forgetting learning model obtained As a forgotten vehicle A new global model after forgetting learning.

[0113] To better illustrate the technical effects of this invention, specific examples are used to experimentally verify it. This embodiment tests four traffic scenarios: rainy days, nighttime, foggy days, and snowy days. In this embodiment, the invention (denoted as FedCoffs) sets the federated learning model as an MLP model and a CNN model respectively, uses a spatiotemporal awareness algorithm to train the federated learning model, and after data forgetting, uses collaborative optimization to calibrate the model training, and uses the remaining dataset for accurate knowledge compensation. Then, three existing models are selected as comparison models for experimental comparison with this invention. The comparison models include:

[0114] FedRetrain: After deleting the data from the target client, the model is trained from scratch on the remaining dataset as a baseline model.

[0115] FedEraser: After deleting data from the target client, it performs round-by-round calibration updates using historical global parameters and the remaining local dataset.

[0116] FedAccum: Updates the global model by directly accumulating the local model parameters of the calibration client and using the accumulated updates.

[0117] This embodiment conducted comparative experiments on four datasets: MNIST, GTSRB, CIFAR-10, and Cityscapes. The experiment involved 100 vehicle clients, with 20 training rounds (10 epochs per round, 5 calibration rounds), and a knowledge distillation temperature of 2. Evaluation metrics included mAP (Mean Average Precision), accuracy, forgetting accuracy, forgetting time, member inference attack metrics, mIoU, and performance recovery rounds.

[0118] Figure 5 This is a comparison chart of the accuracy of the federated learning algorithm of the present invention under different Non-IID data partitions in this embodiment. Figure 5 As shown, in a typical Non-IID vehicle networking scenario, the accuracy of the MLP model of this invention reaches 84.06%, which is 13.62% higher than the 70.44% of FedAvg (uniform partitioning); the accuracy of the CNN model is 86.84%, which is 9.12% higher than the 77.62% of FedAvg.

[0119] Figure 6 This is a comparison chart of the forgetting accuracy of the present invention and the comparison method under different tasks in this embodiment. For example... Figure 6 As shown, the forgetting accuracy of this invention is 0.963 in the MNIST task, which is 4.3% higher than FedEraser (0.920); the forgetting accuracy is 0.940 in the GTSRB task, which is significantly better than FedEraser (0.924); and the forgetting accuracy is 0.594 in the CIFAR-10 task, which is better than FedRetrain (0.577).

[0120] Figure 7 This is a performance comparison chart of the present invention under continuous forgetting in this embodiment. For example... Figure 7 As shown, in scenarios where forgetting occurs continuously in different regions, the performance of other methods declines in stages, while the present invention can stably maintain the model performance at a level close to the initial model performance.

[0121] Figure 8 This is a comparison chart of the forgetting time of the present invention and the comparison method under different tasks in this embodiment. For example... Figure 8 As shown, the forgetting times of this invention on the three datasets are 44.04s, 57.80s and 42.86s, respectively, accounting for only 28.5%, 31.2% and 27.8% of the total reconstruction time, and the reconstruction speed is about three times that of FedRetrain.

[0122] Figure 9 This is a supplementary training convergence speed diagram of the present invention and the comparison method under different tasks in this embodiment. For example... Figure 9As shown, in the MNIST task, the present invention achieved an accuracy of 0.964 after only 2 rounds of supplementary training; in the CIFAR-10 task, it also achieved an accuracy of 0.507 in the second round and stabilized at 0.612 in the fifth round.

[0123] Figure 10 This is a comparison chart of MIA attack metrics for the present invention and the comparison method under different tasks in this embodiment. For example... Figure 10 As shown, the attack accuracy (0.4582-0.4725) and recall (0.52-0.5815) of this invention on all datasets are very close to FedRetrain, close to the level of random guessing (50%), and significantly better than FedEraser.

[0124] To verify the effectiveness of this invention in complex perception tasks and highly spatiotemporally heterogeneous data that more closely resemble real-world vehicle-to-everything (V2X) scenarios, this embodiment conducts systematic ablation experiments on the Cityscapes dataset. The natural inter-city scene differences in the Cityscapes dataset are used to simulate the distribution of different RSU coverage areas in V2X. The experiment employs the lightweight segmentation network Fast-SCNN, dividing clients according to road scene categories to construct a spatiotemporally heterogeneous federated learning setting, and focuses on evaluating the model's mIoU change on the validation set before and after forgetting, as well as the number of rounds required for performance recovery. As a control, the same ablation process was performed on the GTSRB and MNIST datasets to evaluate the impact of each module on time efficiency and model accuracy in classification tasks. Training hyperparameters and strategies were kept consistent across all datasets to ensure comparability. Table 1 shows the configuration table of the ablation strategy in this embodiment. Table 2 shows the ablation experiment results of this invention on the GTSRB dataset in this embodiment. Table 3 shows the ablation experiment results of this invention on the MNIST dataset in this embodiment. Table 4 shows the ablation experiment results of this invention on the Cityscapes dataset in this embodiment.

[0125] FULL √ √ √ √ w / o scheduling × √ √ √ w / o compensation √ × √ √ w / o progressive √ √ × √ w / o distillation √ √ √ ×

[0126] Table 1

[0127] FULL 284.5 62.8 0.952 0.948 w / o scheduling 254.2 69.4 0.937 0.924 w / o compensation 276.4 54.6 0.944 0.941 w / o progressive 274.5 57.5 0.956 0.944 w / o distillation 268.6 79.54 0.952 0.953

[0128] Table 2

[0129] FULL 224.26 44.04 0.984 0.974 w / o compensation 223.4 36.7 0.978 0.961

[0130] Table 3

[0131] FULL 3384 748 -0.89% 3 w / o compensation 3421 917 -2.47% 5 w / o scheduling 3116 846 -1.07% 3

[0132] Table 4

[0133] As shown in Table 2, this invention improves both time and accuracy on the GTSRB dataset. Table 3 also demonstrates that the compensation mechanism helps improve the stability of the model after forgetting. Table 4 shows a comparison between the complete method and two key variants. It can be seen that the complete method only causes a 0.89% decrease in mIoU after forgetting, and performance can be largely recovered after 3 rounds of compensated training; while the uncompensated variant experiences a 2.47% decrease and requires 5 rounds to recover, further verifying the necessity of the knowledge compensation mechanism in segmentation tasks. Although the scheduling mechanism has a limited impact on the final accuracy, it helps improve the convergence efficiency during the training phase. Within a reasonable parameter range, the model performance shows a gradual trend with parameter changes, indicating that the proposed method has good robustness to key hyperparameters.

[0134] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A federated forgetting method for vehicle-to-everything (V2X) networks based on precise knowledge compensation, characterized in that, Includes the following steps: S1: Divide the entire vehicle-to-everything (V2X) service area into several zones based on the physical coverage of the roadside units; S2: Each vehicle Continuously collect local datasets during driving. , It indicates the number of vehicles and continuously uploads status data to the central server through the roadside units; ,in, Indicates vehicle The historical area that last participated in federal training, Indicates vehicle The current region to which it belongs; This represents a forgetting counter used to record vehicle information. The number of forget requests that have been successfully executed; Indicates vehicle In historical areas The identifier for participation in federal training, with a default value of 0, is used when the vehicle... In historical areas Participation in federal training would lead to ; This represents the training counter, used to record vehicle data. In historical areas The number of times China participated in federal training; S3: In each federal training round The central server determines the set of federally trained vehicles from all vehicles, and then processes the global model. The global model is distributed to each vehicle in the federated training vehicle set via roadside units. After receiving the global model, each vehicle trains on its local data and updates its local model via roadside units. Uploaded to the central server, the central server performs global aggregation to obtain the global model. And save the global model parameters and the local model update amount for each vehicle according to the preset round interval; S4: When a vehicle To perform data forgetting, a forgetting request is sent to the central server. The central server then performs forgetting learning according to the following process: S4.1: The central server filters out forgotten vehicles from the saved historical update information. The global model set that participated ,in Indicates historical cycle Update the obtained global model, , Indicates forgetting the vehicle The number of global models participated in; S4.2: Forget the cycle The latest global model will be used as the teacher model. ; S4.3: The central server never participated in any previous rounds. Selected from other vehicles in federal training One vehicle was used as a calibration vehicle, and then the teacher model was used. Send to each calibration vehicle , Each calibration vehicle Use the local model as the student model Teacher model For knowledge supervision, sample local datasets from local data. Perform knowledge distillation calibration training on the local model to obtain the local model update. And upload it to the central server; S4.4: The central server calibrates each vehicle. Local model update volume Generate historical parameter update calibration values : ， in, Indicates historical cycle The historical global model update volume. This indicates the goal of finding the norm. The central server updates the calibration values ​​for all historical parameters. Aggregate to obtain the global model update amount Then update the current model to be forgotten to obtain the initial forgetting learning model. : ； S4.5: Retrieve forgotten vehicles from stored historical parameters In the cycle of history The local model gradient generated in Gradient reversal yields the forgetting learning model. : ， in, Indicates the reversal coefficient. ; S4.6: The central server selects and forgets vehicles from other vehicles. The vehicles have similar spatiotemporal characteristics and the forgetting count does not exceed a preset threshold. One vehicle was used as a compensation vehicle, and then compensation calibration was performed. The specific method was as follows: The central server will model the teacher. and forgetting learning models Send to compensation vehicle , Each compensation vehicle Use the local model as the student model Teacher model For knowledge supervision, sample local datasets from local data. Perform knowledge distillation training on the local model to obtain the local model update. And upload it to the central server; the central server... Local model update volume for each compensation vehicle Aggregate the data to obtain the global model update amount after compensation training. This leads to the final forgetting learning model. : ； S4.7: Determine if If yes, proceed to step S4.8; otherwise, proceed to step S4.

9. S4.8: Order Teacher Model Return to step S4.3; S4.9: Transfer historical rounds The final forgetting learning model obtained As a forgotten vehicle A new global model after forgetting learning.

2. The vehicle networking federated forgetting method according to claim 1, characterized in that, The region clustering partitioning in step S1 involves partitioning the region into regions. The amount of data in each region tends to be consistent within the error range.

3. The vehicle networking federated forgetting method according to claim 1, characterized in that, In step S2, the vehicle maintains its state data using the following state transition rules. : When the vehicle When a roadside unit registers for the first time, the historical area is initialized. and the current area All of these are areas where the roadside unit is located, and then the forgetting counter is initialized based on the current participation in federated training. Logo ,frequency And will be updated based on subsequent participation in federal training; When the vehicle From the current area Switch to new region When, if its identifier Then let If its identifier Then let ; When the vehicle If a forget request is initiated and executed successfully, then the state data is reset. ; If the vehicle When switching regions during the forgetting process, first update the forgetting counter. Logo ,frequency Update the historical area again. and the current area .

4. The vehicle networking federated forgetting method according to claim 1, characterized in that, The specific method for determining the set of federal training vehicles in step S3 is as follows: S3.1: For each vehicle, its location is determined based on its real-time location information. Then, the vehicle's behavioral temporal features are obtained. The vehicle's regional and behavioral temporal features are encoded to obtain the vehicle's spatiotemporal features. Based on the vehicle's spatiotemporal features, the vehicles are clustered to obtain... One cluster; S3.2: Record each vehicle In the current round The spatiotemporal characteristics of the vehicle are For each cluster, the spatiotemporal features of all vehicles within it are averaged to obtain the cluster center features. , Then, the vehicle-cluster center similarity is calculated. ,in, This indicates the calculation of feature similarity. Indicates vehicle Cluster center characteristics of the cluster to which it belongs; S3.3: The vehicle is calculated using the following formula. priority : ， in, Indicates the vehicles in the previous round priority, This represents the historical priority decay coefficient. ; Indicates vehicle Local data is concentrated in the current round. The amount of data that has not yet participated in federal training; This indicates the preset balance weights. ; This indicates frequent participation in penalty weighting. Indicates the current round vehicle The dynamic update factor is calculated using the following formula: ， in, Indicates the current round Vehicles within the previously reserved time period The number of rounds of federal training to participate in is determined based on status data; S3.4: Sort all vehicles in descending order of priority, then iterate through each vehicle and check its forgetting counter. If the number of vehicles exceeds a preset threshold, skip the vehicle; otherwise, add the vehicle to the federal training vehicle set. Continue this process until the number of vehicles in the federal training vehicle set reaches the preset value.

5. The vehicle networking federated forgetting method according to claim 1, characterized in that, When a global model is aggregated, a sample-weighted aggregation method is used: based on the proportion of local training samples of each vehicle in this calibration training to the total samples, corresponding aggregation weights are assigned, and a weighted federated average aggregation is performed.

Images