A method and apparatus for protecting privacy data in autonomous driving based on federated learning

Abstract

This invention discloses a method for protecting privacy data in autonomous driving based on federated learning. It includes a first device applied to collaborating parties and at least two second devices, each second device applied to a target automaker. The first and second devices are communicatively connected. The method, applied to the first device, includes: for any one of the at least two target automakers, using the autonomous driving database corresponding to the target automaker and the model parameters of the automaker's pre-trained model as a shared privacy database; performing federated co-creation encryption on the at least two privacy databases to generate an encrypted sample database; each privacy database corresponding to one encrypted sample; and performing federated learning on the collaborating party's pre-trained model based on the encrypted sample database to generate a federated encrypted model. This ensures that federated modeling is performed without sharing privacy databases among the automakers, solving the data silo problem, protecting privacy, and improving machine learning efficiency.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, and particularly relates to a method and apparatus for protecting privacy data in autonomous driving based on federated learning. Background Technology

[0002] Currently, the focus of artificial intelligence research both domestically and internationally remains on leveraging next-generation information technologies, represented by AI, to address the digitalization and intelligentization of functions, applications, and services across various fields. The secure use and sharing of privacy-related data has not yet been fully considered. To address the problem of data silos while protecting data privacy, researchers have proposed federated learning. Federated learning allows multiple participants to build a general and robust machine learning model without sharing data, thereby solving the privacy leakage problem caused by centralized data. It has become the most widely used privacy data protection solution.

[0003] Most existing federated learning research employs single-key homomorphic encryption schemes, where all clients share the same key pair. If a client actively intercepts encrypted data sent by other clients, it can directly decrypt the data using its own key, leading to the leakage of the intercepted client's model and local privacy data. While some federated learning schemes are based on multi-key encryption, they still suffer from problems such as complex ciphertext structures causing the ciphertext size to grow exponentially with the number of participants, and the need to pre-define the number of participants, failing to meet flexible multi-hop requirements. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention provides a method and apparatus for protecting privacy data in autonomous driving based on federated learning. This method enables the privacy databases of multiple automakers to be available but not visible during the creation of a federated encrypted model. This ensures that participating automakers can perform federated modeling without sharing their privacy databases, thus solving the data silo problem in the federated modeling process, achieving privacy protection, and improving machine learning efficiency.

[0005] According to a first aspect of the present invention, a method for protecting privacy data in autonomous driving based on federated learning is provided, characterized in that it includes a first device applied to a collaborating party and at least two second devices, each second device applied to a target vehicle manufacturer; the first device and the second devices are communicatively connected; the method is applied to the first device and includes: for any one of the at least two target vehicle manufacturers: using the autonomous driving database corresponding to the target vehicle manufacturer and the model parameters of the vehicle manufacturer's pre-trained model as the privacy database corresponding to the target vehicle manufacturer; wherein the vehicle manufacturer's pre-trained model is used to instruct the second device corresponding to the target vehicle manufacturer to generate a model based on the autonomous driving database through machine learning and / or deep learning; the autonomous driving database includes at least: vehicle manufacturer name, user ID, and autonomous driving-related agency regulations; performing federated co-creation encryption processing on the at least two privacy databases to generate an encrypted sample database; wherein the encrypted sample database includes at least two encrypted samples; each privacy database corresponds to one encrypted sample; and performing federated learning on the collaborating party's pre-trained model based on the encrypted sample database to generate a federated encrypted model.

[0006] Optionally, the step of performing federated co-creation encryption processing on the at least two privacy databases to generate an encrypted sample database includes: for any one of the at least two privacy databases, performing vector transformation processing on the privacy database to generate a transformed vector; performing centering processing on the transformed vector to output a centralized vector; combining the centralized vector and the transformed vectors of the remaining privacy databases to form a candidate sample data group corresponding to the privacy database; performing weighted processing on the at least two candidate sample data groups to generate a quasi-sample data group; wherein, the quasi-sample data group includes the database vector corresponding to the autonomous driving database of each target car company and the model parameter vector corresponding to the pre-trained model of the car company; obtaining the model parameter vector corresponding to each target car company from the quasi-sample data group, and calculating the average value of the model parameter vectors of all target car companies to obtain the total model parameters; and generating an encrypted sample database based on the total model parameters and the database vector corresponding to each target car company.

[0007] Optionally, when the at least two candidate sample data groups include a first candidate sample data group corresponding to a first target automaker, a second candidate sample data group corresponding to a second target automaker, and a third candidate sample data group corresponding to a third target automaker; the first candidate sample data group includes a first centered vector, a second transformed vector, and a third transformed vector; the second candidate sample data group includes a first transformed vector, a second centered vector, and a third transformed vector; and the third candidate sample data group includes a first transformed vector, a second transformed vector, and a third centered vector.

[0008] The step of weighting at least two candidate sample data groups to generate quasi-sample data groups includes: determining a first weight corresponding to the first candidate sample data group, a second weight corresponding to the second candidate sample data group, and a third weight corresponding to the third candidate sample data group based on an attention mechanism or preset rules; applying the first weight to the first centered vector in the first candidate sample data group, applying the second weight to the first transformed vector in the second candidate sample data group, and applying the third weight to the first transformed vector in the third candidate sample data group, and performing summation processing to output a first database vector and a first model parameter vector corresponding to the first target car company; applying the second weight to the second centered vector in the second candidate sample data group and applying the first weight to the second transformed vector in the first candidate sample data group, so as to... A third weight is applied to the second transformed vector in the third candidate sample data group, and the vectors are summed to output the second database vector and the second model parameter vector corresponding to the second target car company. A third weight is applied to the third centered vector in the third candidate sample data group, a first weight is applied to the third transformed vector in the first candidate sample data group, and a second weight is applied to the third transformed vector in the second candidate sample data group, and the vectors are summed to output the third database vector and the third model parameter vector corresponding to the third target car company. The first database vector and the first model parameter vector corresponding to the first target car company, the second database vector and the second model parameter vector corresponding to the second target car company, and the third database vector and the third model parameter vector corresponding to the third target car company are determined as the quasi-sample data group.

[0009] Optionally, before performing vector transformation processing on the privacy database, the method further includes: performing homomorphic encryption processing on the privacy database to output homomorphically encrypted data; and performing multi-party computation on the homomorphically encrypted data to output the multi-party computed privacy database.

[0010] Optionally, the privacy database sent by the second device is encrypted based on the federated encryption model to output encrypted data; an encrypted database is generated based on several of the encrypted data.

[0011] Optionally, the method further includes: receiving a data request sent by a second device of the requesting vehicle manufacturer; determining the requested vehicle manufacturer based on the data request; sending the data request to the second device corresponding to the requested vehicle manufacturer; upon receiving a consent to share instruction sent by the requested vehicle manufacturer, selecting encrypted data corresponding to the data request from the encrypted database; and sending the encrypted data to the second device corresponding to the requesting vehicle manufacturer.

[0012] According to a second aspect of the present invention, a method for protecting privacy data in autonomous driving based on federated learning is provided, applied to a second device of a target automaker; comprising: performing compliance data processing on the autonomous driving database corresponding to the target automaker to output compliant federated data; performing machine learning and / or deep learning on the compliant federated data to generate a pre-trained model of the automaker; and sending the model parameters corresponding to the pre-trained model of the automaker and the autonomous driving database to the first device.

[0013] Optionally, the method further includes: receiving encrypted data sent by the first device; optimizing the pre-trained model of the car manufacturer based on the encrypted data to generate a quasi-model of the car manufacturer.

[0014] Optionally, the step of performing compliance data processing on the autonomous driving database corresponding to the target automaker and outputting compliant federated data includes: performing multi-party data standardization processing on the autonomous driving database corresponding to the target automaker based on the local data standardization compliance agreement signed by all parties of the target automaker, and outputting a standardized database; performing local preprocessing on the standardized database and outputting preprocessed data; and performing federated view standardization compliance processing on the preprocessed data and outputting compliant federated data.

[0015] According to a third aspect of the present invention, an autonomous driving privacy data protection device based on federated learning is also provided, comprising a first device applied to a collaborating party and at least two second devices, each second device applied to a target vehicle manufacturer; the first device and the second devices are communicatively connected; the device is applied to the first device; comprising: a first determining module, configured to, for any one of the at least two target vehicle manufacturers, use the autonomous driving database corresponding to the target vehicle manufacturer and the model parameters of the vehicle manufacturer's pre-trained model as the privacy database corresponding to the target vehicle manufacturer; wherein, the vehicle manufacturer's pre-trained model is used to instruct the second device corresponding to the target vehicle manufacturer to generate a model based on the autonomous driving database through machine learning and / or deep learning; the autonomous driving database includes at least: vehicle manufacturer name, user ID, and autonomous driving-related agency regulations; an encryption processing module, configured to perform federated co-creation encryption processing on the privacy databases corresponding to the at least two target vehicle manufacturers to generate an encrypted sample database; and a model generation module, configured to perform federated learning on the collaborating party's pre-trained model based on the encrypted sample database to generate a federated encrypted model.

[0016] According to a fourth aspect of the present invention, an autonomous driving privacy data protection device based on federated learning is applied to a second device of a target automaker; comprising: a compliance data module for performing compliance data processing on the autonomous driving database corresponding to the target automaker and outputting compliant federated data; a pre-training module for performing machine learning and / or deep learning on the compliant federated data to generate a pre-trained model of the automaker; and a sending module for sending the model parameters corresponding to the pre-trained model of the automaker and the autonomous driving database to the first device.

[0017] According to a fifth aspect of the present invention, an electronic device is also provided, comprising: a processor; a memory for storing executable instructions of the processor; the processor being configured to read the executable instructions from the memory and execute the instructions to implement the method as described in the first or second aspect.

[0018] According to a sixth aspect of the present invention, a computer-readable medium is also provided, on which a computer program is stored, wherein the program, when executed by a processor, implements the method as described in the first or second aspect.

[0019] This invention provides a method for protecting privacy data in autonomous driving based on federated learning, including a first device applied to a collaborating party and at least two second devices, each second device applied to a target automaker; the first device and the second devices are communicatively connected; the method is applied to the first device and includes: for any one of the at least two target automakers: using the autonomous driving database corresponding to the target automaker and the model parameters of the automaker's pre-trained model as the privacy database corresponding to the target automaker; wherein, the automaker's pre-trained model is used to instruct the second device corresponding to the target automaker to generate a model based on the autonomous driving database through machine learning and / or deep learning; the autonomous driving database includes at least: the automaker's name, user ID, and autonomous driving-related agency regulations; performing federated co-creation encryption processing on the at least two privacy databases to generate an encrypted sample database; wherein, the encrypted sample database includes at least two encrypted samples; each privacy database corresponds to one encrypted sample; and performing federated learning on the collaborating party's pre-trained model based on the encrypted sample database to generate a federated encrypted model. This embodiment employs federated co-creation encryption technology to encrypt the privacy databases generated during the autonomous driving process of multiple target automakers, and performs federated learning based on the encrypted sample database generated by the encryption process. This ensures that each participating automaker can perform federated modeling without sharing the privacy database, thus solving the data silo problem, achieving privacy protection, and improving machine learning efficiency. Attached Figure Description

[0020] The following sections will describe some specific embodiments of the invention in detail by way of example and not limitation, with reference to the accompanying drawings. The same reference numerals in the drawings denote the same or similar parts or portions. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. In the drawings:

[0021] Figure 1 A flowchart illustrating an embodiment of the autonomous driving privacy data protection method based on federated learning provided by the present invention;

[0022] Figure 2 This is a schematic diagram of the process for generating an encrypted sample database according to an embodiment of the present invention;

[0023] Figure 3 A flowchart illustrating a method for protecting privacy data in autonomous driving based on federated learning, provided as another embodiment of the present invention;

[0024] Figure 4 A schematic diagram of a federated encryption model framework based on federated co-creation provided in an embodiment of the present invention;

[0025] Figure 5 A schematic diagram of the structure of an autonomous driving privacy data protection device based on federated learning provided in an embodiment of the present invention;

[0026] Figure 6 This is a schematic diagram of a privacy data protection device for autonomous driving based on federated learning, provided as another embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0028] In 2016, Google proposed the concept of federated learning (Konen et al., 2016). Federated learning is a form of distributed computing that enables multiple parties to collaborate on highly efficient machine learning while ensuring data exchange security and individual privacy. Its key characteristic is that participant data is available but not visible. It involves training on local data, uploading updated parameters to a server, and then aggregating the parameters to obtain the total set. Compared to traditional machine learning methods, federated learning not only improves learning efficiency but also ensures that participants can perform federated modeling without sharing data, thus solving the data silo problem and protecting privacy.

[0029] Assume N automakers, each with access to autonomous driving user data, participate in modeling. Each automaker participant, Ui, possesses its own dataset Di. Federated learning no longer integrates all data for unified training to obtain a model. Instead, each automaker participant trains its local data based on the initial parameters wG transmitted from the server, obtaining new parameters wGi′. The updated parameter values ​​δi = wGi′ - wG are then transmitted to the server, where they are aggregated to obtain new overall parameters wG′ = wG + f(δ1; δ2; ...; δn). After multiple iterations, the overall trained model MFed is finally obtained. Federated learning requires ensuring that the difference between the performance of MFed (VFed) and the performance of model MSum (VSum) is sufficiently small, i.e., |VFed - VSum| < ε, where ε is an arbitrarily small positive value (Yang et al., 2019).

[0030] like Figure 1 The diagram shown is a flowchart illustrating a method for protecting privacy data in autonomous driving based on federated learning, according to an embodiment of the present invention.

[0031] A method for protecting privacy data in autonomous driving based on federated learning includes a first device applied to a collaborating party and at least two second devices, each second device applied to a target automaker; the first device and the second devices are communicatively connected; the method is applied to the first device; and includes at least the following steps:

[0032] S101, for any one of at least two target car companies: the autonomous driving database corresponding to the target car company and the model parameters of the car company's pre-trained model are used together as the privacy database corresponding to the target car company; wherein, the car company's pre-trained model is used to instruct the second device corresponding to the target car company to generate a model based on the autonomous driving database through machine learning and / or deep learning; the autonomous driving database includes at least: car company name, user ID, and autonomous driving-related agency regulations;

[0033] S102, perform federated co-creation encryption processing on at least two privacy databases to generate an encrypted sample database; wherein, the encrypted sample database includes at least two encrypted samples; each privacy database corresponds to one encrypted sample;

[0034] S103, based on the encrypted sample database, performs federated learning on the pre-trained model of the collaborators to generate a federated encrypted model.

[0035] In S101, regulations related to autonomous driving include at least: data privacy protection laws, relevant autonomous driving regulations, and traffic laws. The target automakers' second devices locally store their privacy databases and are authorized to access their own privacy databases. To protect privacy data and meet the requirements of the current Personal Information Protection Act (Draft), target automakers typically do not share or actively exchange their privacy databases; therefore, the privacy databases of each target automaker used for federated learning remain independent.

[0036] It should be noted that the pre-trained models for each target automaker can be the same or different. For example, the pre-trained model for the first target automaker is used to instruct the first target automaker to train its autonomous driving database based on machine learning. The pre-trained model for the second target automaker is used to instruct the second target automaker to train its autonomous driving database based on machine learning and deep learning.

[0037] In S102 and S103, each target automaker has a corresponding second device.

[0038] There are no restrictions on the specific process of federal co-creation encryption; the goal is simply to prevent the target automakers from speculating on the original privacy databases. For example, the first device can achieve federal co-creation encryption by aligning and encrypting the privacy databases of multiple target automakers. Through federal co-creation encryption, the first device not only encrypts the autonomous driving databases of each target automaker but also encrypts the model parameters of each automaker.

[0039] The pre-trained model of the collaborating party can be an initialized random model or a model trained based on the autonomous driving database of at least one target car company.

[0040] Federated learning can be categorized based on the overlap between user and feature dimensions, and according to the data distribution in encrypted sample databases. These categories include horizontal federated learning, vertical federated learning, and federated transfer learning. Vertical federated learning, based on different datasets with significant overlap in data features but minimal overlap in users, splits the dataset according to the user dimension and uses the portion of data with the same features across participants but not identical user characteristics for training. Horizontal federated learning, also based on different datasets with significant overlap in users but minimal overlap in data features, splits the dataset according to the data feature dimension and uses the portion of data from participants targeting the same user but with not identical data features for training. Federated transfer learning, based on multiple datasets with minimal overlap in users and data features, does not split the data but instead uses transfer learning to overcome insufficient data or labels. In the practice of protecting user privacy data in autonomous driving, federated learning applications can also be categorized into ToB and ToC scenarios. For ToB scenarios, the main service targets are car manufacturers and related institutions; for ToC scenarios, the federated modeling participants are mainly edge computing devices, typically involving a large number of participants with low computing power.

[0041] In this embodiment, the first device performs federated co-creation encryption processing on the privacy databases corresponding to at least two target car companies; it uses the generated encrypted sample database to perform federated learning on the pre-trained models of the collaborators to generate federated encrypted models; thus, while ensuring that each participating car company performs federated modeling without sharing the privacy database, it solves the data silo problem and achieves privacy protection.

[0042] like Figure 2 The diagram shown is a flowchart illustrating the process of generating an encrypted sample database according to an embodiment of the present invention.

[0043] Federated co-creation encryption is performed on the at least two privacy databases to generate an encrypted sample database; including:

[0044] S201, for any one of at least two privacy databases: perform vector transformation on the privacy database to generate a transformed vector; perform centering on the transformed vector to output a centered vector; and combine the centered vector with the transformed vectors of the remaining privacy databases to form a candidate sample data group corresponding to the privacy database.

[0045] S202, weighting at least two candidate sample data groups to generate quasi-sample data groups; wherein, the quasi-sample data groups include the database vector corresponding to the autonomous driving database of each target car company and the model parameter vector corresponding to the car company's pre-trained model;

[0046] S203: Obtain the model parameter vector corresponding to each target car company from the quasi-sample data group, and calculate the average value of the model parameter vectors of all target car companies to obtain the total model parameters;

[0047] S204 generates an encrypted sample database based on the total parameters of the model and the database vector corresponding to each target car company.

[0048] In this embodiment, there are at least two target automakers for federated learning.

[0049] For example, when the at least two candidate sample data groups include a first candidate sample data group corresponding to a first target automaker, a second candidate sample data group corresponding to a second target automaker, and a third candidate sample data group corresponding to a third target automaker; the first candidate sample data group includes a first centered vector, a second transformed vector, and a third transformed vector; the second candidate sample data group includes a first transformed vector, a second centered vector, and a third transformed vector; the third candidate sample data group includes a first transformed vector, a second transformed vector, and a third centered vector.

[0050] The step of weighting at least two candidate sample data groups to generate quasi-sample data groups includes: determining a first weight corresponding to the first candidate sample data group, a second weight corresponding to the second candidate sample data group, and a third weight corresponding to the third candidate sample data group based on an attention mechanism or preset rules; applying the first weight to the first centered vector in the first candidate sample data group, applying the second weight to the first transformed vector in the second candidate sample data group, and applying the third weight to the first transformed vector in the third candidate sample data group, and performing summation processing to output a first database vector and a first model parameter vector corresponding to the first target car company; applying the second weight to the second centered vector in the second candidate sample data group and applying the first weight to the second transformed vector in the first candidate sample data group, so as to... A third weight is applied to the second transformed vector in the third candidate sample data group, and the vectors are summed to output the second database vector and the second model parameter vector corresponding to the second target automaker. A third weight is applied to the third centralized vector in the third candidate sample data group, a first weight is applied to the third transformed vector in the first candidate sample data group, and a second weight is applied to the third transformed vector in the second candidate sample data group, and these are summed to output the third database vector and the third model parameter vector corresponding to the third target automaker. The first database vector and the first model parameter vector corresponding to the first target automaker, the second database vector and the second model parameter vector corresponding to the second target automaker, and the third database vector and the third model parameter vector corresponding to the third target automaker are determined as the quasi-sample data group. The average value of the first model parameter vector, the second model parameter vector, and the third model parameter vector is calculated to obtain the total model parameters. Based on the first database vector, the second database vector, the third database vector, and the total model parameters, a federated co-creation encrypted database is generated.

[0051] Here, "Federated Co-creation Encryption Processing" is also known as "Centralized Data and Weighted Average Algorithm-Federated Learning" (CWA-FL).

[0052] This embodiment employs federated co-creation encryption to encrypt the privacy databases of multiple target automakers. This effectively protects data privacy, prevents both target automakers and multiple target automakers from speculating on the original privacy database, and enables in-depth analysis and protection of distributed data. It also improves the accuracy of training the federated encryption model.

[0053] In a preferred embodiment of this example, before performing vector transformation processing on the privacy database, the method further includes: performing homomorphic encryption processing on the privacy database to output homomorphically encrypted data; and performing multi-party computation on the homomorphically encrypted data to output the multi-party computed privacy database.

[0054] Therefore, by combining homomorphic encryption, multi-party computation, and federated co-creation encryption, the privacy databases of all target automakers can be effectively protected, preventing them from obtaining the original privacy databases through reverse engineering techniques. This enhances the level of protection for the original privacy databases and enables the federated encryption model to be usable but not visible to the privacy databases of each automaker during training.

[0055] In a preferred embodiment of this example, the method further includes: encrypting the privacy database sent by the second device based on the federated encryption model, and outputting encrypted data; and generating an encrypted database based on a plurality of the encrypted data.

[0056] After receiving the privacy database from the second device, the first device directly encrypts the privacy database and outputs encrypted data. Since each privacy database corresponds to one piece of encrypted data, at least two pieces of encrypted data form an encrypted database.

[0057] In a preferred embodiment of this example, the method further includes: receiving a data request sent by a second device requesting a car manufacturer; determining the requested car manufacturer based on the data request; sending the data request to the second device corresponding to the requested car manufacturer; upon receiving a consent to share instruction sent by the requested car manufacturer, selecting encrypted data corresponding to the data request from the encrypted database; and sending the encrypted data to the second device corresponding to the requesting car manufacturer.

[0058] Here, since the requested automaker is used to instruct other target automakers besides the requesting automaker for federated learning, the encrypted data corresponding to the data request does not include the encrypted data corresponding to the requesting automaker.

[0059] In this embodiment, the first device sends encrypted data corresponding to the data request to the requested car company, making the encrypted data usable but not visible to the requested car company. This effectively protects the privacy databases of all car companies, enabling them to share training data. This is beneficial for training the quasi-models of all car companies, solving the problem of data silos when car companies are training models, and improving the efficiency of machine learning.

[0060] like Figure 3 The diagram shown is a flowchart illustrating a method for protecting privacy data in autonomous driving based on federated learning, provided in another embodiment of the present invention.

[0061] A federated learning-based method for protecting privacy data in autonomous driving, applied to a second device of a target automaker; including:

[0062] S301 performs compliance data processing on the autonomous driving database of the target automaker and outputs compliant federated data;

[0063] S302, perform machine learning and / or deep learning on compliant federal data to generate pre-trained models for automakers;

[0064] S303, send the model parameters corresponding to the pre-trained model of the car manufacturer and the autonomous driving database to the first device.

[0065] Specifically, firstly, the second device performs multi-party data standardization processing on the autonomous driving database corresponding to the target automaker based on the local data standardization and compliance agreement signed by all parties, outputting a standardized database; secondly, the second device performs local preprocessing on the standardized database, outputting preprocessed data; then, the second device performs federated view standardization and compliance processing on the preprocessed data, outputting compliant federated data. Here, local preprocessing includes: deduplication, standardization, and error correction, etc.

[0066] This solves the problem of data silos during the training of federated encryption models and improves the accuracy of federated encryption model training.

[0067] In a preferred embodiment of this example, the method further includes: receiving encrypted data sent by the first device; and optimizing the pre-trained model of the car manufacturer based on the encrypted data to generate a quasi-model of the car manufacturer.

[0068] After receiving encrypted data from the first device, the second device optimizes or trains the pre-trained model of the car company based on the encrypted data, generating a quasi-model of the car company. Thus, the second device can retrain the pre-trained model of the car company based on the privacy database of multiple target car companies, improving the prediction coverage of the quasi-model of the car company.

[0069] A privacy data protection system for autonomous driving based on federated learning includes a first device applied to a collaborating party and at least two second devices, each of the second devices applied to a target automaker; the first device and the second devices are communicatively connected; and the system includes at least the following steps:

[0070] S1, the second device performs compliance data processing on the autonomous driving database corresponding to the target car company and outputs compliant federated data; performs machine learning and / or deep learning on the compliant federated data to generate a pre-trained model for the car company; and sends the model parameters corresponding to the pre-trained model for the car company and the autonomous driving database to the first device.

[0071] S2, for any one of at least two target car companies: the first device uses the autonomous driving database corresponding to the target car company and the model parameters of the car company's pre-trained model as the privacy database corresponding to the target car company; wherein, the car company's pre-trained model is used to instruct the second device corresponding to the target car company to generate a model based on the autonomous driving database through machine learning and / or deep learning; the autonomous driving database includes at least: car company name, user ID, and autonomous driving-related agency regulations.

[0072] S3, for any one of at least two privacy databases: the first device performs vector transformation processing on the privacy database to generate a transformed vector; the transformed vector is centered to output a centered vector; the centered vector and the transformed vectors of the remaining privacy databases are used to form a candidate sample data group corresponding to the privacy database; the at least two candidate sample data groups are weighted to generate a quasi-sample data group; wherein, the quasi-sample data group includes the database vector corresponding to the autonomous driving database of each target car company and the model parameter vector corresponding to the pre-trained model of the car company; the model parameter vector corresponding to each target car company is obtained from the quasi-sample data group, and the average value of the model parameter vectors of all target car companies is calculated to obtain the total model parameters; based on the total model parameters and the database vector corresponding to each target car company, an encrypted sample database is generated. The encrypted sample database includes at least two encrypted samples; each privacy database corresponds to one encrypted sample.

[0073] S4, the first device performs federated learning on the pre-trained model of the collaborator based on the encrypted sample database to generate a federated encrypted model.

[0074] S5, the first device encrypts the privacy database sent by the second device based on the federated encryption model and outputs encrypted data; based on several of the encrypted data, an encrypted database is generated.

[0075] S6, the first device receives a data request sent by the second device of the requesting car manufacturer; determines the requested car manufacturer based on the data request; sends the data request to the second device corresponding to the requested car manufacturer; upon receiving the consent sharing instruction sent by the requested car manufacturer, selects encrypted data corresponding to the data request from the encrypted database; and sends the encrypted data to the second device corresponding to the requesting car manufacturer.

[0076] S7, the second device receives encrypted data sent by the first device; and optimizes the pre-trained model of the car manufacturer based on the encrypted data to generate a quasi-model of the car manufacturer.

[0077] In this embodiment, the first device interacts with at least two second devices. The first device has full authority to manage the entire federated learning process. When processing large-scale distributed data, CWA-FL can avoid the direct transmission of sensitive data by performing model pre-training and updating in a local privacy database, thereby protecting the user's data privacy.

[0078] The following section provides a detailed description of a federated learning-based method for protecting privacy data in autonomous driving, based on specific application scenarios.

[0079] A method for protecting privacy data in autonomous driving based on federated learning, the specific operation process of which includes:

[0080] (1) Initialization: All second devices initialize the random model at the start of the task to obtain the initial task model;

[0081] (2) First device selects second device: Randomly select a portion of the second devices from all the second devices for model pre-training;

[0082] (3) Model pre-training: The selected second device trains the initial task model on its local autonomous driving database until the gradient directions of each device converge, generating the pre-trained model of the car manufacturer.

[0083] (4) Model update and upload: Each second device uploads the trained pre-trained model of the car manufacturer to the first device;

[0084] (5) Federated co-creation encryption processing: After the first device is pre-trained, the first device will convert all the uploaded car manufacturers' pre-trained models into vectors and use the centralized data exchange fusion weighted average federated algorithm to group and calculate all the second devices and their corresponding models.

[0085] (6) Model optimization: The first device performs federated learning on the pre-trained model of the collaborating party based on the encrypted sample database generated after the federated co-creation encryption process, and adjusts its parameters to optimize performance.

[0086] (7) Model convergence detection: The first device checks whether the pre-trained model of its collaborators has converged on the data of the new task. If it has converged, a federated encryption model is generated; and the first device will begin training for the next task. Otherwise, return to step (2) and continue the training process.

[0087] This embodiment provides an efficient learning strategy based on federated learning. Its main goal is to effectively protect data privacy, prevent two-party and multi-party speculation on the original data, and perform in-depth analysis and protection of distributed data. CWA-FL adopts a centralized federated learning architecture, in which the first device interacts with multiple second devices and manages the entire federated learning process.

[0088] like Figure 4 The diagram shown is a schematic of a federated encryption model framework based on federated co-creation provided in an embodiment of the present invention.

[0089] In the practical application of solutions for protecting user privacy data in autonomous driving, we focus on the underlying code, algorithm optimization, and intelligent platform construction, through independent research and development innovation and intelligent management. The model for protecting user privacy data in autonomous driving, based on the federated co-creation approach, mainly includes a customized framework consisting of a parent framework (a one-stop intelligent collaboration system) and a sub-framework (a federated co-creation modeling framework platform).

[0090] Main framework: One-stop intelligent collaboration system.

[0091] This one-stop intelligent collaboration system, based on federated co-creation learning, achieves intelligent prediction and interpretable innovative attempts. It solidifies the collaboration chain from the data layer, federation layer, algorithm layer, and optimization layer. In terms of practical functionality, it is implemented by connecting and parallelizing the multi-party data source module, multi-party federated co-creation module, local pre-training and formal training module, and model optimization and update module through a privacy and security module. The first device includes: the multi-party data source module, the multi-party federated co-creation module, the local pre-training and formal training module, the privacy and security module, and the model optimization and update module; the second device includes: the multi-party data source module, the multi-party federated co-creation module, the local pre-training and formal training module, and the model optimization and update module.

[0092] In the multi-source data generation module, autonomous driving data is categorized into structured and unstructured data. This data originates from automakers, autonomous driving users, and relevant regulations and structures. The multi-source data generation module then sends the autonomous driving database, formed from this data, to the multi-party federated co-creation module.

[0093] The multi-party federated co-creation module performs compliance data processing on the autonomous driving database. This compliance data processing includes multi-party data standardization, local preprocessing, and federated view standardization compliance. For example: first, based on the local data standardization compliance agreements signed by all parties, the autonomous driving database undergoes multi-party data standardization processing to obtain standardized data; second, based on methods such as deduplication, standardization, and error correction, the standardized data undergoes local preprocessing to obtain preprocessed data; finally, the preprocessed data undergoes federated view standardization compliance processing through federated normalization to output compliant federated data. The multi-party federated co-creation module then delivers the compliant federated data to the local pre-training and formal training modules.

[0094] The local pre-training and formal training modules include first performing machine learning on compliant federated data; then performing deep learning on the machine learning results to output a pre-trained model for the car company; among them, machine learning includes: functionalization, regression, and multi-classification, etc.; deep learning includes CNN, RNN, and BERT, etc.

[0095] The local pre-training and formal training modules send the pre-trained models from car manufacturers or collaborators to the model optimization and update module.

[0096] The model optimization and update process includes several units; these units include: asymmetric encryption of model information, compliant transmission, collaborative integration of gradient parameters, optimization of the federated co-creation learning model, and updates to each model.

[0097] The privacy and security module includes homomorphic encryption, multi-party computation, and a centralized data exchange fusion weighted average federated algorithm.

[0098] Sub-framework: Federation co-creation model framework platform

[0099] The federated co-creation modeling framework platform, through parallel encryption algorithms, asynchronous computation, and federated co-creation learning technologies, improves learning efficiency while ensuring data security, and gradually achieves the ability to integrate data on a scale of hundreds of millions. This framework more efficiently and accurately utilizes and mines the value of data from all parties, achieving a win-win situation. It has three main characteristics: First, regarding data and model privacy, different active and passive data participants (see Figure 1: Database 1 - Data Controller 1; Database 2 - Data Controller 2; ... Database q - Data Controller q) do not directly exchange local autonomous driving databases and model parameters, but instead exchange intermediate values ​​needed to update parameters. Second, in terms of information transmission, it proposes the concept of a "centralized data exchange, fusion, and weighted average algorithm," where collaborating parties complete data processing, making data exchange independent of participating parties and making it difficult for other collaborating parties to infer or guess the data. Third, it adopts an asynchronous computation framework to improve the training speed of the federated encrypted model. The parent framework serves as the foundation of the data system, while the sub-frameworks complete the federated co-creation modeling computation. The parent and sub-frameworks work together to complete the task of protecting the privacy data of autonomous driving users.

[0100] This embodiment shares the privacy databases of various car manufacturers while effectively protecting their privacy databases, thereby solving the problem of data silos during the training of federated encryption models and car manufacturer quasi-models, and improving the accuracy of model training.

[0101] like Figure 5 The diagram shown is a structural schematic of an autonomous driving privacy data protection device based on federated learning provided in an embodiment of the present invention.

[0102] An autonomous driving privacy data protection device based on federated learning includes a first device applied to a collaborating party and at least two second devices, each second device applied to a target automaker; the first device and the second devices are communicatively connected; the device 500 is applied to the first device; and includes: a first determining module 501, used for, for any one of the at least two target automakers: using the autonomous driving database corresponding to the target automaker and the model parameters of the automaker's pre-trained model as the privacy database corresponding to the target automaker; wherein, the automaker's pre-trained model is used to instruct the second device corresponding to the target automaker to generate a model based on the autonomous driving database through machine learning and / or deep learning; the autonomous driving database includes at least: automaker name, user ID, and autonomous driving-related agency regulations; an encryption processing module 502, used for performing federated co-creation encryption processing on the privacy databases corresponding to the at least two target automakers to generate an encrypted sample database; and a model generation module 503, used for performing federated learning on the collaborating party's pre-trained model based on the encrypted sample database to generate a federated encrypted model.

[0103] In a preferred embodiment of this example, the encryption module includes: a first encryption unit, configured to: perform vector transformation processing on any one of at least two privacy databases to generate a transformed vector; perform centering processing on the transformed vector to output a centralized vector; and combine the centralized vector with the transformed vectors of the remaining privacy databases to form a candidate sample data group corresponding to the privacy database; a second encryption unit, configured to perform weighted processing on at least two candidate sample data groups to generate a quasi-sample data group; wherein the quasi-sample data group includes a database vector corresponding to the autonomous driving database of each target automaker and a model parameter vector corresponding to the pre-trained model of the automaker; a third encryption unit, configured to obtain the model parameter vector corresponding to each target automaker from the quasi-sample data group and calculate the average value of the model parameter vectors of all target automakers to obtain the total model parameters; and a fourth encryption unit, configured to generate an encrypted sample database based on the total model parameters and the database vector corresponding to each target automaker.

[0104] In a preferred embodiment of this example, when the at least two candidate sample data groups include a first candidate sample data group corresponding to a first target automaker, a second candidate sample data group corresponding to a second target automaker, and a third candidate sample data group corresponding to a third target automaker; the first candidate sample data group includes a first centered vector, a second transformed vector, and a third transformed vector; the second candidate sample data group includes a first transformed vector, a second centered vector, and a third transformed vector; and the third candidate sample data group includes a first transformed vector, a second transformed vector, and a third centered vector.

[0105] The second encryption unit includes: a first determining subunit, used to determine, based on an attention mechanism or preset rules, a first weight corresponding to the first candidate sample data group, a second weight corresponding to the second candidate sample data group, and a third weight corresponding to the third candidate sample data group; a first weighting subunit, used to apply the first weight to the first centered vector in the first candidate sample data group, the second weight to the first transformed vector in the second candidate sample data group, and the third weight to the first transformed vector in the third candidate sample data group, and perform summation processing to output a first database vector and a first model parameter vector corresponding to the first target car company; and a second weighting subunit, used to apply the second weight to the second centered vector in the second candidate sample data group, the first weight to the second transformed vector in the first candidate sample data group, and the third weight to the third candidate sample data group. A third weight is applied to the second transformed vector in the sample data set, and the vectors are summed to output the second database vector and the second model parameter vector corresponding to the second target car company. A third weighting subunit is used to apply the third weight to the third centered vector in the third candidate sample data set, the first weight to the third transformed vector in the first candidate sample data set, and the second weight to the third transformed vector in the second candidate sample data set, and then sum them to output the third database vector and the third model parameter vector corresponding to the third target car company. A second determining unit is used to determine the first database vector and the first model parameter vector corresponding to the first target car company, the second database vector and the second model parameter vector corresponding to the second target car company, and the third database vector and the third model parameter vector corresponding to the third target car company as the quasi-sample data set.

[0106] In a preferred embodiment of this example, the encryption module further includes: a homomorphic encryption unit, used to perform homomorphic encryption processing on the privacy database and output homomorphically encrypted data; and a multi-party computation unit, used to perform multi-party computation on the homomorphically encrypted data and output the multi-party computed privacy database.

[0107] In a preferred embodiment of this invention, the device further includes: an encryption prediction module, used to encrypt the privacy database sent by the second device based on the federated encryption model and output encrypted data; and an encryption database generation module, used to generate an encryption database based on a plurality of the encrypted data.

[0108] In a preferred embodiment of this invention, the device further includes: a receiving module for receiving a data request sent by a second device requesting a car manufacturer; a second determining module for determining a requested car manufacturer based on the data request; a first sending module for sending the data request to the second device corresponding to the requested car manufacturer; a selecting module for selecting encrypted data corresponding to the data request from the encrypted database when receiving a consent-to-share instruction sent by the requested car manufacturer; and a second sending module for sending the encrypted data to the second device corresponding to the requesting car manufacturer.

[0109] like Figure 6 The diagram shown is a structural schematic of an autonomous driving privacy data protection device based on federated learning, provided in another embodiment of the present invention.

[0110] An autonomous driving privacy data protection device based on federated learning is applied to a second device of a target automaker. The device 600 includes: a compliance data module 601, used to perform compliance data processing on the autonomous driving database corresponding to the target automaker and output compliant federated data; a pre-training module 602, used to perform machine learning and / or deep learning on the compliant federated data to generate a pre-trained model of the automaker; and a sending module 603, used to send the model parameters corresponding to the pre-trained model of the automaker and the autonomous driving database to the first device.

[0111] In a preferred embodiment of this invention, the device further includes: a receiving module for receiving encrypted data sent by the first device; and a model optimization module for optimizing the pre-trained model of the car manufacturer based on the encrypted data to generate a quasi-model of the car manufacturer.

[0112] In a preferred embodiment of this example, the compliance data module includes: a multi-party data normalization unit, used to perform multi-party data normalization processing on the autonomous driving database corresponding to the target automaker based on the local data normalization compliance agreement signed by all parties, and output a normalized database; a local preprocessing unit, used to perform local preprocessing on the normalized database and output preprocessed data; and a compliance processing unit, used to perform federated view normalization compliance processing on the preprocessed data and output compliant federated data.

[0113] The aforementioned apparatus can execute the autonomous driving privacy data protection method based on federated learning provided in an embodiment of the present invention, and possesses the corresponding functional modules and beneficial effects for executing the autonomous driving privacy data protection method based on federated learning. Technical details not described in detail in this embodiment can be found in the autonomous driving privacy data protection method based on federated learning provided in an embodiment of the present invention.

[0114] The present invention also provides an electronic device, comprising: a processor; a memory for storing executable instructions of the processor; the processor being configured to read the executable instructions from the memory and execute the instructions to implement the federated learning-based autonomous driving privacy data protection method of the present invention.

[0115] In addition to the methods and apparatus described above, embodiments of this application may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in the methods according to various embodiments of this application described in the "Exemplary Methods" section above.

[0116] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0117] Furthermore, embodiments of this application may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps in the methods according to the following embodiments of this application described in the "Exemplary Methods" section above.

[0118] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may, for example, include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable 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.

[0119] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the application to the necessity of employing the aforementioned specific details for implementation.

[0120] The block diagrams of devices, apparatuses, devices, and systems involved in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0121] It should also be noted that in the apparatus, equipment, and methods of this application, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of this application.

[0122] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this application. Therefore, this application is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0123] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

[0124] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0125] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0126] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for protecting privacy data in autonomous driving based on federated learning, characterized in that, This includes a first device applied to a collaborating party and at least two second devices, each of which is applied to a target automaker. The first device and the second device are communicatively connected; The method is applied to a first device; including: For any one of at least two target automakers: the autonomous driving database corresponding to the target automaker and the model parameters of the automaker's pre-trained model are used together as the privacy database corresponding to the target automaker; wherein, the automaker's pre-trained model is used to instruct a second device corresponding to the target automaker to generate a model based on the autonomous driving database through machine learning and / or deep learning; the autonomous driving database includes at least: automaker name, user ID, and autonomous driving-related agency regulations; At least two privacy databases are subjected to federated co-creation encryption to generate an encrypted sample database; wherein the encrypted sample database includes at least two encrypted samples; each privacy database corresponds to one encrypted sample; Based on the encrypted sample database, federated learning is performed on the pre-trained model of the collaborating party to generate a federated encrypted model. The step of performing federated co-creation encryption on at least two privacy databases to generate an encrypted sample database includes: For any one of at least two privacy databases: perform vector transformation on the privacy database to generate a transformed vector; perform centering on the transformed vector to output a centered vector; and combine the centered vector with the transformed vectors of the remaining privacy databases to form a candidate sample data group corresponding to the privacy database. At least two candidate sample data groups are weighted to generate quasi-sample data groups; wherein, the quasi-sample data groups include the database vector corresponding to the autonomous driving database of each target car company and the model parameter vector corresponding to the car company's pre-trained model; The model parameter vector corresponding to each target car company is obtained from the quasi-sample data group, and the average value of the model parameter vectors of all target car companies is calculated to obtain the total model parameters. An encrypted sample database is generated based on the total parameters of the model and the database vector corresponding to each target car company.

2. The method according to claim 1, characterized in that, When the at least two candidate sample data groups include a first candidate sample data group corresponding to a first target automaker, a second candidate sample data group corresponding to a second target automaker, and a third candidate sample data group corresponding to a third target automaker; the first candidate sample data group includes a first centered vector, a second transformed vector, and a third transformed vector; the second candidate sample data group includes a first transformed vector, a second centered vector, and a third transformed vector; the third candidate sample data group includes a first transformed vector, a second transformed vector, and a third centered vector. The step of weighting at least two candidate sample data groups to generate quasi-sample data groups includes: Based on attention mechanisms or preset rules, determine the first weight corresponding to the first candidate sample data group, the second weight corresponding to the second candidate sample data group, and the third weight corresponding to the third candidate sample data group. Apply a first weight to the first centered vector in the first candidate sample data group, apply a second weight to the first transformed vector in the second candidate sample data group, and apply a third weight to the first transformed vector in the third candidate sample data group, and perform summation to output the first database vector and the first model parameter vector corresponding to the first target car company. The second centering vector in the second candidate sample data group is given a second weight, the second transformed vector in the first candidate sample data group is given a first weight, and the second transformed vector in the third candidate sample data group is given a third weight. These weights are then summed to output the second database vector and the second model parameter vector corresponding to the second target car company. The third centralized vector in the third candidate sample data group is given a third weight, the third transformed vector in the first candidate sample data group is given a first weight, and the third transformed vector in the second candidate sample data group is given a second weight. These weights are then summed to output the third database vector and the third model parameter vector corresponding to the third target car company. The first database vector and first model parameter vector corresponding to the first target car company, the second database vector and second model parameter vector corresponding to the second target car company, and the third database vector and third model parameter vector corresponding to the third target car company are determined as the quasi-sample data group.

3. The method according to claim 1, characterized in that, Before performing vector transformation on the privacy database, the process also includes: The privacy database is homomorphically encrypted, and homomorphically encrypted data is output. Perform multi-party computation on the homomorphically encrypted data and output the privacy database after multi-party computation.

4. The method according to claim 1, characterized in that, Also includes: The privacy database sent by the second device is encrypted based on the federated encryption model, and encrypted data is output. An encrypted database is generated based on several of the encrypted data.

5. The method according to claim 4, characterized in that, Also includes: Receive data requests sent by the automaker's second device; Based on the data request, the requested automaker is identified; The data request is sent to the second device corresponding to the requested automaker; Upon receiving the consent to share instruction sent by the requested automaker, the encrypted data corresponding to the data request is selected from the encrypted database; The encrypted data is sent to the second device corresponding to the requesting automaker.

6. A method for protecting privacy data in autonomous driving based on federated learning, characterized in that, Secondary equipment applied to target automakers; including: Perform compliance data processing on the autonomous driving database corresponding to the target automaker and output compliant federated data; Machine learning and / or deep learning are performed on the compliant federal data to generate pre-trained models for automakers; The model parameters corresponding to the pre-trained model of the car manufacturer and the autonomous driving database are sent to the first device; The step of performing compliance data processing on the autonomous driving database corresponding to the target automaker and outputting compliant federated data includes: performing multi-party data standardization processing on the autonomous driving database corresponding to the target automaker based on the local data standardization and compliance agreement signed by all parties of the target automaker, and outputting a standardized database; performing local preprocessing on the standardized database and outputting preprocessed data; and performing federated view standardization and compliance processing on the preprocessed data and outputting compliant federated data.

7. The method according to claim 6, characterized in that, Also includes: Receive encrypted data sent by the first device; Based on the encrypted data, the pre-trained model of the car company is optimized to generate a quasi-model of the car company.

8. A privacy data protection device for autonomous driving based on federated learning, characterized in that, This includes a first device applied to a collaborating party and at least two second devices, each of which is applied to a target automaker. The first device and the second device are communicatively connected; The device is applied to the first equipment; comprising: The first determining module is configured to, for any one of at least two target automakers, use the autonomous driving database corresponding to the target automaker and the model parameters of the automaker's pre-trained model as the privacy database corresponding to the target automaker; wherein, the automaker's pre-trained model is used to instruct a second device corresponding to the target automaker to generate a model based on the autonomous driving database through machine learning and / or deep learning; the autonomous driving database includes at least: automaker name, user ID, and autonomous driving-related agency regulations; The encryption processing module is used to perform federated co-creation encryption processing on at least two privacy databases to generate an encrypted sample database. The model generation module is used to perform federated learning on the pre-trained model of the collaborating party based on the encrypted sample database to generate a federated encrypted model. The encryption processing module includes: a first encryption unit, configured to: perform vector transformation processing on any one of at least two privacy databases to generate a transformed vector; perform centering processing on the transformed vector to output a centralized vector; and combine the centralized vector with the transformed vectors of the remaining privacy databases to form a candidate sample data group corresponding to the privacy database; a second encryption unit, configured to perform weighted processing on at least two candidate sample data groups to generate a quasi-sample data group; wherein the quasi-sample data group includes a database vector corresponding to the autonomous driving database of each target automaker and a model parameter vector corresponding to the pre-trained model of the automaker; a third encryption unit, configured to obtain the model parameter vector corresponding to each target automaker from the quasi-sample data group and calculate the average value of the model parameter vectors of all target automakers to obtain the total model parameters; and a fourth encryption unit, configured to generate an encrypted sample database based on the total model parameters and the database vector corresponding to each target automaker.

9. A privacy data protection device for autonomous driving based on federated learning, characterized in that, Secondary equipment applied to target automakers; including: The compliance data module is used to perform compliance data processing on the autonomous driving database corresponding to the target car company and output compliant federated data. The pre-training module is used to perform machine learning and / or deep learning on the compliant federated data to generate pre-trained models for automakers. The sending module is used to send the model parameters corresponding to the pre-trained model of the car manufacturer and the autonomous driving database to the first device; The compliance data module includes: a multi-party data normalization unit, used to perform multi-party data normalization processing on the autonomous driving database corresponding to the target automaker based on the local data normalization compliance agreement signed by all parties, and output a normalized database; a local preprocessing unit, used to perform local preprocessing on the normalized database and output preprocessed data; and a compliance processing unit, used to perform federated view normalization compliance processing on the preprocessed data and output compliant federated data.

10. A computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the method as claimed in any one of claims 1 to 5 or the method as claimed in any one of claims 6 to 7.

Images