Model training method, apparatus, program, device, and medium

Abstract

A model training method, device, program, equipment and medium are disclosed, wherein the method is applied to the technical field of intelligent vehicles, the method comprises: locally training a local intelligent agent model based on local feature data of vehicle use data to obtain model update parameters; generating encrypted update parameters of the model update parameters based on a security encryption rule, and transmitting the encrypted update parameters to a server end; and receiving a global intelligent agent model issued by the server end. The embodiments of the present application can utilize the data of the vehicle owner under the premise of protecting the privacy of the vehicle owner, train the intelligent agent model based on the vehicle use data, and improve the utilization value of the vehicle use data.

Description

Technical Field

[0001] This invention relates to the field of intelligent vehicle technology, and in particular to a model training method, apparatus, program, device, and medium. Background Technology

[0002] With the rapid development of intelligent connected vehicles, vehicles are no longer just transportation tools, but also mobile data collectors. A single intelligent connected vehicle can generate terabytes of data daily, including high-definition video, high-precision location information, and highly sensitive personal privacy information. The collection and processing of this data has raised concerns among car owners about privacy and security. Furthermore, the development of intelligent agents within intelligent connected vehicles cannot be completely separated from the user data, such as vehicle routes and entertainment system usage data. This necessitates that intelligent connected vehicle manufacturers protect the privacy of car owners on the vehicle side. Currently, common mechanisms for protecting car owner privacy include data anonymization, localized data storage and access, and encrypted transmission. While these mechanisms improve the security of car owner data to some extent, they still have some shortcomings. For example, local storage and access can create data silos, preventing collaboration with data from other vehicles and significantly limiting the usability of car owner data. Moreover, although the transmission and storage process is encrypted, servers need to obtain decrypted vehicle usage data when training intelligent agent models using this data, posing a risk of privacy leaks and creating vulnerabilities in the privacy and security protection of vehicle users. Summary of the Invention

[0003] This invention provides a model training method, apparatus, program, device, and medium to enable the utilization of vehicle owner data while protecting the privacy of vehicle owners. By training intelligent agent models based on vehicle usage data, the utilization value of the data can be improved.

[0004] According to one aspect of the present invention, a model training method is provided for application in a vehicle, wherein the method includes:

[0005] Local intelligent agent models are trained locally based on local feature data from vehicle usage data to obtain model update parameters.

[0006] The encrypted update parameters of the model update parameters are generated based on the security encryption rules, and the encrypted update parameters are transmitted to the server.

[0007] The global intelligent agent model is received from the server.

[0008] According to another aspect of the present invention, a model training method is provided, applied on a server side, wherein the method includes:

[0009] Aggregate the encrypted update parameters uploaded by at least two vehicle terminals into model aggregation parameters;

[0010] Update the global agent model based on the model aggregation parameters;

[0011] The global intelligent agent model is sent to the vehicle.

[0012] According to another aspect of the present invention, a model training method is provided, applied to a model training system, the model training system including a server and a vehicle, the method comprising:

[0013] The local intelligent agent model on the vehicle is trained locally based on the local feature data of the vehicle usage data on the vehicle to obtain model update parameters.

[0014] The model update parameters are processed into encrypted update parameters on the vehicle side based on security encryption rules;

[0015] Aggregate at least two of the encrypted update parameters from the vehicle end into model aggregation parameters;

[0016] The global agent model on the server is updated according to the model aggregation parameters, and the global agent model is used as the local agent model on the vehicle.

[0017] According to another aspect of the present invention, a model training method is provided for use in a vehicle, wherein the apparatus includes:

[0018] The local training module is used to train the local agent model locally based on local feature data of vehicle usage data, and obtain model update parameters.

[0019] The upper-level training module is used to generate encrypted update parameters for the model update parameters based on security encryption rules, and transmit the encrypted update parameters to the server.

[0020] The model update module is used to receive the global agent model sent by the server.

[0021] According to another aspect of the present invention, a model training method is provided, applied on a server side, wherein the apparatus includes:

[0022] The parameter aggregation module is used to aggregate encrypted update parameters uploaded by at least two vehicle terminals into model aggregation parameters;

[0023] The model update module is used to update the global agent model based on the model aggregation parameters;

[0024] The model distribution module is used to distribute the global intelligent agent model to the vehicle.

[0025] According to another aspect of the present invention, a model training apparatus is provided, characterized in that it is applied to a model training system, the model training system comprising a server and a vehicle, the apparatus comprising:

[0026] The local training module is used to train the local intelligent agent model on the vehicle end based on the local feature data of the vehicle usage data, and obtain the model update parameters.

[0027] The data desensitization module is used to process the model update parameters into encrypted update parameters on the vehicle side based on security encryption rules;

[0028] The parameter aggregation module is used to aggregate at least two of the encrypted update parameters from the vehicle end into model aggregation parameters;

[0029] The local update module is used to update the global agent model on the server side according to the model aggregation parameters, and use the global agent model as the local agent model on the vehicle side.

[0030] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising:

[0031] At least one processor; and

[0032] A memory communicatively connected to the at least one processor; wherein,

[0033] The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the model training method according to any embodiment of the present invention.

[0034] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the model training method described in any embodiment of the present invention.

[0035] According to another aspect of the present invention, a computer program product is provided, the computer program product comprising a computer program, which, when executed by a processor, provides a model training method according to any embodiment of the present invention.

[0036] The technical solution of this invention involves acquiring vehicle usage data and extracting local feature data from it. This local feature data is then used to train a local intelligent agent model, generating model update parameters. These parameters are then encrypted using secure encryption rules and uploaded to a server to train a global intelligent agent model. The resulting global intelligent agent model is then used as the local intelligent agent model. This invention generates corresponding model update parameters locally from vehicle owner data through model training, eliminating the need to directly transmit the data to a server. This reduces the probability of privacy leaks in vehicle usage data. Encrypting the model update parameters using secure encryption rules enhances data transmission security. Furthermore, the model training process does not require transmitting complete vehicle usage data, reducing data transmission volume and lowering system requirements for network and storage.

[0037] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart of a model training method provided in Embodiment 1 of the present invention;

[0040] Figure 2 This is a flowchart of another model training method provided in Embodiment 2 of the present invention;

[0041] Figure 3 This is a flowchart of another model training method provided in Embodiment 3 of the present invention;

[0042] Figure 4 This is a flowchart of a model training method provided in Embodiment 4 of the present invention;

[0043] Figure 5 This is a flowchart of another model training method provided in Embodiment 5 of the present invention;

[0044] Figure 6 This is a flowchart of another model training method provided in Embodiment Six of the present invention;

[0045] Figure 7 This is a schematic diagram of a model training architecture provided in Embodiment 7 of the present invention;

[0046] Figure 8 This is an example diagram of a model training method provided in Embodiment 7 of the present invention;

[0047] Figure 9 This is a schematic diagram of the structure of a model training device according to Embodiment 8 of the present invention;

[0048] Figure 10 This is a schematic diagram of the structure of a model training device according to Embodiment Nine of the present invention;

[0049] Figure 11 This is a schematic diagram of the structure of a model training device according to Embodiment 10 of the present invention;

[0050] Figure 12 This is a schematic diagram of the structure of an electronic device that implements the model training method of this invention. Detailed Implementation

[0051] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0052] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0053] Example 1

[0054] Figure 1This is a flowchart of a model training method according to Embodiment 1 of the present invention. This embodiment is applicable to the training of intelligent agent models. The method can be executed by a model training device, which can be implemented in hardware and / or software, and can be configured in a vehicle. Figure 1 As shown, the method includes:

[0055] Step 110: Train the local agent model locally based on local feature data of vehicle usage data to obtain model update parameters.

[0056] Vehicle usage data can include information on vehicle use by the owner and / or users, including driving behavior data, in-vehicle environment data, vehicle operating status data, device connection data, and other sensitive data. Driving behavior data can include driving trajectory data and driving habit data. Driving trajectory data can include information such as driving routes, speed, and location, while driving habit data can include habitual data such as acceleration, braking, and steering. In-vehicle environment data can include in-vehicle audio and video data. In-vehicle audio data can include in-vehicle voice conversations and music playback records, while in-vehicle video data can include video recorded by in-vehicle cameras. Vehicle operating status data can include vehicle status and fault diagnosis information. Device connection data can include mobile phone connection information and third-party application data. Other sensitive data can include information on occupants and data about the vehicle's surrounding environment. Local feature data can be data obtained by extracting features from vehicle usage data. Local feature data can be feature data extracted from vehicle usage data in dimensions such as time, driving habits, road, and user role. Local feature data extraction methods can include, but are not limited to, extraction using neural network models and mathematical statistical methods.

[0057] The local intelligent agent model can be an intelligent agent model deployed on the vehicle side. The local intelligent agent model can include, but is not limited to, voice interaction models, autonomous driving models, intelligent cockpit models, vehicle control models, multimodal fusion models, reasoning and decision-making models, etc.

[0058] In this embodiment of the invention, vehicle usage data can be collected. This vehicle usage data may include driving behavior data, in-vehicle environment data, vehicle operating status data, device connection data, and other sensitive data that are securely stored in the vehicle. The collected vehicle usage data can be used to extract features through neural network models or data statistics. The extracted feature data can be used as local feature data. It is understood that the feature extraction process of vehicle usage data can be executed locally in the vehicle.

[0059] Specifically, local feature data can be input into a local agent model deployed on the vehicle. This local feature data can be used to train the local agent model for one or more rounds. The model state information value at the end of training can be used as the model update parameter. This model state information value can include the gradient value or model parameter value of the local agent model. The training process of the local agent model can be completed on the vehicle, and can be implemented through rule-based model training methods, supervised learning model training methods, unsupervised learning model training methods, reinforcement learning-based model training methods, context-based model training methods, etc.

[0060] Step 120: Generate encrypted update parameters for model update parameters based on security encryption rules, and transmit the encrypted update parameters to the server to train the global agent model.

[0061] The security encryption rules can be rules for securely encrypting model update parameters, and can include homomorphic encryption rules, symmetric encryption rules, asymmetric encryption rules, etc. The server can be a server that globally trains the agent models for each vehicle. The server can act as a model training server for each vehicle and can deploy a global agent model. This global agent model can be the aggregated result of updates to the local agent models of all vehicles. The global agent model represents the locally and globally optimal model state. The local agent models deployed on each vehicle can be trained and generated based on local vehicle usage data, and can reflect the local data characteristics of the vehicle.

[0062] In this embodiment of the invention, the model update parameters can be encrypted using security encryption rules. The encrypted ciphertext generated from the encrypted model update parameters can be used as the encrypted update parameters. The encrypted update parameters can be uploaded to the server. The global intelligent agent model can be trained by the server using the model update parameters from at least two vehicles.

[0063] Step 130: Receive the global intelligent agent model sent by the server.

[0064] In this embodiment of the invention, the server can distribute the trained global agent model to each vehicle terminal. The vehicle terminal can receive the global agent model and use it as the local agent model of the vehicle terminal, thereby completing the training of the local agent model. It can be understood that the server can distribute the global agent model to the vehicle terminal at regular intervals, or it can distribute the global agent model to the vehicle terminal as the local agent model of the vehicle terminal after completing the training of the global agent model in sequence.

[0065] This invention, in its embodiments, acquires vehicle usage data and extracts local feature data from this data. A local intelligent agent model is trained using this local feature data, and model update parameters for the local intelligent agent model are obtained. These model update parameters are then encrypted using secure encryption rules and uploaded to a server to train a global intelligent agent model. The resulting global intelligent agent model is then used as the local intelligent agent model. This invention generates corresponding model update parameters locally from vehicle owner data through model training, eliminating the need to directly transmit vehicle owner data to a server. This reduces the probability of user privacy leaks within the vehicle usage data. Encrypting the model update parameters using secure encryption rules enhances data transmission security. Furthermore, the model training process does not require the transmission of complete vehicle usage data, reducing the amount of data transmitted during training and lowering the system's network and storage requirements.

[0066] Example 2

[0067] Figure 2 This is a flowchart of another model training method provided in Embodiment 2 of the present invention. This embodiment is a concretization based on the above-described embodiments, describing the acquisition of local feature data and the training process of the local agent model. See [link to documentation]. Figure 2 The method provided in this embodiment of the invention specifically includes the following steps:

[0068] Step 210: Extract at least one of the following from the vehicle's driving route data and entertainment function usage data as vehicle usage data.

[0069] The data related to the driving route can include GPS location information, speed, timestamp, acceleration, direction, etc., while the data on entertainment functions can include audio playback data, video playback data, multimedia entertainment function data, game and interactive entertainment data, etc.

[0070] In this embodiment of the invention, driving video data and entertainment function usage data stored locally on the vehicle can be extracted as vehicle usage data. It is understood that the storage and processing of vehicle usage data can be performed locally on the vehicle.

[0071] Step 220: Perform data preprocessing on the vehicle usage data, wherein data preprocessing includes at least one of data cleaning, data denoising, and outlier detection.

[0072] Specifically, vehicle usage data collected can be preprocessed locally on the vehicle. This preprocessing can include data cleaning, denoising, and outlier detection. Data cleaning aims to identify and correct errors, duplications, and inconsistencies in the data, improving its quality and usability. Data cleaning involves preliminary checks to identify null values, outliers, and noisy data. Before cleaning, the original data can be backed up to prevent data loss. Based on the cleaning objectives, the data is initially screened to remove obviously unsuitable data. For missing values, deletion, imputation, and interpolation can be used to restore the data. Data denoising aims to remove noise from the data, improving its purity and quality. The denoising process can use regular expressions or other methods to identify invalid characters, duplicates, outliers, and format errors. The cleaned data can be validated to ensure its effectiveness. If the cleaning effect is unsatisfactory, the regular expressions need to be optimized and the data denoised again. Outlier detection identifies and handles anomalies in the data. Outlier detection can be achieved using methods such as Z-Score, IQR, and model-based detection.

[0073] Step 230: Extract the road features, time features, and driving behavior features of the vehicle usage data as local feature data.

[0074] Among them, road features can be feature data reflecting vehicle usage data in the road dimension, time features can be feature data reflecting vehicle usage data in the time dimension, and driving behavior features can be feature data reflecting vehicle usage data in the driver's driving behavior dimension. Road feature data, time feature data, and driving behavior feature data can be processed by the same or different neural network models.

[0075] In this embodiment of the invention, road features, time features, and driving behavior features can be extracted from vehicle usage data in terms of road dimension, time and temperature dimension, and driving behavior dimension, respectively. The extracted road features, time features, and driving behavior features can be used as local feature data. It is understood that the extraction process of ontological feature data can be implemented by a machine learning model or deep learning model trained locally on the vehicle. For example, a recurrent neural network model trained locally on the vehicle can be used to extract features from vehicle usage data to obtain local feature data.

[0076] Step 240: Input the local feature data of vehicle usage data into the local intelligent agent model for training.

[0077] In this embodiment of the invention, local feature data can be input into the deployed local intelligent agent model at the vehicle end, thereby training the local intelligent agent model.

[0078] Step 250: Determine that the local agent model has completed training, and obtain the gradient value or model parameter value of the local agent model as the model update parameter.

[0079] The gradient value can be the derivative of the loss function with respect to the model parameters during the training process of the local agent model. In each iteration, the gradient value indicates the direction and magnitude in which the model parameters need to be adjusted to minimize the loss function. The model update parameters can be the weights and biases learned by the model during training. These parameter values ​​determine how the model maps input data to output data. The model parameter values ​​at the end of training represent the final state of the model, which are obtained by adjusting based on the training data and the optimization algorithm.

[0080] In this embodiment of the invention, when the local agent model completes training, the gradient value or model parameter value of the local agent model can be extracted. The gradient value may include all gradient values ​​of the model during the entire training process or the gradient value when the model training is completed. The model parameter value may be the weights and biases learned during the training process. The obtained model parameter value and / or gradient value can be used as model update parameters.

[0081] Step 260: Generate encrypted update parameters for model update parameters based on security encryption rules, and transmit the encrypted update parameters to the server.

[0082] Step 270: Receive the global intelligent agent model sent by the server.

[0083] This invention, in its embodiments, extracts vehicle-side route-related data and entertainment function usage data as vehicle usage data. This vehicle usage data undergoes preprocessing to extract road features, time features, and driving behavior features as local feature data. A local intelligent agent model is trained based on this local feature data. The gradient values ​​or model parameter values ​​of the trained local intelligent agent model are used as model update parameters. These updated parameters are then encrypted using secure encryption rules and uploaded to a server to train a global intelligent agent model. The resulting global intelligent agent model is then used as the local intelligent agent model. This invention, by generating corresponding model update parameters locally from vehicle owner data through model training, reduces the probability of vehicle usage data leakage while decreasing the amount of data transmitted during model training. Encrypting the model update parameters using secure encryption rules further enhances data transmission security.

[0084] Example 3

[0085] Figure 3This is a flowchart of another model training method provided in Embodiment 3 of the present invention. This embodiment is a specific modification based on the above-described embodiments, describing the process of updating model parameters. See [link to documentation]. Figure 3 The method provided in this embodiment of the invention specifically includes the following steps:

[0086] Step 310: Determine the local characteristic data of vehicle usage data.

[0087] Step 320: Train the local agent model locally based on local feature data of vehicle usage data to obtain model update parameters.

[0088] Step 330: Perform differential privacy processing on the model update parameters according to the security encryption rules to obtain preliminary desensitized data with added noise.

[0089] Differential privacy processing can be a process for updating model parameters based on differential privacy technology. Differential privacy technology can protect privacy by adding noise to the data while allowing statistical analysis of the data. Differential privacy processing can add Laplace noise or Gaussian noise to the model update parameters.

[0090] In this embodiment of the invention, noise can be added to the model update parameters according to the differential privacy processing rules in the security encryption rules. The model update parameters with added noise can be used as preliminary decrypted data. It is understood that the degree of noise addition can be determined by the privacy budget configured in the differential privacy processing rules. The privacy budget can control the strength of privacy protection. The smaller the privacy budget value, the stronger the privacy protection, but the utility of the data may be reduced. An appropriate privacy budget value can be set in the differential privacy processing rules based on experience or latency.

[0091] Step 340: Perform homomorphic encryption on the initially decrypted data according to the security encryption rules to obtain the encryption update parameters.

[0092] Homomorphic encryption can be an encryption technique that allows specific operations to be performed directly on encrypted data. Homomorphic encryption can include additive homomorphic encryption, multiplicative homomorphic encryption, and fully homomorphic encryption.

[0093] In this embodiment of the invention, a homomorphic encryption algorithm in the security encryption rules can be obtained, and an encryption key and a decryption key generated according to the homomorphic encryption algorithm can be used to perform homomorphic encryption on the initially decrypted data according to the encryption key, thereby using the generated ciphertext as an encryption update parameter. It is understood that the encryption update parameter can perform the addition operation and / or multiplication operation corresponding to homomorphic encryption.

[0094] Step 350: Use the secure communication protocol to upload the encrypted update parameters to the server.

[0095] Among them, secure communication protocols can include a set of rules and standards used in network communication to protect the security of data transmission. Secure communication protocols can ensure the confidentiality, integrity and authenticity of data during transmission through technical means such as encryption, authentication and integrity verification. Secure communication protocols may include, but are not limited to, transport layer security protocols, IPSec security protocols, Kerberos network authentication protocols, etc.

[0096] In this embodiment of the invention, the vehicle can invoke a secure communication protocol to transmit encrypted update parameters, enabling the server to securely receive the encrypted update parameters. The server can then perform federated learning on the global agent model based on the received encrypted update parameters from one or more vehicle terminals. Federated learning technology can be a distributed machine learning technique that allows multiple vehicle terminals to collaboratively train a global agent model without sharing the original data. Federated learning technology can include horizontal federated learning, vertical federated learning, and federated transfer learning, among others.

[0097] Step 360: Receive the global intelligent agent model sent by the server.

[0098] This invention, in its embodiments, extracts local feature data from vehicle usage data, trains a local intelligent agent model locally based on this local feature data, and obtains model update parameters during the training process. These parameters are then processed using differential privacy to create preliminary de-identified data, which is further processed into encrypted update parameters using homomorphic encryption. The encrypted update parameters are then uploaded to a server via a secure communication protocol, resulting in the acquisition of the global intelligent agent model from the server. This invention integrates federated learning with multiple privacy protection technologies, combining federated learning with differential privacy and homomorphic encryption to construct a multi-layered privacy protection mechanism. This mechanism protects the privacy of vehicle owners during model training, thereby enhancing the security of vehicle usage data.

[0099] Furthermore, based on the above embodiments of the invention, differential privacy processing is performed on the model update parameters according to security encryption rules to obtain preliminary desensitized data with added noise, including:

[0100] Obtain the local feature data corresponding to the model update parameters and determine the gradient vector of the local feature data; apply L2 norm constraint to the gradient vector according to the gradient boundary value to obtain the clipped gradient vector; add Gaussian noise to the clipped gradient vector to obtain the preliminary desensitized data.

[0101] The gradient vector can be composed of gradient values ​​generated from local feature data used to train the local agent model for each set of data. Each element in the gradient vector can be a gradient value generated from one training iteration of the local agent model. The gradient bound can be a threshold for clipping the gradient vector, and the gradient bound can be set empirically. The L2 norm, also known as the Euclidean norm or 2-norm, is the square root of the sum of the squares of the vector elements. The L2 norm constraint restricts the gradient vector to the gradient bound, ensuring that the gradient vector of each sample has a uniform sensitivity before noise is added, thus providing a robust boundary for subsequent noise addition.

[0102] In this embodiment of the invention, the local feature data used by the local agent model when generating model update parameters can be determined. The gradient values ​​generated each time during the iterative training of the local agent model using the local feature data can be recorded, and a gradient vector can be constructed based on these gradient values. To prevent gradient explosion caused by excessively large gradients, the gradient vector can be subject to L2 norm constraints. The gradient boundary value corresponding to the L2 norm constraint can be obtained. The L2 norm of the gradient vector can be compared with the gradient boundary value. If the L2 norm of the gradient vector exceeds the gradient boundary value, the gradient vector is scaled down to within the gradient boundary value, thereby ensuring that the L2 norm of the gradient vector does not exceed the gradient boundary value. The clipped gradient vector can be used as the clipped gradient vector. Gaussian noise can be added to the clipped gradient vector to ensure that the noise can mask the gradient information of each local feature data. The noise addition formula can be as follows:

[0103]

[0104] in, This represents the gradient update value used in the global agent model training during the t-th iteration. N(0,σ) represents the gradient update value of the local agent model in the t-th iteration of vehicle-side i, where σ is a scale parameter related to the privacy budget. 2 C 2 I) indicates a value with zero mean and σ 2 C 2 The variance is a multidimensional Gaussian distribution, where C represents the gradient bound and I represents the identity matrix.

[0105] Example 4

[0106] Figure 4 This is a flowchart of a model training method according to Embodiment 4 of the present invention. This embodiment is applicable to the training of intelligent agent models. The method can be executed by a model training device, which can be implemented in hardware and / or software and can be configured on a server. Figure 4 As shown, the method includes:

[0107] Step 410: Aggregate the encrypted update parameters uploaded by at least two vehicle terminals into model aggregation parameters.

[0108] The model aggregation parameters can be generated by aggregating encrypted update parameters uploaded from different vehicles or the same vehicle at different times. This aggregation operation can include methods such as simple averaging, weighted averaging, federated averaging, secure aggregation, and model decoupling.

[0109] In this embodiment of the invention, at least two encrypted update parameters uploaded by vehicle terminals can be obtained. These encrypted update parameters can be aggregated, and the resulting aggregation result can be used as the model aggregation parameter for each encrypted update parameter. For example, different weight values ​​can be configured for different vehicle terminals. The encrypted update parameters for each vehicle terminal can be obtained, and each encrypted update parameter can be weighted according to its weight value. The average value of the weighted encrypted update parameters can be used as the model aggregation parameter.

[0110] Step 420: Update the global agent model based on the model aggregation parameters.

[0111] In this embodiment of the invention, the global agent model can be updated using model aggregation parameters. The update process may include increasing or decreasing the weights of the global agent model using gradient values ​​in the model aggregation parameters, or configuring the model parameter values ​​in the model aggregation parameters as model parameters of the global agent model, etc.

[0112] Step 430: Deploy the global intelligent agent model to the vehicle.

[0113] In this embodiment of the invention, the server can distribute the trained global agent model to the vehicle. The vehicle can receive the global agent model and use it as its local agent model, thereby completing the training of the local agent model. It is understood that the server can distribute the global agent model to the vehicle at regular intervals, or it can distribute the global agent model to the vehicle as its local agent model after completing the training of the global agent model in sequence. The vehicle that distributes the global agent model can be all or part of the vehicle that has uploaded the encrypted update parameters, or it can be other vehicle that has not uploaded the encrypted update parameters.

[0114] In this embodiment of the invention, encrypted update parameters from different vehicle terminals are obtained and aggregated into model aggregation parameters. The global agent model is then updated based on these aggregation parameters, and the updated global agent model is distributed to the vehicle terminals. By transmitting encrypted update parameters between the server and the vehicle terminals, this embodiment of the invention can reduce the amount of data transmitted during model training while ensuring the privacy of vehicle usage data. This reduces the system's network and storage requirements. Training the global agent model based on the encrypted update parameters from different vehicle terminals can improve the intelligence level of the agent model, enhance model performance, and improve the user experience of the agent model.

[0115] Example 5

[0116] Figure 5 This is a flowchart of another model training method provided in Embodiment 5 of the present invention. This embodiment is a concretization based on the above-described embodiments, describing the aggregation process of encrypted update parameters. See [link to documentation]. Figure 5 The method provided in this embodiment of the invention specifically includes the following steps:

[0117] Step 510: Obtain the encrypted update parameters uploaded by at least two vehicle terminals, and determine the aggregation calculation coefficients for each vehicle terminal.

[0118] Among them, the aggregation calculation coefficient can be the weight value used by each vehicle terminal when aggregating the encrypted update parameters. The aggregation calculation coefficient can reflect the importance of each vehicle terminal's encrypted update parameters to the training of the global intelligent agent model. The aggregation calculation coefficient can be determined based on the actual state of the vehicle terminal. For example, the aggregation calculation coefficient can be determined by whether the vehicle terminal is in an online state, or by the evaluation value of the vehicle terminal's historical encrypted update parameters, etc.

[0119] In this embodiment of the invention, an aggregate calculation coefficient corresponding to each vehicle terminal can be obtained. The aggregate calculation coefficient of each vehicle terminal can be pre-configured or determined in real time according to the current state of the vehicle terminal. For example, a multimodal fusion model can be invoked for each vehicle terminal to generate an aggregate calculation coefficient through the public data or de-identified data of each vehicle terminal, and the aggregate calculation coefficient can be stored on the server.

[0120] Step 520: For each vehicle, the encrypted update parameters are weighted based on the aggregation calculation coefficients to obtain the weighted update parameters.

[0121] In this embodiment of the invention, the aggregate calculation coefficient of each vehicle can be used to weight its own encryption update parameters. For example, the product or sum of the aggregate calculation coefficient and the encryption update parameters can be determined, and the encryption update parameters of each vehicle after being weighted by the aggregate calculation coefficient can be used as the weighted update parameters.

[0122] Step 530: Perform secure multi-party computation on each weighted update parameter to obtain the model aggregation parameters.

[0123] In this embodiment of the invention, secure multi-party computation allows multiple vehicle terminals to collaboratively complete computation tasks without disclosing their respective privacy data. Each vehicle terminal only receives the computation result and cannot obtain the weighted update parameters input by other vehicle terminals. Secure multi-party computation can include semi-honest computation models, malicious computation models, and covert attack computation models. For example, the computation model configured according to secure multi-party computation can be a computation circuit diagram. Each weighted update parameter can be mapped to different computation gates within the computation circuit diagram. Aggregate computation can be performed on each weighted update parameter according to the computation circuit diagram to obtain model aggregate parameters.

[0124] Specifically, a computational model based on secure multi-party computation can be obtained. Each weighted update parameter can be assigned a corresponding position within the computational model according to its respective vehicle-side or vehicle-side identifier. The weighted update parameters at each position can be calculated according to the computational model. The output of the computational model based on secure multi-party computation can be used as the model aggregation parameter.

[0125] Step 540: Update the global agent model based on the model aggregation parameters.

[0126] Step 550: Extract the model weight values ​​of the global intelligent agent model; call the secure communication protocol to send the model weight values ​​to the vehicle.

[0127] In this embodiment of the invention, the model weight values ​​of the trained global agent model can be extracted, and these model weight values ​​can be sent to one or more vehicle terminals through a secure communication protocol. Each vehicle terminal can update its local agent model according to the received model weight values. That is, the original model weight values ​​of the local agent model can be replaced with the model weight values ​​sent by the server, thereby realizing the update of the local agent model and further reducing the amount of data transmission for updating the local agent model.

[0128] This invention, in its embodiments, obtains encrypted update parameters from multiple vehicle-side endpoints, determines corresponding aggregation calculation coefficients for each endpoint, weights the encrypted update parameters according to these coefficients to obtain weighted update parameters, and uses the result of secure multi-party computation on these weighted update parameters as model aggregation parameters. These aggregation parameters are then used to update the global agent model. The updated global agent model's model weights are extracted and distributed to the vehicle-side endpoints to update the local agent model, thus achieving local agent model updates. This invention, through the aggregation calculation coefficients on the vehicle-side endpoints, reflects the importance of the global agent model training process. The global agent model can be flexibly adjusted based on these coefficients, ensuring its training closely aligns with real-world scenarios, improving its reliability, and enhancing resource utilization efficiency.

[0129] Furthermore, based on the above embodiments of the invention, the aggregation calculation coefficients for each vehicle end are determined, including:

[0130] The online status, network conditions, and computing resources of each vehicle are extracted as vehicle performance indicators; the aggregation calculation coefficients of each vehicle are determined according to the vehicle performance indicators.

[0131] Among them, online status can represent the connection status between each vehicle and the server. Online status can be represented by being online or offline. Network status can be information about the network connection quality of the vehicle. Network status can include the maximum length of network messages, bandwidth, latency, signal strength, etc. Computing resources can be the computing power of the vehicle to support various in-vehicle applications and services. Computing resources can be various software and hardware resources in the vehicle used for processing data and computing tasks.

[0132] In this embodiment of the invention, information such as online status, network status, and computing resources of each vehicle terminal can be obtained. By normalizing the above-mentioned online status, network status, and computing resources, the collected data can be normalized to the same range. The normalized values ​​of online status, network status, and computing resources of each vehicle terminal can be used as vehicle performance indicators. The proportion of each vehicle performance indicator to the overall performance can be converted into the aggregation calculation coefficient of each vehicle terminal. For example, the total indicator value of each vehicle performance indicator can be determined, and the proportion of each vehicle performance indicator to the total indicator value can be used as the aggregation calculation coefficient of the corresponding vehicle terminal.

[0133] Based on the above embodiments of the invention, the aggregation calculation coefficients for each vehicle end are determined according to vehicle performance indicators, including at least one of the following:

[0134] The aggregate calculation coefficients corresponding to each vehicle performance index are determined by using a neural network model or a machine learning model.

[0135] Determine the value of each vehicle performance index, and use the proportion of each index value to the sum of all index values ​​as the aggregation calculation coefficient.

[0136] In this embodiment of the invention, the aggregation calculation coefficient can be determined using a neural network model or a machine learning model. Vehicle performance data from each vehicle can be intelligently input into the neural network model or machine learning model, and the output of the neural network model or machine learning model can be used as the aggregation calculation coefficient. Alternatively, the index values ​​of each vehicle performance indicator can be extracted, and the sum of these index values ​​can be determined. The proportion of each vehicle's index value to the sum of all index values ​​can be used as the aggregation calculation coefficient.

[0137] Example 6

[0138] Figure 6 This is a flowchart of another model training method provided in Embodiment Six of the present invention. This embodiment is applicable to the training of intelligent agent models. The method includes execution by a model training device, which can be implemented in hardware and / or software. The device can be configured in a model training system, which includes a server and a vehicle. See [link to documentation]. Figure 6 As shown, the method includes:

[0139] In this embodiment of the application, feature extraction can be performed on the vehicle usage data of each vehicle terminal. The extracted feature data can be used as the local feature data of the corresponding vehicle terminal. It is understood that the extraction process of the local feature data can be performed by the vehicle terminal or other terminal devices. For example, the vehicle usage data of the vehicle terminal can be saved to a third-party trusted platform, and the third-party trusted platform can extract features from the vehicle usage data of each vehicle terminal to obtain the local feature data of each vehicle terminal.

[0140] In some embodiments of the invention, determining local feature data based on vehicle usage data includes: extracting at least one of the vehicle's driving route related data and entertainment function usage data as vehicle usage data; performing data preprocessing on the vehicle usage data, wherein the data preprocessing includes at least one of data cleaning, data denoising, and outlier detection; and extracting the road features, time features, and driving behavior features of the vehicle usage data as local feature data.

[0141] Step 610: Train the local intelligent agent model on the vehicle side locally based on the local feature data of the vehicle usage data on the vehicle side to obtain the model update parameters.

[0142] In this embodiment of the application, feature extraction can be performed on the vehicle usage data of each vehicle terminal. The extracted feature data can be used as the local feature data of the corresponding vehicle terminal. It is understood that the extraction process of the local feature data can be performed by the vehicle terminal or other terminal devices. For example, the vehicle usage data of the vehicle terminal can be saved to a third-party trusted platform, and the third-party trusted platform can extract features from the vehicle usage data of each vehicle terminal to obtain the local feature data of each vehicle terminal.

[0143] In some embodiments of the invention, determining local feature data based on vehicle usage data includes: extracting at least one of the vehicle's driving route related data and entertainment function usage data as vehicle usage data; performing data preprocessing on the vehicle usage data, wherein the data preprocessing includes at least one of data cleaning, data denoising, and outlier detection; and extracting the road features, time features, and driving behavior features of the vehicle usage data as local feature data.

[0144] In this embodiment of the invention, for each vehicle terminal, local feature data can be input into the local intelligent agent model deployed on the corresponding vehicle terminal. The local intelligent agent model is trained using the local feature data, and model update parameters are extracted from the trained local intelligent agent model. The training of the local intelligent agent model using local feature data can be performed in one or more rounds, and the model update parameters can be model parameter values ​​of the local intelligent agent model or gradient values ​​from the model training process, etc.

[0145] In some embodiments of the invention, the local intelligent agent model on the vehicle is trained locally based on local feature data to obtain model update parameters, including:

[0146] Local feature data is input into the local agent model on the vehicle for training; once the local agent model has completed training, the gradient value or model parameter value of the local agent model is obtained as the model update parameter.

[0147] Step 620: Based on security encryption rules, process the model update parameters into encrypted update parameters for the vehicle.

[0148] In this embodiment of the invention, the model update parameters are encrypted using a secure encryption rule. The ciphertext generated from the encrypted model update parameters can be used as the encrypted update parameters. This secure encryption rule can be used to prevent the leakage of vehicle usage data, further enhancing the security of vehicle owner privacy. Specifically, the secure encryption rule can include homomorphic encryption rules, symmetric encryption rules, asymmetric encryption rules, etc.

[0149] In some embodiments of the invention, the model update parameters are processed into encrypted update parameters on the vehicle side based on secure encryption rules, including: performing differential privacy processing on the model update parameters on the vehicle side according to secure encryption rules to obtain preliminary desensitized data with added noise; performing homomorphic encryption on the preliminary desensitized data according to secure encryption rules to obtain encrypted update parameters; and calling a secure communication protocol to upload the encrypted update parameters to the server side, so that the server side can perform federated learning on the global intelligent agent model based on the encrypted update parameters.

[0150] Step 630: Aggregate the encrypted update parameters of at least two vehicle ends into model aggregation parameters.

[0151] In this embodiment of the invention, at least two vehicle-side encrypted update parameters can be obtained. These multiple encrypted update parameters can be aggregated to obtain model aggregate parameters. The model aggregate parameters can be generated by performing aggregation operations on multiple encrypted update parameters. These aggregation operations can include methods such as simple averaging, weighted averaging, federated averaging, secure aggregation, and model decoupling. For example, the weight value of each vehicle-side parameter can be obtained, and the encrypted update parameters of each vehicle-side parameter can be weighted and averaged according to the corresponding vehicle-side weight value. The resulting weighted average value is then used as the aggregated model parameters.

[0152] In some embodiments of the invention, the encrypted update parameters from at least two vehicle terminals are aggregated as model aggregation parameters, including: obtaining encrypted update parameters uploaded by at least two vehicle terminals and determining the aggregation calculation coefficients for each vehicle terminal; for each vehicle terminal, weighting the encrypted update parameters based on the aggregation calculation coefficients to obtain weighted update parameters; and performing secure multi-party computation on each weighted update parameter to obtain model aggregation parameters.

[0153] Based on the above embodiments of the invention, the aggregation calculation coefficient of each vehicle terminal is determined, including: extracting the online status, network status and computing resources of each vehicle terminal as vehicle performance indicators; and determining the aggregation calculation coefficient of each vehicle terminal according to the vehicle performance indicators.

[0154] In some embodiments of the invention, the aggregation calculation coefficients for each vehicle end are determined according to vehicle performance indicators, including at least one of the following:

[0155] The aggregate calculation coefficients corresponding to each vehicle performance index are determined by using a neural network model or a machine learning model.

[0156] Determine the value of each vehicle performance index, and use the proportion of each index value to the sum of all index values ​​as the aggregation calculation coefficient.

[0157] Step 640: Update the global agent model on the server side according to the model aggregation parameters, and use the global agent model as the local agent model on the vehicle side.

[0158] In this embodiment of the invention, the global agent model deployed on the server can be updated using model aggregation parameters. For example, the update process may include retraining the global agent model using model aggregation parameters, or directly adjusting the model parameter values ​​of the global agent model using model aggregation parameters, thereby updating the global agent model. The updated global agent model can then be used as the local agent model for each vehicle.

[0159] In this embodiment of the invention, vehicle usage data from the vehicle terminal is acquired and processed into local feature data for the corresponding vehicle terminal. A local intelligent agent model deployed within the corresponding vehicle terminal is trained based on the local feature data of each vehicle terminal. Model update parameters of the local intelligent agent model are obtained. Based on security encryption rules, the model update parameters of each vehicle terminal are processed into encrypted update parameters. The encrypted update parameters of at least two vehicle terminals are aggregated into model aggregation parameters. A global intelligent agent model deployed on the server side is trained according to the model aggregation parameters. The trained global intelligent agent model is used as the local intelligent agent model for each vehicle terminal. This embodiment of the invention utilizes vehicle owner data... The local agent model on the vehicle serves as the model update parameter, eliminating the need to directly transmit vehicle owner data to the server. This reduces the probability of vehicle usage data leakage. Encrypting the model update parameters based on security encryption rules enhances data transmission security. The model training process does not require the transmission of complete vehicle usage data, reducing the amount of data transmitted during training and lowering the system's network and storage requirements. Training the global agent model based on model update parameters from multiple vehicles allows the global agent model to reflect the local characteristics of different vehicles, improving its versatility and enhancing model performance.

[0160] Example 7

[0161] This invention aims to protect the privacy of vehicle usage data, achieving efficient utilization of vehicle data without uploading original driving route data and preventing the leakage of sensitive personal information. While protecting the privacy of vehicle usage data, it enables cross-vehicle data collaboration, improving the model's generalization ability and accuracy. Reducing data transmission volume during model training lowers communication and storage costs, reduces network bandwidth and storage resource requirements, and improves system efficiency. See also Figure 7 The model training method provided in this embodiment of the invention can be jointly executed by the vehicle end and the federated server end.

[0162] The vehicle-side architecture can include a sensor layer, a data processing layer, a model training layer, a security processing layer, a communication layer, and a model update layer. The sensor layer includes a GPS module, speedometer, accelerometer, etc., used to collect vehicle driving data. The data processing layer includes a data acquisition and preprocessing model, and a feature extraction module. The data acquisition and preprocessing module is responsible for real-time data acquisition and preprocessing; the feature extraction module extracts features from the preprocessed data to generate feature vectors for model training. The model training layer includes a local model training module, which trains a local model based on the feature vectors. The security processing layer includes a differential privacy processing module and a homomorphic encryption module. The differential privacy processing module adds noise to the model parameters, and the homomorphic encryption module encrypts the processed model parameters. The secure communication module in the communication layer uploads the encrypted model parameters to the federated server via a secure communication protocol and receives the global model. The local model update module in the model update layer receives the global model and updates the local model parameters based on the global model.

[0163] The federated server can include a communication layer, a model aggregation layer, and a management layer. The secure communication module in the communication layer receives encrypted model parameters uploaded by the vehicle-side and distributes the global model to the vehicle-side. The model aggregation layer includes a secure multi-party computation module and a model update module. The secure multi-party computation module performs secure aggregation of the encrypted model parameters, and the model update module generates the global model. The management layer includes a node management module and a logging module. The node management module manages the vehicle-side nodes participating in the federated computation, while the logging module records the model aggregation and communication processes to ensure system traceability and compliance. See also... Figure 8 The model training method provided in this embodiment of the invention may include the following steps:

[0164] 1. Data Acquisition and Preprocessing: Real-time local data collection on the vehicle's driving route, including GPS location information, speed, timestamp, acceleration, and direction. The collected data undergoes preprocessing such as cleaning, noise reduction, and outlier detection to ensure data quality.

[0165] 2. Local Feature Extraction and Modeling: Extract road features, time features, driving behavior features, etc. from the collected data, and generate feature vectors for model training based on the road features, time features, and driving behavior features. Based on the extracted features, train machine learning models or deep learning models, such as recurrent neural networks, locally on the vehicle.

[0166] 3. Model Parameter Encryption and Upload: Differential Privacy Processing: Differential privacy technology is applied to the model parameters obtained from local model training, adding noise to prevent reverse engineering of the original data through the parameters. Homomorphic Encryption: The model parameters after differential privacy processing are homomorphically encrypted to ensure that model aggregation calculations can still be performed in the encrypted state. Secure Communication: Secure communication protocols such as Transport Layer Security (TLS / SSL) are used to upload the encrypted model parameters to the federated server.

[0167] 4. Model Aggregation: Receives encrypted model parameters from multiple vehicle terminals, uses secure multi-party computation technology to aggregate the encrypted model parameters, generates a global model, and ensures that the federated server cannot decrypt the model parameters of a single vehicle terminal.

[0168] 5. Global Model Distribution Module: Distributes the aggregated global model to each vehicle terminal via a secure communication protocol.

[0169] 6. Model Update: After receiving the global model, each vehicle updates its local model parameters to improve the model's generalization ability and accuracy.

[0170] Furthermore, the model training method can be iteratively updated and dynamically managed. The iterative training mechanism allows for continuous iteration of the training process, with model performance constantly improving as more data is added. The dynamic node management mechanism dynamically adjusts the vehicle-end nodes participating in federated learning based on the vehicle's online status, network conditions, and computing resources, optimizing system performance and resource utilization.

[0171] This invention, by storing the original driving route data locally on the vehicle and not uploading it to the cloud, complies with data protection regulations. Through technologies such as differential privacy, homomorphic encryption, and secure multi-party computation, the invention protects the vehicle owner's privacy throughout the entire process from data collection to model aggregation.

[0172] Example 8

[0173] Figure 9 This is a schematic diagram of the structure of a model training device according to Embodiment 8 of the present invention, as shown below. Figure 9 As shown, the device includes:

[0174] The local training module 710 is used to train the local agent model locally based on local feature data of vehicle usage data, and obtain model update parameters.

[0175] The upper-level training module 720 is used to generate encrypted update parameters of the model update parameters based on security encryption rules, and transmit the encrypted update parameters to the server.

[0176] The model update module 730 is used to receive the global intelligent agent model sent by the server.

[0177] In this embodiment of the invention, a feature extraction module acquires local feature data of vehicle usage data, a local training module trains a local intelligent agent model using the local feature data, and obtains the model update parameters of the local intelligent agent model. A higher-level training module encrypts the model update parameters into encrypted update parameters using secure encryption rules and uploads these encrypted update parameters to the server to train the global intelligent agent model. The model update module obtains the global intelligent agent model issued by the server and uses this global intelligent agent model as the local intelligent agent model. This embodiment of the invention generates corresponding model update parameters locally from vehicle owner data through model training, eliminating the need to directly transmit vehicle owner data to the server. This reduces the probability of user privacy leakage within vehicle usage data. Encrypting the model update parameters based on secure encryption rules improves data transmission security. The model training process does not require the transmission of complete vehicle usage data, reducing the amount of data transmitted during model training and lowering the system's network and storage requirements.

[0178] Based on the above embodiments of the invention, a feature extraction module is also included, specifically used for:

[0179] The data extraction unit is used to extract at least one of the following as vehicle usage data: driving route related data and entertainment function usage data.

[0180] The preprocessing unit is used to preprocess vehicle usage data, wherein the data preprocessing includes at least one of data cleaning, data denoising, and outlier detection.

[0181] The feature extraction unit is used to extract road features, time features, and driving behavior features from vehicle usage data as local feature data.

[0182] Based on the above embodiments of the invention, the local training module 710 includes:

[0183] The model training unit is used to input local feature data into the local agent model for training.

[0184] The parameter determination unit is used to determine whether the local agent model has completed training and to obtain the gradient value or model parameter value of the local agent model as model update parameters.

[0185] Based on the above embodiments of the invention, the upper-level training module 720 includes:

[0186] Differential privacy units are used to perform differential privacy processing on model update parameters according to security encryption rules, resulting in preliminary desensitized data with added noise.

[0187] The homomorphic encryption unit is used to homomorphically encrypt the initially decrypted data according to the security encryption rules to obtain the encryption update parameters.

[0188] The secure transmission unit is used to call the secure communication protocol to upload encrypted update parameters to the server, so that the server can perform federated learning on the global agent model based on the encrypted update parameters.

[0189] Based on the above embodiments of the invention, the differential privacy unit is specifically used to obtain local feature data corresponding to the model update parameters and determine the gradient vector of the local feature data; to restrict the gradient vector according to the gradient boundary value using the L2 norm to obtain the clipped gradient vector; and to add Gaussian noise to the clipped gradient vector to obtain preliminary desensitized data.

[0190] The model training apparatus provided in this embodiment of the invention can execute the model training method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0191] Example 9

[0192] Figure 10 This is a schematic diagram of a model training device according to Embodiment Nine of the present invention. See also... Figure 10 The apparatus provided in this embodiment of the invention specifically includes:

[0193] The parameter aggregation module 810 is used to aggregate encrypted update parameters uploaded by at least two vehicle terminals into model aggregation parameters.

[0194] The model update module 820 is used to update the global agent model based on the model aggregation parameters.

[0195] The model distribution module 830 is used to distribute the global intelligent agent model to the vehicle.

[0196] In this embodiment of the invention, encrypted update parameters from different vehicle terminals are obtained through a parameter aggregation module, and these encrypted update parameters are aggregated into model aggregation parameters. The model update module updates the global agent model based on the model aggregation parameters, and the model distribution module distributes the updated global agent model to the vehicle terminals. By transmitting encrypted update parameters between the server and the vehicle terminals, this embodiment of the invention can reduce the amount of data transmitted during model training while ensuring the availability of vehicle usage data. This reduces the system's network and storage requirements. Training the global agent model based on the encrypted update parameters from different vehicle terminals can improve the intelligence level of the agent model, enhance model performance, and improve the user experience of the agent model.

[0197] Furthermore, based on the above embodiments of the invention, the parameter aggregation module 810 includes:

[0198] The coefficient determination unit is used to obtain encrypted update parameters uploaded by at least two vehicle terminals and determine the aggregate calculation coefficients for each vehicle terminal.

[0199] The parameter weighting unit is used to weight the encrypted update parameters based on the aggregated calculation coefficients for each vehicle end, so as to obtain the weighted update parameters.

[0200] The aggregation execution unit is used to perform secure multi-party computation on each weighted update parameter to obtain the model aggregation parameters.

[0201] Based on the above embodiments of the invention, the coefficient determination unit is specifically used to extract the online status, network status and computing resources of each vehicle terminal as vehicle performance indicators; and to determine the aggregate calculation coefficients of each vehicle terminal according to the vehicle performance indicators.

[0202] Based on the above embodiments of the invention, the coefficient determination unit is used for at least one of the following: determining the aggregate calculation coefficients corresponding to each vehicle performance index through a neural network model or a machine learning model; determining the index value of each vehicle performance index, and using the proportion of each index value to the sum of all index values ​​as the aggregate calculation coefficient.

[0203] Based on the above embodiments of the invention, the model distribution module 830 is specifically used to extract the model weight values ​​of the global intelligent agent model and call the secure communication protocol to distribute the model weight values ​​to the vehicle.

[0204] The model training apparatus provided in this embodiment of the invention can execute the model training method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0205] Example 10

[0206] Figure 11 This is a schematic diagram of a model training device according to Embodiment 10 of the present invention. See also... Figure 11 The apparatus provided in this embodiment of the invention includes:

[0207] The local training module 910 is used to train the local intelligent agent model on the vehicle based on the local feature data of vehicle usage data, and obtain the model update parameters.

[0208] The data desensitization module 920 is used to process model update parameters into encrypted update parameters for the vehicle based on security encryption rules.

[0209] The parameter aggregation module 930 is used to aggregate the encrypted update parameters of at least two vehicle ends into model aggregation parameters.

[0210] The local update module 940 is used to update the global agent model on the server side according to the model aggregation parameters, and use the global agent model as the local agent model on the vehicle side.

[0211] In this embodiment of the invention, a feature extraction module acquires vehicle usage data from the vehicle terminal and processes it into local feature data for the corresponding vehicle terminal. A local training module trains a local intelligent agent model deployed within the corresponding vehicle terminal based on the local feature data of each vehicle terminal, and obtains the model update parameters of the local intelligent agent model. A data anonymization module processes the model update parameters of each vehicle terminal into encrypted update parameters based on security encryption rules. A parameter aggregation module aggregates the encrypted update parameters of at least two vehicle terminals into model aggregation parameters. The global intelligent agent model deployed on the server is trained according to the model aggregation parameters. The local update module uses the trained global intelligent agent model as the local intelligent agent model for each vehicle terminal. This invention, through its embodiment, uses the vehicle owner data in the local intelligent agent model on the vehicle as the model update parameters, eliminating the need to directly transmit the vehicle owner data to the server. This reduces the probability of vehicle usage data leakage. Encrypting the model update parameters based on security encryption rules enhances data transmission security. The model training process does not require the transmission of complete vehicle usage data, reducing the amount of data transmitted during training and lowering the system's network and storage requirements. Training the global intelligent agent model based on model update parameters from multiple vehicle terminals allows the global intelligent agent model to reflect the local characteristics of different vehicle terminals, improving its versatility and enhancing model performance.

[0212] Example 11

[0213] Figure 12 This is a schematic diagram of the structure of an electronic device implementing the model training method of embodiments of the present invention. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0214] like Figure 12As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 may also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0215] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0216] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as model training methods.

[0217] In some embodiments, the model training method may be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program may be loaded and / or mounted on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the model training method described above may be performed. Alternatively, in other embodiments, processor 11 may be configured to execute the model training method by any other suitable means (e.g., by means of firmware).

[0218] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0219] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0220] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0221] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0222] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0223] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0224] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0225] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A model training method, characterized in that, Applied to the vehicle end, the method includes: Local intelligent agent models are trained locally based on local feature data from vehicle usage data to obtain model update parameters. The encrypted update parameters of the model update parameters are generated based on the security encryption rules, and the encrypted update parameters are transmitted to the server. Receive the global intelligent agent model sent by the server.

2. The method according to claim 1, characterized in that, Also includes: The local feature data for determining the vehicle usage data includes: Extract at least one of the vehicle's driving route data and entertainment function usage data as the vehicle usage data; The vehicle usage data is preprocessed, wherein the preprocessing includes at least one of data cleaning, data denoising, and outlier detection. The road features, time features, and driving behavior features of the vehicle usage data are extracted as the local feature data.

3. The method according to claim 1 or 2, characterized in that, The local intelligent agent model is trained locally using local feature data based on vehicle usage data to obtain model update parameters, including: The local feature data is input into the local intelligent agent model for training; Once the local agent model has completed training, the gradient value or model parameter value of the local agent model is obtained as the model update parameter.

4. The method according to any one of claims 1-3, characterized in that, The step of generating encrypted update parameters for the model update parameters based on security encryption rules and transmitting the encrypted update parameters to the server includes: Differential privacy processing is performed on the model update parameters according to the security encryption rules to obtain preliminary desensitized data with added noise; The initially decrypted data is homomorphically encrypted according to the security encryption rules to obtain the encryption update parameters; The encrypted update parameters are uploaded to the server by invoking a secure communication protocol.

5. The method according to claim 4, characterized in that, The step of performing differential privacy processing on the model update parameters according to the security encryption rules to obtain preliminary desensitized data with added noise includes: Obtain the local feature data corresponding to the model update parameters, and determine the gradient vector of the local feature data; The gradient vector is subjected to L2 norm constraint according to the gradient boundary value to obtain the clipped gradient vector; Gaussian noise is added to the clipped gradient vector to obtain the preliminary desensitized data.

6. A model training method, characterized in that, Applied to the server side, the method includes: Aggregate the encrypted update parameters uploaded by at least two vehicle terminals into model aggregation parameters; Update the global agent model based on the model aggregation parameters; The global intelligent agent model is sent to the vehicle.

7. The method according to claim 6, characterized in that, The aggregated encrypted update parameters uploaded by at least two vehicle terminals are the model aggregation parameters, including: Obtain encrypted update parameters uploaded by at least two vehicle terminals, and determine the aggregation calculation coefficients for each vehicle terminal; For each of the aforementioned vehicle terminals, the encrypted update parameters are weighted based on the aggregated calculation coefficients to obtain weighted update parameters; Secure multi-party computation is performed on each of the weighted update parameters to obtain the model aggregation parameters.

8. The method according to claim 7, characterized in that, The determination of the aggregation calculation coefficients for each of the vehicle ends includes: The online status, network conditions, and computing resources of each vehicle terminal are extracted as vehicle performance indicators; The aggregation calculation coefficients for each of the vehicle terminals are determined according to the vehicle performance indicators.

9. The method according to claim 8, characterized in that, The determination of the aggregate calculation coefficient for each of the vehicle terminals according to the vehicle performance indicators includes at least one of the following: The aggregate calculation coefficients corresponding to each of the vehicle performance indicators are determined by a neural network model or a machine learning model. The index values ​​of each of the vehicle performance indicators are determined, and the proportion of each index value to the sum of all the index values ​​is used as the aggregation calculation coefficient.

10. The method according to any one of claims 6-9, characterized in that, The step of sending the global intelligent agent model to the vehicle includes: Extract the model weight values ​​of the global agent model; The model weight values ​​are sent to the vehicle by invoking a secure communication protocol.

11. A model training method, characterized in that, Applied to a model training system, the model training system including a server and a vehicle, the method includes: The local intelligent agent model on the vehicle is trained locally based on the local feature data of vehicle usage data to obtain model update parameters; The model update parameters are processed into encrypted update parameters on the vehicle side based on security encryption rules; Aggregate at least two of the encrypted update parameters from the vehicle end into model aggregation parameters; The global agent model on the server is updated according to the model aggregation parameters, and the global agent model is used as the local agent model on the vehicle.

12. A model training device, characterized in that, Applied to the vehicle end, the device includes: The local training module is used to train the local agent model locally based on local feature data of vehicle usage data, and obtain model update parameters. The upper-level training module is used to generate encrypted update parameters for the model update parameters based on security encryption rules, and transmit the encrypted update parameters to the server. The model update module is used to receive the global agent model sent by the server.

13. A model training device, characterized in that, The device, applied to the server side, includes: The parameter aggregation module is used to aggregate encrypted update parameters uploaded by at least two vehicle terminals into model aggregation parameters; The model update module is used to update the global agent model based on the model aggregation parameters; The model distribution module is used to distribute the global intelligent agent model to the vehicle.

14. A model training device, characterized in that, Applied to a model training system, the model training system including a server and a vehicle, the device includes: The local training module is used to train the local intelligent agent model on the vehicle based on the local feature data of the vehicle usage data on the vehicle, and to obtain the model update parameters. The data desensitization module is used to process the model update parameters into encrypted update parameters on the vehicle side based on security encryption rules; The parameter aggregation module is used to aggregate at least two of the encrypted update parameters from the vehicle end into model aggregation parameters; The local update module is used to update the global agent model on the server side according to the model aggregation parameters, and use the global agent model as the local agent model on the vehicle side.

15. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the model training method according to any one of claims 1-11.

16. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the model training method according to any one of claims 1-11.

17. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the model training method according to any one of claims 1-11.

Images