A Data Privacy Preservation Method for Packet Negotiation Conditional Asynchronous Federated Learning in Vehicle-to-Everything (V2X)

Abstract

This invention relates to a data privacy protection method for conditional asynchronous federated learning in vehicle-to-everything (V2X) networks, which coordinates vehicle mobility and privacy protection. Addressing the issues of low training efficiency and privacy leaks caused by roadside units, it establishes an integrated mechanism: dwell time grouping, conditional triggering asynchronous aggregation, dynamic mask negotiation, and secret sharing reconstruction. Vehicles are grouped based on their dwell time in the roadside unit service area, building a relatively stable multi-round collaboration within each group. Aggregation is triggered when a certain number of vehicles in a group complete local training. A mask seed is generated using a pseudo-random function based on the shared key and training rounds within the group. Vehicles then generate random masks and superimpose them onto their local model parameters for local encryption. During the aggregation phase, a dynamic rotating mask cancellation mechanism is used to complete secure aggregation and suppress reconstruction inference. When only some vehicles participate in aggregation, a threshold-based secret sharing mechanism is used to reconstruct missing masks, ensuring the correctness and robustness of the aggregation.

Description

Technical Field

[0001] This invention relates to the field of vehicle networking technology, and in particular to a data privacy protection method for packet negotiation conditional asynchronous federated learning in vehicle networking. Background Technology

[0002] Vehicle-to-everything (V2X) communication relies on next-generation mobile communication and intelligent driving technologies to achieve vehicle-to-vehicle and vehicle-to-infrastructure (V2I) collaboration and real-time traffic data collection and transmission, providing the communication and data foundation for the training and deployment of intelligent service application models. Due to the high-speed movement and fluctuating scale of vehicles, wireless link conditions change rapidly, and vehicles move frequently, putting pressure on data exchange in terms of latency and collaboration overhead. In a centralized paradigm, raw data must be aggregated to a central server for processing, which is constrained by latency, privacy compliance, and single-point-of-failure bottlenecks, making it difficult to simultaneously meet the requirements of real-time performance, privacy, and collaboration. To address these constraints, distributed learning has become a feasible path; among them, federated learning, characterized by "data not leaving the source and only uploading local model parameters," is gradually becoming a key paradigm for intelligent V2X applications.

[0003] The privacy protection mechanism of federated learning lies in placing the training task at the vehicle terminal. The server only collects and aggregates local model parameters, and the original data can complete distributed collaboration without leaving the vehicle. In the highly dynamic environment of the Internet of Vehicles, existing federated learning still faces two core challenges: (1) System stability and timeliness. The high speed of vehicle movement causes fluctuations in link conditions, and vehicles are prone to disconnection or upload failure. Synchronous federated learning (FL) training requires waiting for lagging vehicles to be aggregated, resulting in significant long-tail latency; asynchronous FL training produces aggregation bias due to incomplete updates or uneven latency distribution, affecting convergence and accuracy. (2) Privacy security. Even without uploading the original data, the local model parameters still carry exploitable statistical information. Semi-honest servers can reconstruct individual vehicle privacy data through model parameter difference, model inversion, or member inference, increasing the risk of leakage. The highly dynamic characteristics of the Internet of Vehicles will amplify the above problems, requiring stronger and provable privacy protection under the constraints of low latency and high communication efficiency.

[0004] Existing technologies often enhance the protection of local model parameters in FL (Flexible Learning) through differential privacy, homomorphic encryption, or secure aggregation. However, they generally lack a collaborative design for mobility and privacy to address the highly dynamic characteristics of vehicular networks, making it difficult to simultaneously achieve convergence efficiency and privacy strength under conditions such as link instability and time lag in local model parameters. For example, Chinese patent application publication number CN119743249A discloses an asynchronous federated learning method for lightweight privacy protection in vehicular networks: before uploading local model parameters, vehicles add random numbers generated based on a key and training epochs to encrypt the local model parameters. The vehicle information, encrypted model parameters, and training epochs are then sent to the Roadside Unit (RSU). The RSU forwards the vehicle information and training epochs to the base station, which calculates the sum of random numbers based on the corresponding shared key and epoch and returns it. The RSU then decrypts the encrypted local model parameters and completes global model aggregation. However, this scheme does not adequately consider the impact of vehicle mobility on the encryption process. During training, vehicles may leave the RSU coverage area due to path termination, link interruption, or insufficient energy. During the aggregation phase, because the random number corresponding to the departing vehicle is missing, the local update with added random numbers cannot cancel the random numbers, which can easily lead to inconsistencies in summation and decryption failures, and trigger re-upload, resulting in increased transmission and computation overhead. At the same time, the sum of random numbers is calculated centrally by the base station and then returned to the RSU. The round trip between base stations introduces additional latency, which amplifies the risk of desynchronization and aggregation deviation in high mobility scenarios, thus affecting training efficiency and accuracy.

[0005] In summary, current solutions for using federated learning to protect data privacy in connected vehicles lack consideration for vehicle mobility and fail to achieve a synergistic design between mobility and privacy protection. Specifically: ① The coupling between privacy mechanisms and mobility is not properly addressed. For example, vehicle entry and exit alter the allocation and accounting processes of differential privacy cross-wheel privacy budgets, increasing the complexity of homomorphic encryption key negotiation, rotation, and management, and weakening the fault tolerance threshold and reliability of secure aggregation. ② The assumption of static vehicle terminals deviates from reality, assuming vehicles are stably online and can fully participate in the training and privacy processes, ignoring vehicle mobility. Based on these issues, it is necessary to present a federated learning method and system that synergistically considers mobility and provides verifiable privacy guarantees in connected vehicles, achieving synergistic optimization of privacy strength and learning efficiency under low latency and limited communication conditions. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a data privacy protection method for group negotiation conditional asynchronous federated learning in vehicle networking, so as to solve or partially solve the problems of existing solutions not fully considering vehicle mobility, high risk of privacy leakage, and unsatisfactory training effect.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] This invention provides a data privacy protection method for packet negotiation conditional asynchronous federated learning in vehicle-to-everything (V2X) systems, applicable to V2X systems including roadside units and vehicle terminals. The method includes the following steps:

[0009] Step S1: The roadside unit predicts the expected dwell time of each vehicle terminal in the coverage area based on the status information of each vehicle terminal, divides the vehicles into at least one vehicle group based on the similarity of the expected dwell time, and publishes the public key of the vehicle in the group to the vehicle terminals in the group.

[0010] Step S2: For each vehicle terminal in the vehicle group, a mask seed is randomly generated, encrypted using the public key of the vehicle in the group, and forwarded to other vehicle terminals in the same vehicle group through the roadside unit. The vehicle-to-everything (V2X) terminals in the vehicle group cooperate to calculate a set of random values ​​that sum to zero as the vehicle mask for each vehicle terminal.

[0011] Step S3: For each vehicle terminal in the vehicle group, its own vehicle mask is split into multiple secret shares through threshold secret sharing and securely distributed to other vehicle terminals in the same vehicle group, and the secret shares of other vehicle terminals in the same vehicle group are pre-stored.

[0012] In step S4, the roadside unit asynchronously sends the latest global model to the vehicle terminals within the same vehicle group. The vehicle terminals perform training based on the received global model and local data, and complete the calculation of local model parameters.

[0013] Step S5: For each vehicle terminal, encrypt the local model parameters based on the vehicle mask and send them to the roadside unit.

[0014] Step S6: After the encrypted model parameters received by the roadside unit reach the threshold, asynchronous global aggregation is performed and the vehicle mask is canceled to obtain a new global model.

[0015] Step S7: In the asynchronous global aggregation, in response to the current training time window closing and the number of vehicle terminals that have collected encrypted model parameters not reaching the threshold, the vehicle mask of the missing vehicle terminal is reconstructed by requesting the secret share stored by the vehicle terminals in the vehicle group.

[0016] As a preferred technical solution, the following are also included:

[0017] In step S8, the roadside unit broadcasts the new global model to the vehicle terminals. The vehicle terminals that participate in the training enter the next round of local training, while the vehicle terminals that do not participate in the training continue to train based on their existing version, forming a closed-loop process of continuous iteration.

[0018] Step S9: Repeat steps S1-S8 until the preset stopping condition is met to obtain the final trained global model parameters.

[0019] As a preferred technical solution, step S1 includes:

[0020] Step S101: The roadside unit calculates the expected dwell time of each vehicle based on the vehicle speed of each vehicle terminal and the path length of the navigation software.

[0021] Step S102: Based on the expected dwell time of each vehicle terminal, the newly added vehicle terminals in this round of training are grouped by K-means clustering.

[0022] Step S103: Construct a vehicle group instance based on the grouping results, and generate a group code, creation time, and expected lifespan, wherein the lifespan is calculated based on the expected dwell time and the number of vehicle terminals within the vehicle group;

[0023] Step S104: Based on the proportion of active vehicles and the expected remaining lifespan, the roadside unit evaluates each existing vehicle group and determines whether to disband the vehicle group based on preset disbandment conditions.

[0024] Step S105: Assign a public key for each vehicle group and publish it to the vehicle terminal.

[0025] As a preferred technical solution, step S2 includes:

[0026] Step S201: The vehicle terminal generates a set of random numbers locally as a mask seed, encrypts them based on the vehicle public key in the set, and uploads them to the roadside unit.

[0027] In step S202, the roadside unit forwards the encrypted mask seed to the vehicle terminals within the same vehicle group, and collaboratively calculates a set of random values ​​that sum to zero to obtain the vehicle mask for each vehicle terminal.

[0028] As a preferred technical solution, step S3 includes:

[0029] Step S301: The vehicle terminal generates a Lagrange interpolation polynomial of the same order locally, calculates the corresponding secret share based on the vehicle mask, encrypts it with the vehicle public key in the group, and uploads it to the roadside unit.

[0030] In step S302, the vehicle terminal securely distributes the secret share to other vehicle terminals in the same vehicle group through the roadside unit, so that each vehicle terminal pre-stores the secret share of other vehicle terminals in the same vehicle group.

[0031] As a preferred technical solution, in step S6, a threshold is calculated based on the number of active vehicles in the current group.

[0032] As a preferred technical solution, step S7 includes:

[0033] Step S701: In response to the expiration of the current training time window and the fact that the number of vehicle terminals that have collected encrypted model parameters has not reached the threshold, the roadside unit generates a list of vehicle terminals that are missing encrypted model parameters and broadcasts a vehicle mask reconstruction request within the vehicle group.

[0034] Step S702: Within the vehicle group, the secret share corresponding to the pre-stored missing vehicle terminal is summed locally. The summation result is encrypted using the vehicle public key and uploaded to the roadside unit. After receiving a secret share greater than or equal to the threshold, the roadside unit reconstructs the missing vehicle mask and cancels it, thereby realizing asynchronous global aggregation.

[0035] As a preferred technical solution, the process of obtaining the new global model in step S6 includes:

[0036] Step S601: Based on the time-delayed rounds of local model updates in the vehicle terminal, calculate the aggregate weight of the vehicle terminal using a nonlinear penalty function.

[0037] In step S602, the roadside unit aggregates the encrypted model parameters of the vehicle terminal based on the aggregated weights to obtain a new global model.

[0038] As a preferred technical solution, in step S3, the secret share is calculated using the following formula:

[0039]

[0040]

[0041] in, Vehicle terminal The Lagrange interpolation function, Vehicle identification number. Indicates vehicle Vehicle mask, For vehicle group The number of occupants in the vehicle, The first one randomly generated for the vehicle terminal Interpolation function coefficients, The secret share is sent to the vehicle terminal. And the secret share receiving vehicle terminal is The secret share.

[0042] As a preferred technical solution, in step S7, the vehicle mask for reconstructing the missing vehicle terminal is calculated using the following formula:

[0043]

[0044] in, For the number of vehicle groups. For vehicle group The sum of the mask values ​​of the vehicle terminals that did not upload local model parameters. Indicates connection with vehicle terminal The sum of related, but not received by the RSU secret shares. Indicates vehicle group Excluding vehicle terminals Any vehicle terminal outside the group. Using this additive homomorphic property, the sum of the mask values ​​of multiple vehicle terminals can be recovered at once, completing the reconstruction of the lost mask within the group and ensuring the correctness of the aggregation result when mask loss occurs.

[0045] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0046] (1) Reduce the risk of system privacy leakage and improve model accuracy: This method adds a random mask that satisfies the zero-sum constraint to the local model parameters on the vehicle side, and uses end-to-end public key encryption and signature transmission to make it impossible for the semi-honest RSU to recover the original update of any vehicle, effectively resisting inference attacks such as model inversion and member inference. The mask cancels out during aggregation within the group, and the aggregation result is equivalent to the real unmasked summation, avoiding statistical bias and maintaining training accuracy and convergence stability. In this method, in order to adapt to the frequent changes of vehicles participating in training in the Internet of Vehicles, a dynamic rotating vehicle group mask negotiation method is established by predicting the dwell time of moving vehicles and performing dynamic grouping. The mask of each vehicle is determined by group negotiation to achieve efficient and secure aggregation in multiple rounds. Considering the impact of vehicle dynamics, in order to avoid the failure of local model parameter transmission of some vehicles leading to errors in the aggregation result at the RSU end, a lost mask reconstruction method based on secret sharing is designed to effectively reconstruct the lost mask within each vehicle group, ensuring that the data privacy of vehicles in the Internet of Vehicles scenario is not leaked. This meets the requirements of low latency, high communication efficiency and privacy protection in the Internet of Vehicles scenario and has high practicality.

[0047] (2) High model training efficiency and good performance: This invention uses the "expected dwell time" of vehicles as a feature for prediction and dynamic grouping. Combined with conditional asynchronous aggregation triggered by time windows or thresholds, it can maintain group stability and training continuity under conditions of frequent vehicle entry and exit and link fluctuations, improve intra-group synchronization and resource utilization, and improve training efficiency. The estimation of vehicle dwell time is accurate. By using the similarity of expected dwell time for grouping, the stability is high when establishing training units, and the coordination between vehicles during the training phase is good. Overall, the training efficiency and model performance have been significantly improved.

[0048] (3) High deployment efficiency: This method adopts the deployment of actual roadside units of vehicles, 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 effect. Attached Figure Description

[0049] Figure 1 This is a schematic diagram illustrating the data privacy protection process for packet negotiation conditional asynchronous federated learning in the embodiment for vehicle-to-everything (V2X) communication.

[0050] Figure 2 This is a schematic diagram of the vehicle networking system in the embodiment;

[0051] Figure 3 This is a performance comparison chart of different schemes on the Vehicle-10 traffic dataset. Detailed Implementation

[0052] 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.

[0053] To address the aforementioned problems in the existing technology, this embodiment provides a data privacy protection method for group negotiation conditional asynchronous federated learning in vehicle-to-everything (V2X) systems. This method is applied to V2X systems including vehicle terminals and roadside units, aiming to consider the vehicle's mobility characteristics during the design of privacy protection strategies, reduce the risk of system privacy leakage, and improve model training efficiency.

[0054] like Figure 2 As shown, the architecture of the vehicle-to-everything (V2X) system includes a Roadside Unit (RSU) located at the center of the area and multiple vehicles traveling on the road. In the considered system architecture, the vehicles and the RSU are equipped with storage units, computing units, and communication units, respectively. Considering the communication environment, vehicle mobility, and the size of the transmitted data, wireless transmission is specified between the vehicles and the RSU server. The system has a two-layer architecture, specifically including a Roadside Unit (RSU) located at the center of the area and a large number of moving vehicles. Using sets This indicates that each vehicle completes model training based on local data and uploads its local model parameters. The RSU aggregates the received local models, completes global model aggregation, and distributes updates.

[0055] This method addresses the privacy threats and insufficient training efficiency of semi-honest roadside units (RSUs) in dynamic vehicle-to-everything (V2X) scenarios. It proposes: (1) an intelligent grouping mechanism based on the expected dwell time of vehicles, dividing vehicles with similar movement characteristics into stable cooperative groups to limit the impact of dynamic vehicle entry and exit on the training process, thereby improving group stability and cooperative efficiency. (2) a threshold-triggered conditional asynchronous secure aggregation strategy, initiating the aggregation process when the number of vehicles in a group completing local training reaches a preset threshold, thus reducing waiting latency while maintaining the real-time performance and accuracy of model training. (3) a secret-sharing mask reconstruction mechanism for dynamic environments, achieving secure model aggregation by combining intra-group mask perturbation, public-key encryption, and secret sharing, and introducing dynamically rotating masks to resist inference attacks by semi-honest RSUs; even when some vehicles fail to complete local model updates in time, mask calculation and recovery can still be completed, ensuring the correctness of aggregation. Under verifiable privacy constraints, this mechanism effectively reduces communication overhead, improves training efficiency and robustness, meets the requirements of low latency and high dynamic characteristics of V2X, and further enhances data privacy protection capabilities.

[0056] See Figure 1 The method in this embodiment mainly includes the following parts:

[0057] (1) Dynamic vehicle addition and grouping on the vehicle terminal side, and movement management of roadside units. Corresponding step S1.

[0058] (2) Mask generation and negotiation on the vehicle terminal side, and secure forwarding by the roadside unit. Corresponding steps S2-S3.

[0059] (3) Local training on the vehicle terminal side and safety aggregation of roadside units. Corresponding steps S4-S5.

[0060] (4) Mask reconstruction on the vehicle terminal side and mask cancellation on the roadside unit. Corresponding steps S6-S8.

[0061] The method in this embodiment specifically includes the following steps:

[0062] Step S1: The RSU uses the vehicle's (i.e., vehicle terminal's) movement status information (such as speed) as a basis. With preset path information The system predicts the expected dwell time of vehicles in the RSU service area, divides vehicles into relatively stable groups based on the similarity of expected dwell time, and publishes a set of public keys for vehicles within each group.

[0063] Specifically, step S1 may include steps S101-S105:

[0064] Step S101, the RSU determines the speed of each vehicle. and the route length information preset by navigation software Calculate the expected dwell time for each vehicle. The optimal expected dwell time is obtained using the following formula:

[0065]

[0066] Step S102: After obtaining the expected stay time, the vehicles are grouped using the K-means algorithm based on the expected stay time, resulting in relatively stable vehicle groups. The group numbers are denoted as follows: ,in This indicates the number of vehicle groups. Grouping takes into account the number of newly arriving vehicles, and the optimal number of groups is obtained using the following formula:

[0067]

[0068] in, This indicates the number of new vehicles added within the current time window. This means that newly added vehicles must be divided into at least one group to avoid having 0 vehicles in each group. This indicates a maximum of 3 groups, controlling for excessive subdivision and computational overhead.

[0069] Step S103: Create a vehicle group instance and record the group code ID and creation time. Initial vehicle set and expected lifecycle The expected lifespan is determined by the average expected dwell time of vehicles within the group. The expected lifespan is obtained using the following formula:

[0070]

[0071] in, Indicates the number is The vehicle group, This indicates the number of vehicles in the vehicle group.

[0072] Step S104: The RSU periodically assesses the status of each vehicle group to determine whether the group should remain in place or be disbanded. This determination is based on two independent criteria:

[0073] (1) Proportion of active vehicles: When the proportion of active vehicles in the group is lower than the threshold (usually 30%), the group is disbanded. The active vehicle proportion is obtained by the following formula:

[0074]

[0075] in, Indicates vehicle group The number of active vehicles in the system.

[0076] (2) Expected Remaining Lifetime: Disbandment is triggered when the expected remaining lifetime of the vehicle group is lower than a threshold (usually 10 seconds). The expected remaining lifetime is obtained using the following formula:

[0077]

[0078] in, Indicates vehicle group Expected remaining lifespan Indicates the vehicle group creation time. Indicates the current time of the vehicle group.

[0079] In step S105, when the preset disbanding conditions are met, the group status is set to "disbanded" and the aggregation request queue is cleared; the remaining members in the group are not forcibly removed, and their eligibility for subsequent reorganization and training is retained.

[0080] Step S2: Each vehicle in the group generates a set of random numbers locally as a mask seed; the random seed associated with that member is encrypted using the public key of the group member, and forwarded by the RSU within the group. Each vehicle calculates its own random value within the group, and together they calculate a set of random values ​​that sum to zero as the group mask, thereby determining the vehicle mask for this round group.

[0081] Specifically, step S2 may include steps S201-S202:

[0082] Step S201: The vehicle generates a set of random numbers locally as a mask seed, and uses the public keys of the group members. encryption.

[0083] Each vehicle in the vehicle group generates a pseudo-random number locally using a pseudo-random number generator. There are 1 mask seed, among which Indicates the number of vehicles in the group. Vehicles The generated mask seed is represented as:

[0084]

[0085] in, This indicates a surrounding number within the group to prevent out-of-bounds numbering. It is the group number Vehicles with the number [number] need to [do something]. The vehicle sent the seed, the vehicle Send the seeds to the front of the group in a circular order. Adjacent vehicles.

[0086] The vehicle uses these mask seeds via the group's public key After encryption, it is uploaded to the RSU, which then forwards it within the group. Public key encryption can be represented as: ,in This indicates that the mask seed is encrypted.

[0087] In step S202, the vehicle uploads a set of messages composed of all random mask seeds to the RSU, which then forwards the messages within the set.

[0088] Step S3: To address the failure of the zero-sum property caused by dynamic changes in group members, the vehicle splits its own mask into multiple secret shares according to the threshold secret sharing scheme and distributes them securely, while pre-storing the necessary shares of other members in the group; if a vehicle leaves or the link is interrupted during the aggregation process, the vehicles present will reconstruct the missing mask based on the shares they already hold, maintaining the zero-sum property and the correctness of the aggregation.

[0089] Specifically, step S3 includes steps S301-S302:

[0090] Step S301: The vehicle calculates the mask locally using the Lagrange interpolation function. The corresponding secret share is encrypted using the public key and uploaded to the RSU. The calculation process for the secret share is as follows:

[0091] Each vehicle A Lagrange interpolation polynomial of the same order will be generated locally for subsequent evaluation / reconstruction. The Lagrange interpolation function can be written in the following form:

[0092]

[0093] in, Indicates vehicle The mask, For vehicle group The number of members in a vehicle group is determined by the RSU during the vehicle grouping phase and assigned to the vehicles within that group. The first one randomly generated for the vehicle Interpolation function coefficients, It is the order of the Lagrange interpolation function, and also represents the threshold of the secret share required for Shamir's secret share to recover the original secret, which determines the number of secret shares required for the mask reconstruction process.

[0094] vehicle The mask is calculated using various forms of Lagrange interpolation generated locally. Corresponding secret share ,in Representing the sender of the secret share, This represents the recipient of the secret share. The generation of a secret share can be written in the following form:

[0095]

[0096] in, A mask secret share Send to all vehicles in the group, among which Representing the sender of the secret share, Represents the recipient of the secret share.

[0097] After that, the vehicle The secret share is encrypted using the public key and uploaded to the RSU.

[0098] In step S302, the RSU forwards the secret share to other vehicles in the group.

[0099] Step S4: The RSU asynchronously distributes the latest global model to the vehicles in the group; after receiving it, the vehicles use it as the initialization parameters of their local models, perform several rounds of training based on local data, and complete the calculation of local model parameters.

[0100] Step S5: The vehicle adds a random mask to its local model parameters to form encrypted local model parameters, and then sends the encrypted local model parameters to the RSU. The mask is a vector of random values ​​that applies the same-dimensional perturbation to the floating-point model parameters, changing their statistical distribution and ensuring that the local model parameters of a single vehicle cannot be restored before aggregation.

[0101] Step S6: RSU adopts an asynchronous aggregation strategy based on threshold triggering conditions; when the encrypted updates received by RSU reach the set upload threshold, RSU immediately starts global aggregation.

[0102] Specifically, RSU does not need to wait for all vehicles in the group to complete the upload of their masked local models. It can aggregate as soon as a preset threshold is reached. The optimal threshold is obtained using the following formula:

[0103]

[0104] in, It is the threshold number of vehicles that triggers aggregation. It is the proportionality coefficient (set to 0.7). This represents the current number of active vehicles in the group. This indicates rounding down to the nearest integer.

[0105] Step S7: If a vehicle fails to upload on time, resulting in the mask not being offset, the system, based on Shamir threshold secret sharing, requires online vehicles within the group to submit the missing share to restore the corresponding mask.

[0106] Specifically, step S7 includes steps S701-S702:

[0107] Step S701: If the group's reports are not collected by the end of the current training time window, the RSU generates a list of missing vehicles and broadcasts a mask reconstruction request to the online vehicles in the group.

[0108] In step S702, after receiving a request from the RSU, the online vehicles in the group locally sum the pre-stored group mask secret shares corresponding to the missing vehicles, encrypt them using a public key, attach a signature, and securely upload them to the RSU; the RSU receives no less than a threshold The secret share is used to reconstruct the missing mask; then the residual mask is canceled to ensure successful secure aggregation. The original mask is obtained using the following formula:

[0109]

[0110] in, This refers to the number of vehicle groups. For vehicle group The sum of the mask values ​​of the vehicle terminals that did not upload local model parameters. Indicates with vehicles The sum of related, but not received by the RSU secret shares. Indicates vehicle group Excluding vehicles Any vehicle terminal outside the group. Using this additive homomorphic property, the sum of the mask values ​​of multiple vehicles can be recovered at once, completing the reconstruction of lost masks within the group and ensuring the correctness of the aggregation result when mask loss occurs.

[0111] In one specific embodiment, a vehicle group is provided. The lack of vehicles during the internal aggregation process The mask value is calculated by summing the pre-stored group mask secret shares of the remaining online vehicles in the group after they receive the mask reconstruction request sent by the RSU. ,in, Indicates with vehicles Related, but not necessarily the sum of the secret shares received by the RSU. And using the public key pair Perform public key encryption and signing.

[0112] Step S8: RSU generates new global model parameters through weighted aggregation and broadcasts them to the vehicles; participating vehicles enter the next round of local training, while non-participating vehicles continue to train based on their existing versions, forming a closed-loop process of continuous iteration.

[0113] Specifically, step S8 includes steps S801-S802:

[0114] Step S801: Adjust the weights of vehicles in the aggregation process using a nonlinear penalty function.

[0115] Asynchronous uploading of vehicle local models is prone to time lag. The older global model generates local model parameters late, leading to version inconsistencies, aggregation bias, and slower convergence. A non-linear penalty function is used to adjust the weights of vehicles during the aggregation process. The optimal weights are obtained using the following formula:

[0116]

[0117] in, Indicates vehicle Aggregate weights, This represents the base weight (usually the mean of the vehicles participating in the aggregation). ), This indicates the time lag round for local model updates.

[0118] Step S802: Once the threshold condition is met, RSU begins executing global model aggregation. Global model aggregation is obtained using the following formula:

[0119]

[0120] in, Indicates vehicle The masked local model parameters have a mask sum of 0 in the aggregation.

[0121] Step S9: Repeat steps S1-S8 until the preset stopping condition is met. Obtain the final trained global model parameters.

[0122] Preferably, the stopping condition is that the number of communication cycles between the vehicle and the roadside unit (RSU) reaches a preset communication cycle threshold.

[0123] To verify the effectiveness of this method, it was tested in a simulation environment. The radius of the RSU coverage area was 500m, containing 20 vehicles traveling at different speeds, with the vehicle speeds ranging from [missing information]. The vehicle local model training iterations are set to 20, and the threshold for vehicle-RSU communication rounds is set to 200. For classification and recognition tasks in intelligent driving, a convolutional neural network model is used on a real-world vehicle-to-everything (V2X) image classification dataset. Vehicle data is partitioned using a non-independent, identically distributed (i.i.d.) partitioning method. Under this configuration, a comparative experiment is presented between this method and synchronous federated learning and asynchronous federated learning methods. Figure 3 The simulation results are for this scenario. From Figure 3 Two main conclusions can be drawn: (1) The method proposed in this invention achieves higher test accuracy than synchronous federated learning and asynchronous federated learning strategies; (2) The method of this invention achieves higher test accuracy with fewer communication rounds, demonstrating higher convergence efficiency.

[0124] In summary, this method collaboratively designs vehicle mobility and privacy protection, addressing the issues of low training efficiency and privacy leakage caused by semi-honest RSU. It establishes an integrated mechanism of "dwelling time grouping—conditionally triggered asynchronous aggregation—dynamic mask negotiation—secret sharing reconstruction": vehicles are grouped based on their dwell time in the RSU service area, constructing relatively stable multi-round collaboration within each group; aggregation is triggered when a preset number of vehicles in a group complete local training, reducing global waiting time and improving efficiency; a mask seed is generated using a pseudo-random function based on the shared key and training rounds within the group; vehicles generate random masks accordingly and superimpose them on local model parameters for local encryption; a dynamic rotating mask cancellation mechanism is used during the aggregation phase to complete secure aggregation and suppress reconstruction inference from semi-honest RSU; when only some vehicles participate in aggregation, a threshold-based secret sharing mechanism is used to reconstruct missing masks, ensuring aggregation correctness and robustness. This method achieves collaborative optimization of mobility features and privacy protection under controllable communication and computational overhead, adapting to the highly dynamic and real-time scenarios of vehicle-to-everything (V2X) networks.

[0125] This method overcomes the performance degradation problem caused by vehicles joining or leaving at any time due to path termination, link interruption, or energy depletion in federated learning data privacy protection methods in the Internet of Vehicles scenario. It proposes intelligent grouping based on expected dwell time, grouping vehicles with similar expected dwell time, thus constraining the dynamic of frequent switching to a local range, maintaining stable multi-round collaboration within the group, and effectively improving training efficiency and convergence accuracy. At the same time, it maintains training continuity and service real-time performance, enhancing the overall security and availability of the federated learning process in dynamic scenarios.

[0126] This invention considers the impact of vehicle mobility on aggregation efficiency and proposes a conditionally triggered asynchronous safe aggregation strategy. This strategy effectively addresses the characteristics of vehicle link fluctuations in the Internet of Vehicles (IoV). By combining threshold triggering with a weighting strategy that considers time lag rounds, continuous aggregation can be achieved without all users being online simultaneously: aggregation is initiated immediately when a preset number of uploads is reached, eliminating delays caused by waiting for slow vehicles; the time-lag local model is weighted according to version differences to suppress aggregation bias caused by outdated parameters, ensuring stable convergence, significantly improving learning efficiency and model accuracy, and enhancing the adaptability and usability of federated learning in IoV scenarios.

[0127] Simultaneously considering the dual impact of vehicle mobility on privacy and efficiency, a lost mask reconstruction based on secret sharing is constructed to meet the high dynamics and strong real-time requirements of the Internet of Vehicles. When only a subset of vehicles participate, threshold-based secret sharing can reconstruct the missing mask, ensuring aggregation correctness and training continuity. This achieves improved learning efficiency and enhanced privacy protection, making it suitable for online business scenarios such as intelligent driving.

[0128] The Roadside Unit (RSU) strategy is adopted, which is more efficient than the traditional cloud-vehicle architecture and is more in line with the infrastructure deployment in actual vehicle-to-everything (V2X) application scenarios.

[0129] 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 data privacy protection method for packet negotiation conditional asynchronous federated learning in vehicle-to-everything (V2X) networks, characterized in that, The method, applied to a vehicle-to-everything (V2X) system including roadside units and vehicle terminals, comprises the following steps: Step S1: The roadside unit predicts the expected dwell time of each vehicle terminal in the coverage area based on the status information of each vehicle terminal, divides the vehicles into at least one vehicle group based on the similarity of the expected dwell time, and publishes the public key of the vehicle in the group to the vehicle terminals in the group. Step S2: For each vehicle terminal in the vehicle group, a mask seed is randomly generated, encrypted using the public key of the vehicle in the group, and forwarded to other vehicle terminals in the same vehicle group through the roadside unit. The vehicle-to-everything (V2X) terminals in the vehicle group cooperate to calculate a set of random values ​​that sum to zero as the vehicle mask for each vehicle terminal. Step S3: For each vehicle terminal in the vehicle group, its own vehicle mask is split into multiple secret shares through threshold secret sharing and securely distributed to other vehicle terminals in the same vehicle group, and the secret shares of other vehicle terminals in the same vehicle group are pre-stored. In step S4, the roadside unit asynchronously sends the latest global model to the vehicle terminals within the same vehicle group. The vehicle terminals perform training based on the received global model and local data, and complete the calculation of local model parameters. Step S5: For each vehicle terminal, encrypt the local model parameters based on the vehicle mask and send them to the roadside unit. Step S6: After the encrypted model parameters received by the roadside unit reach the threshold, asynchronous global aggregation is performed and the vehicle mask is canceled to obtain a new global model. Step S7: In the asynchronous global aggregation, in response to the current training time window closing and the number of vehicle terminals that have collected encrypted model parameters not reaching the threshold, the vehicle mask of the missing vehicle terminal is reconstructed by requesting the secret share stored by the vehicle terminals in the vehicle group.

2. The data privacy protection method for packet negotiation conditional asynchronous federated learning for vehicle networking according to claim 1, characterized in that, Also includes: In step S8, the roadside unit broadcasts the new global model to the vehicle terminals. The vehicle terminals that participate in the training enter the next round of local training, while the vehicle terminals that do not participate in the training continue to train based on their existing version, forming a closed-loop process of continuous iteration. Step S9: Repeat steps S1-S8 until the preset stopping condition is met to obtain the final trained global model parameters.

3. The data privacy protection method for packet negotiation conditional asynchronous federated learning for vehicle networking according to claim 1, characterized in that, Step S1 includes: Step S101: The roadside unit calculates the expected dwell time of each vehicle based on the vehicle speed of each vehicle terminal and the path length of the navigation software. Step S102: Based on the expected dwell time of each vehicle terminal, the newly added vehicle terminals in this round of training are grouped by K-means clustering. Step S103: Construct a vehicle group instance based on the grouping results, and generate a group code, creation time, and expected lifespan, wherein the lifespan is calculated based on the expected dwell time and the number of vehicle terminals within the vehicle group; Step S104: Based on the proportion of active vehicles and the expected remaining lifespan, the roadside unit evaluates each existing vehicle group and determines whether to disband the vehicle group based on preset disbandment conditions. Step S105: Assign a public key for each vehicle group and publish it to the vehicle terminal.

4. The data privacy protection method for packet negotiation conditional asynchronous federated learning for vehicle networking according to claim 1, characterized in that, Step S2 includes: Step S201: The vehicle terminal generates a set of random numbers locally as a mask seed, encrypts them based on the vehicle public key in the set, and uploads them to the roadside unit. In step S202, the roadside unit forwards the encrypted mask seed to the vehicle terminals within the same vehicle group, and collaboratively calculates a set of random values ​​that sum to zero to obtain the vehicle mask for each vehicle terminal.

5. A data privacy protection method for packet negotiation conditional asynchronous federated learning for vehicle networks according to claim 1, characterized in that, Step S3 includes: Step S301: The vehicle terminal generates a Lagrange interpolation polynomial of the same order locally, calculates the corresponding secret share based on the vehicle mask, encrypts it with the vehicle public key in the group, and uploads it to the roadside unit. In step S302, the vehicle terminal securely distributes the secret share to other vehicle terminals in the same vehicle group through the roadside unit, so that each vehicle terminal pre-stores the secret share of other vehicle terminals in the same vehicle group.

6. A data privacy protection method for packet negotiation conditional asynchronous federated learning for vehicle networks according to claim 1, characterized in that, In step S6, a threshold is calculated based on the number of active vehicles in the current group.

7. A data privacy protection method for packet negotiation conditional asynchronous federated learning for vehicle networks according to claim 1, characterized in that, Step S7 includes: Step S701: In response to the expiration of the current training time window and the fact that the number of vehicle terminals that have collected encrypted model parameters has not reached the threshold, the roadside unit generates a list of vehicle terminals that are missing encrypted model parameters and broadcasts a vehicle mask reconstruction request within the vehicle group. Step S702: Within the vehicle group, the secret share corresponding to the pre-stored missing vehicle terminal is summed locally. The summation result is encrypted using the vehicle public key and uploaded to the roadside unit. After receiving a secret share greater than or equal to the threshold, the roadside unit reconstructs the missing vehicle mask and cancels it, thereby realizing asynchronous global aggregation.

8. A data privacy protection method for packet negotiation conditional asynchronous federated learning for vehicle networks according to claim 1, characterized in that, In step S6, the process of obtaining the new global model includes: Step S601: Based on the time-delayed rounds of local model updates in the vehicle terminal, calculate the aggregate weight of the vehicle terminal using a nonlinear penalty function. In step S602, the roadside unit aggregates the encrypted model parameters of the vehicle terminal based on the aggregated weights to obtain a new global model.

9. A data privacy protection method for packet negotiation conditional asynchronous federated learning for vehicle networks according to claim 1, characterized in that, In step S3, the secret share is calculated using the following formula: in, Vehicle terminal The Lagrange interpolation function, This indicates the vehicle's identification. Indicates vehicle Vehicle mask, For vehicle group The number of occupants in the vehicle, The first one randomly generated for the vehicle terminal Interpolation function coefficients, The secret share is sent to the vehicle terminal. And the secret share receiving vehicle terminal is The secret share.

10. A data privacy protection method for packet negotiation conditional asynchronous federated learning for vehicle networking according to claim 9, characterized in that, In step S7, the vehicle mask for reconstructing the missing vehicle terminal is calculated using the following formula: in, This refers to the number of vehicle groups. For vehicle group The sum of the mask values ​​of the vehicle terminals that did not upload local model parameters. Indicates connection with vehicle terminal The sum of relevant, but not received by the roadside unit secret shares. Indicates vehicle group Excluding vehicle terminals Any vehicle terminal outside of the vehicle.

Images