Generative model privacy protection method suitable for intelligent Internet of Vehicles and related device

By using energy models and Hamiltonian dynamics simulation in intelligent vehicle networking, mapping high-dimensional data to low-dimensional space and introducing perturbations, the limitations of differential privacy generation models in high-dimensional data processing are solved, and privacy protection is improved and data quality, suitable for smart vehicle networking scenarios.

CN118940308AActive Publication Date: 2024-11-12ELECTRIC POWER RES INST CHINA SOUTHERN POWER GRID CO LTD +1
View PDF 7 Cites 0 Cited by

Patent Information

Application Number
CN202410928238.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-12
Estimated Expiration
2044-07-11

AI Technical Summary

Technical Problem

The existing differential privacy generation model has limitations when processing high-dimensional data, including difficulties in processing high-dimensional data, high computational complexity and low generated data quality, making it difficult to effectively protect privacy in smart vehicle networking scenarios.

Method used

By using energy models to generalize the popular features of high-dimensional data sets, projection technology is used to map to low-dimensional space, combining pre-trained generative models and residual enhancement networks, introducing perturbation projection vectors and Hamiltonian dynamics simulations to generate and meet the data distribution requirements. sample data.

Benefits of technology

It effectively reduces the risk of sensitive information leakage in high-dimensional data, enhances privacy protection, improves the quality of the generated data, and synthesizes data suitable for intelligent Internet of Vehicles, protects user privacy and promotes the safe development of Internet of Vehicles technology.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940308A_ABST
    Figure CN118940308A_ABST
Patent Text Reader

Abstract

The invention provides a generative model privacy protection method suitable for an intelligent Internet of Vehicles and a related device, and the method comprises the steps: generalizing a prevalence feature of a given private data set through an energy model, and mapping the private data set to a low-dimensional space, thereby obtaining a low-dimensional prevalence feature; a pre-trained first generation model is utilized to encode the low-dimensional popular features, and a coded mask version is fused into features of an intermediate layer of the generation model through a residual enhancement network; training the first generation model based on a gradient estimation process of score matching, and introducing a disturbance projection vector in the training process to obtain a second generation model; and by utilizing Hamiltonian dynamics simulation and combining gradient information of the model, determining a sampling path and sampling to obtain sample data meeting data distribution requirements of the private data set. According to the method, the original high-dimensional private data set is mapped to the low-dimensional space, the popular features are extracted, meanwhile, the data disturbance is further increased by using the disturbance projection vector, and the privacy protection strength is enhanced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent vehicle networking, and specifically relates to a privacy protection method and related devices for a generation model suitable for intelligent vehicle networking. Background Art

[0002] In Intelligent Vehicular Networks (IVN), vehicles are connected to each other and communicate with infrastructure through the network to achieve information sharing and collaborative control, thereby improving traffic efficiency and safety. Generative models are widely used in IVN, such as generating virtual data to improve the training effect of autonomous driving algorithms, simulating traffic scenarios to test new algorithms, and detecting abnormal behaviors. However, these models may involve a large amount of sensitive data during the training process, such as the owner's location information, driving behavior data, etc., which may lead to privacy leakage if not handled properly.

[0003] Differential Privacy is one of the mainstream technologies for protecting data privacy. It protects personal privacy by injecting noise into data or calculation results to ensure that the impact of any single data point does not significantly change the results in a statistical sense. Introducing differential privacy in the training process of generative models can effectively prevent the model from leaking sensitive information in the training data. In the field of generative models, the research on Differentially Private Generative Models (DPGMs) mainly focuses on how to generate high-quality data while ensuring privacy.

[0004] Although existing differentially private generative models (DPGMs) have made some progress in protecting privacy, they still have some significant shortcomings in practical applications, especially when dealing with high-dimensional data, such as difficulty in processing high-dimensional data, high computational complexity, and low quality of generated data. Summary of the invention

[0005] In view of this, the present invention aims to provide a privacy protection method and related device for a generative model suitable for intelligent vehicle networking, which is used to solve the above-mentioned problems existing in the existing differential privacy generative model when processing high-dimensional data.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a privacy protection method for a generation model applicable to an intelligent vehicle network, comprising the following steps:

[0008] The energy model is used to generalize the popular features of a given private dataset, and the projection technology is used to map the private dataset to a low-dimensional space to obtain low-dimensional popular features. The private dataset is a dataset with privacy protection requirements.

[0009] The low-dimensional popular features are encoded using the pre-trained first generative model, and the encoded masked version is integrated into the features of the intermediate layer of the generative model through the residual enhancement network;

[0010] The first generative model is trained based on the gradient estimation process of score matching, and a perturbation projection vector is introduced in the training process to obtain a second generative model, where the perturbation projection vector is a vector obtained by perturbing the projection vector obtained by the projection technology;

[0011] By using Hamiltonian dynamics simulation and combining the gradient information of the second generation model, the sampling path is determined and sampling is performed to obtain sample data that meets the data distribution requirements of the private dataset.

[0012] Furthermore, the expression of low-dimensional popular features obtained from the private dataset is as follows:

[0013]

[0014] In the formula, represents Fisher divergence, E(θ; x) represents energy model, θ and x represent model parameters and model input, Express expectations, represents the logarithmic gradient, p(x) represents the data distribution, U(x) represents the potential energy function, and H(θ; x) represents e E(θ;x)

[0015] Furthermore, the training loss function of the first generative model is as follows:

[0016]

[0017] In the formula, represents the loss function, v represents the mapping vector, Represents the transpose of the mapped vector.

[0018] Furthermore, when the first generative model is trained, a noise vector sampled by a Gaussian distribution is added to the features of the intermediate layer of the generative model.

[0019] Furthermore, a random perturbation mechanism is used to perturb the projection vector. The expression of the random perturbation mechanism is as follows:

[0020]

[0021] Where Pr[·] represents the probability distribution function, v iRepresents the elements in the mapped vector, v o represents a random variable, P(·) represents a random perturbation mechanism, v′ i Indicates to remove {v i}, v - is a subset of the mapping vector, and ε represents the privacy overhead parameter.

[0022] Furthermore, the training loss function of the second generative model is as follows:

[0023]

[0024] In the formula, represents the loss function, q θ (x) represents the convergence of the potential energy function U(x) on the model θ.

[0025] Furthermore, the sampling process is carried out as follows:

[0026]

[0027] Where c(·) represents the sampling function, t represents the sampling element, and λ represents the step size.

[0028] In a second aspect, the present invention provides a privacy protection device for a generation model applicable to an intelligent vehicle network, comprising:

[0029] The first data processing module is used to generalize the popular features of a given private data set by using an energy model, and map the private data set to a low-dimensional space by using a projection technique to obtain low-dimensional popular features, where the private data set is a data set with privacy protection requirements;

[0030] A residual enhancement module is used to encode the low-dimensional popular features using the pre-trained first generative model, and integrate the encoded masked version into the features of the intermediate layer of the generative model through the residual enhancement network;

[0031] A projection perturbation module is used to train the first generation model based on the gradient estimation process of score matching, and introduce a perturbation projection vector in the training process to obtain a second generation model, where the perturbation projection vector is a vector obtained by perturbing the projection vector obtained by the projection technology;

[0032] The sampling module is used to use Hamiltonian dynamics simulation, combined with the gradient information of the second generation model, to determine the sampling path and perform sampling to obtain sample data that meets the data distribution requirements of the private dataset.

[0033] Accordingly, the present invention also provides a computer device, the device comprising a processor and a memory:

[0034] The memory is used to store the computer program and send the instructions of the computer program to the processor;

[0035] The processor executes a privacy protection method for a generation model applicable to an intelligent vehicle network as described in the first aspect according to the instructions of the computer program.

[0036] Correspondingly, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, a privacy protection method for a generation model applicable to an intelligent vehicle network as described in the first aspect is implemented.

[0037] In summary, the present invention provides a privacy protection method and related device for a generative model suitable for intelligent vehicle networking, including using an energy model to generalize the popular features of a given private data set, and using a projection technique to map the private data set to a low-dimensional space to obtain low-dimensional popular features, where the private data set is a data set with privacy protection requirements; using a pre-trained first generative model to encode the low-dimensional popular features, and integrating the encoded masked version into the features of the middle layer of the generative model through a residual enhancement network; training the first generative model based on a gradient estimation process of score matching, and introducing a perturbation projection vector during the training process to obtain a second generative model, where the perturbation projection vector is a vector obtained by perturbing the projection vector obtained by the projection technology; using Hamiltonian dynamics simulation, combined with the gradient information of the second generative model, determining the sampling path and sampling, and obtaining sample data that meets the data distribution requirements of the private data set. The present invention reduces the risk of sensitive information leakage in the data by mapping the original high-dimensional private data set to a low-dimensional space and extracting popular features, and at the same time further increases data perturbation by using the perturbation projection vector, thereby enhancing privacy protection. This method is particularly suitable for intelligent Internet of Vehicles scenarios. It can generate privacy-protected synthetic data that meets the characteristics of Internet of Vehicles data. It is conducive to data analysis, model training and system testing while protecting user privacy, and promotes the security and development of Internet of Vehicles technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 A flowchart of a privacy protection method for a generation model applicable to an intelligent vehicle network provided by an embodiment of the present invention;

[0040] Figure 2A logic block diagram of a privacy protection method for a generation model applicable to an intelligent Internet of Vehicles provided in an embodiment of the present invention;

[0041] Figure 3 A flowchart of a privacy protection method for a generation model applicable to an intelligent Internet of Vehicles provided in an embodiment of the present invention;

[0042] Figure 4 A block diagram of a privacy protection device for a generation model applicable to an intelligent Internet of Vehicles provided by an embodiment of the present invention;

[0043] Figure 5 A block diagram of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] The prior art is first introduced below.

[0046] Existing differential privacy generation model DPGMs methods can be roughly divided into the following two categories:

[0047] 1. Noise injection-based method: This method achieves differential privacy by injecting noise into data or model parameters. For example, DP-GANs (Differentially Private Generative Adversarial Networks) protects privacy by adding noise to the gradient of the discriminator. However, this method does not work well when processing high-dimensional data, because high-dimensional data requires larger noise to ensure privacy, which affects the quality of generated data.

[0048] 2. Private gradient-based methods: This method protects data by implementing differential privacy during gradient calculation. For example, DP-SGD (Differentially Private Stochastic Gradient Descent) ensures privacy by adding noise every time the gradient is updated. However, due to the complexity of high-dimensional data, existing private gradient methods also face challenges when generating high-dimensional data. High-dimensional data has more features and more complex distributions, which makes it more difficult to protect privacy while maintaining the practicality of the data.

[0049] Although existing differentially private generative models (DPGMs) have made some progress in protecting privacy, they still have some significant shortcomings in practical applications, especially when dealing with high-dimensional data. These shortcomings mainly include:

[0050] Difficulty in processing high-dimensional data: Existing methods generate poor data quality when processing high-dimensional data. The noise injection method requires a large amount of noise to ensure privacy, which will significantly reduce the quality of the generated data. The private gradient method is difficult to effectively capture the characteristics of the data in the face of the complexity of high-dimensional data, and the quality of the generated samples is not high.

[0051] High computational complexity: Some differential privacy techniques have high computational complexity in their implementation, especially when dealing with high-dimensional data, where the computational cost increases significantly. This limits the feasibility and efficiency of these methods in practical applications.

[0052] The quality of generated data is low: Due to the noise introduced by the privacy protection mechanism and the limitation of model complexity, the quality of data generated by existing methods is generally low, and it is difficult to generate high-resolution and high-quality images. This is particularly evident in application scenarios that require high-quality data, such as smart car networking.

[0053] Poor balance between privacy and practicality: Existing differential privacy generation models still have deficiencies in balancing privacy protection and data practicality. In order to ensure privacy, it is often necessary to compromise on data quality, which affects the practicality of the model.

[0054] Based on this, the present invention provides a privacy protection method for a generation model applicable to an intelligent vehicle network, comprising the following steps:

[0055] S11: The energy model is used to generalize the popular features of a given private dataset, and the projection technology is used to map the private dataset to a low-dimensional space to obtain low-dimensional popular features. The private dataset is a dataset with privacy protection requirements.

[0056] It should be noted that since the data in the intelligent vehicle network is generally high-dimensional and complex, it is impractical to directly calculate the gradient of this data. Therefore, this step first uses an energy model (a function that measures the probability of a data point) to capture the most important and commonly occurring features (popular features) in the private dataset. The energy model can evaluate the energy value of a data point, and lower energy values ​​correspond to more typical patterns in the dataset.

[0057] Then, in order to further reduce the computational overhead, dimensionality reduction techniques (such as PCA, t-SNE, etc.) are applied to map the complex private data set from a high-dimensional space to a low-dimensional space. This process retains the main structural information of the data while removing or reducing irrelevant details, which helps privacy protection because low-dimensional representations usually hide some individual identification information.

[0058] S12: Use the pre-trained first generative model to encode the low-dimensional popular features, and integrate the encoded masked version into the features of the intermediate layer of the generative model through the residual enhancement network.

[0059] It should be noted that this step uses a pre-trained generative model (such as GANs, VAEs, etc.) to encode low-dimensional popular features. This step converts data features into a form that can be understood by the generative model.

[0060] The encoded features are processed in some way (such as adding masks to further protect privacy), and then integrated into the intermediate layer features of the generative model through the residual enhancement network (a network structure that can effectively learn residuals). This can improve the expressiveness of the model and the quality of the generated data.

[0061] S13: The first generation model is trained based on the gradient estimation process of score matching, and a perturbation projection vector is introduced in the training process to obtain a second generation model, where the perturbation projection vector is a vector obtained by perturbing the projection vector obtained by the projection technology.

[0062] It should be noted that this step introduces a gradient estimation method based on score matching to optimize the first generative model. Score matching technology is a method for estimating the gradient of a probability density function when the probability density function is unknown. It is suitable for training generative models and can capture data distribution more accurately.

[0063] During the training process, random perturbations are introduced into the projection vectors obtained by the projection technique to generate "perturbed projection vectors". This perturbation increases the randomness of model training, helps improve the generalization ability of the model, and further confuses the characteristics of the original data to enhance privacy protection.

[0064] S14: Using Hamiltonian dynamics simulation and combining the gradient information of the second generation model, determine the sampling path and perform sampling to obtain sample data that meets the data distribution requirements of the private dataset.

[0065] It should be noted that this step uses the Hamiltonian dynamics system theory, combined with the gradient information of the second generative model (i.e., the model obtained through the above training), to simulate the dynamic behavior of the physical system. This step allows exploring the data distribution space in an efficient and natural way.

[0066] Then, samples are extracted from the second generation model by simulating the determined sampling path. These sample data not only conform to the overall statistical characteristics of the private dataset, but also provide rich diversity while maintaining privacy protection, which is suitable for various applications and analysis in the Internet of Vehicles environment.

[0067] This embodiment provides a privacy protection method for a generative model suitable for intelligent Internet of Vehicles. By mapping the original high-dimensional private data set to a low-dimensional space and extracting popular features, the risk of sensitive information leakage in the data is reduced. At the same time, the perturbation projection vector is used to further increase the data perturbation, thereby enhancing the privacy protection. This method is particularly suitable for intelligent Internet of Vehicles scenarios, and can generate privacy-protected synthetic data that meets the characteristics of Internet of Vehicles data. It is conducive to data analysis, model training, and system testing under the premise of protecting user privacy, and promotes the security and development of Internet of Vehicles technology.

[0068] In some embodiments, for step S11, an energy model is used to generalize the manifold features of high-dimensional data (such as a private data set of an intelligent vehicle network), as follows:

[0069]

[0070] in, represents Fisher divergence, E(θ; x) represents energy model, θ and x represent model parameters and model input, Express expectations, represents the logarithmic gradient, p(x) represents the data distribution, U(x) represents the potential energy function, and H(θ; x) represents e E(θ;x) .

[0071] In some embodiments, for step S11, the dimensionality reduction operation is specifically to sample the projection vector v from a standard Gaussian distribution. Then, the projected gradients U(x) and H(θ; x) are in the direction of the projection vector v. Finally, the loss function is calculated as follows:

[0072]

[0073] In the formula, represents the loss function, v represents the mapping vector, Represents the transpose of the mapping vector

[0074] In some embodiments, for step S12, although the private gradient and low-dimensional manifold features of high-dimensional data are obtained in step S11, the reconstruction ability of the model still needs to be further enhanced. The private gradient of high-dimensional data is to ensure the accuracy and privacy of gradient information as much as possible, while the low-dimensional manifold features are to better train and reconstruct data. Therefore, for the low-dimensional feature data obtained in step S11, this embodiment uses a pre-trained generative model such as VQGAN to generate data based on the idea of ​​residual network. (i.e. the data obtained after processing in step S11) is encoded and the encoded masked version is integrated into the model q θ In the features of the middle layer. Through this residual learning of different versions of data, the reconstruction ability of the model will be improved.

[0075] In addition, in some other embodiments, in order to further improve the robustness of model learning, a noise vector sampled from a Gaussian distribution is added to the feature instead of being added to the training model q θ The noise vector is added to the encoded masked features. Compared with adding nothing, adding a noise vector improves the robustness of the model, especially when the data is located in low-density areas. Each residual block consists of the following layers:

[0076] 1. Convolutional layer: extract features.

[0077] 2. Batch Normalization: Stabilizes the training process.

[0078] 3. Activation function (ReLU): introduces nonlinearity.

[0079] 4. Skip connection: add the input directly to the output.

[0080] The specific steps are as follows:

[0081] 1. Convolve the input features: y = Conv(x)

[0082] 2. Perform batch normalization: y = BatchNorm(y)

[0083] 3. Apply activation function: y = ReLU(y)

[0084] 4. Add the input to the output via a skip connection: y = y + x

[0085] By superimposing multiple residual blocks, the reconstruction ability of the model can be effectively enhanced.

[0086] The above embodiment applies the residual learning mechanism on the low-dimensional manifold features, enhances the model through the residual block structure (convolution, batch normalization, activation function and jump connection), and directly adds Gaussian noise to the features instead of operating on the mask features to improve the robustness and generalization ability of the model.

[0087] In order to ensure privacy, existing methods often need to inject large noise, resulting in low quality of generated data, especially in the case of high-dimensional data, where the clarity and resolution of the image are greatly affected. This embodiment introduces a residual enhancement module in the slice score matching process, which significantly improves the reconstruction capability and quality of the generated image by capturing the high-order features of the data. At the same time, by using the efficient sampling of Hamiltonian dynamics, the generated image is closer to the real data distribution.

[0088] In some embodiments, the network q is trained directly using the gradient of the logarithmic data density θ This may introduce privacy risks. To address this issue, this embodiment performs random perturbations to prevent private information leakage. It perturbs the training data so that the trained network does not reveal the true distribution of private data during sampling. Therefore, the trained network can be released without worrying about privacy loss, because it is difficult for the adversary to determine whether the image is in the training data. Specifically, the goal of this embodiment is to privatize the gradient in the equation This is equivalent to privatization. For a batch of data, and its projection vector There is an inherent correspondence between them, so in the absence of any protection, The projection vector v i and In this case, the present embodiment applies a random perturbation mechanism P(·), which privatizes The projection vector of , which is achieved by making the projection vector U(x) and H(θ;x) not strictly aligned. This random perturbation makes the private It indirectly randomizes the real direction towards the perceptually real image. In this way, even if the attacker can obtain the gradient information, it cannot be reused because the added random noise makes it impossible for the attacker to distinguish between real data and perturbed data. In this embodiment, the random perturbation mechanism P(·) can be formulated as follows:

[0089]

[0090] Among them, v - is a subset of v, k = v-, k ≥ 2, v ois a random variable, and ε is a privacy overhead parameter. In order to achieve differential privacy, a perturbed projection vector is introduced in the gradient estimation process of slice score matching. Based on this, in the specific implementation, the generation model training steps are as follows:

[0091] 1. For each data sample x, use step S11 to project the gradients of U(x) and H(θ; x) to the direction of the projection vector v;

[0092] 2. Perform residual enhancement on the obtained low-dimensional manifold features using step S12. During this process, the noise vector sampled by Gaussian distribution needs to be added to the features instead of being added to the training model q θ In the masked features of the encoding;

[0093] 3. To further protect q in step 2 θ The network parameters and features of the model use the differential privacy mechanism to perturb the projection vector, which is equivalent to the perturbation vector v;

[0094] 4. Use Equation 3 to project the gradient.

[0095] Through the above steps, the statistical characteristics of the data can be well preserved under the premise of ensuring differential privacy. In this embodiment, the training loss function of the entire model can be defined as follows:

[0096]

[0097] In the formula, represents the loss function, q θ (x) represents the convergence of the potential energy function U(x) on the model θ.

[0098] In view of the fact that existing methods in the prior art have deficiencies in balancing privacy protection and data practicality, and often need to compromise on data quality, this embodiment introduces a perturbed projection vector in the gradient estimation process in slice score matching, and the present invention implements differential privacy protection. In this process, the statistical characteristics of the data are better preserved, and the balance between privacy protection and data practicality is optimized, so that the generated data has both high quality and effective privacy protection.

[0099] In some embodiments, Hamiltonian dynamics can be used to gradually approximate the manifold of a private data set using the gradient estimated by a trained model during the sampling process, thereby generating high-quality images. Hamiltonian dynamics is a mathematical tool for simulating the evolution of physical systems. Hamiltonian dynamics uses an energy function (HamiltonianFunction) to describe the state changes of the system. For the generative model, an energy function H(x,p) can be defined, where x is a data sample and p is an auxiliary variable (momentum). The evolution process of Hamiltonian dynamics can be described by the following set of differential equations:

[0100]

[0101] By solving the above differential equations, a sampling path can be obtained so that the generated samples gradually approach the target data distribution. First, the present invention uses a well-trained network to estimate the gradient of the private data set and defines the Hamiltonian function:

[0102] H(x,p)=U(x)+K(p)(6)

[0103] Among them, U(x) represents the potential energy of data sample x, and K(p) represents the kinetic energy of auxiliary variable p. Then, the Leapfrog algorithm is used for numerical solution. The specific steps are as follows:

[0104] 1. Initialize the data sample x and auxiliary variable p, that is

[0105] p(x,c)=p(x)·p(c)∝exp(-H(θ;x))·exp(-K(c)).(7)

[0106] 2. Sampling is carried out according to the following method:

[0107]

[0108] 3. Where λ is the step size. According to the initialization method, we can know: as well as Therefore, the above sampling process can be simplified as:

[0109]

[0110] Where c(·) represents the sampling function and t represents the sampling element.

[0111] By iterating the above process multiple times, samples close to the target data distribution can be gradually obtained. This mechanism ensures that the entire distribution of private data is thoroughly explored, expressed as p(x), which helps to generate images with enhanced authenticity. Inspired by the learning rate adjustment technique in machine learning, the present invention adopts a step size decay strategy to speed up the sampling process. Initially, a larger step size is used to accelerate the movement towards the real image. Subsequently, a smaller step size is used to refine the image details. As an additional benefit, because the starting step size is larger, it also prevents the model from collapsing into certain data-intensive areas to a certain extent. It is worth noting that during the sampling process, the present invention no longer uses VQGAN to encode the input of the model, but directly samples from the Gaussian distribution to add to the features extracted by the model. On the one hand, it speeds up the sampling process compared to encoding with VQGAN, and on the other hand, it improves the robustness of the model, especially in increasing the diversity of generated data.

[0112] In view of the problem that the traditional differential privacy generation model in the prior art performs poorly when processing high-dimensional data and the quality of the generated data cannot meet the needs of actual applications, this embodiment introduces Hamiltonian dynamics and performs efficient sampling through the gradient estimated by the trained model, so that the generated samples can gradually approach the manifold of the private data set. This mechanism significantly improves the processing capability of high-dimensional data and can generate high-quality images.

[0113] See also Figure 2 and 3 In combination with the above embodiments, an application example of a privacy protection method for a generation model suitable for intelligent vehicle networking is proposed. Figure 2 and 3 They are respectively the logic block diagram and process implementation block diagram of the application example. Among them, this application example adopts three sub-algorithms, among which sub-algorithm 1 adopts Hamiltonian dynamics sampling to efficiently perform sample sampling to reduce data dimension and computational complexity; sub-algorithm 2 adopts residual enhancement module to capture manifold features in data, thereby achieving high-quality and high-resolution image generation; sub-algorithm 3 adopts a perturbation projection vector method based on differential privacy to protect the privacy of the generated model and the data privacy security of the intelligent vehicle network. Based on this application example, this method has the following characteristics in specific application:

[0114] 1. Hamiltonian dynamics sampling: Hamiltonian dynamics sampling is performed through the gradient estimated by the trained model, so that the generated samples gradually approach the manifold of the private dataset, significantly improving the processing capability of high-dimensional data and the quality of generated data.

[0115] 2. Perturbed projection vector in slice score matching: Introducing perturbed projection vector in the gradient estimation process of slice score matching to achieve differential privacy protection while retaining the statistical characteristics of the data and optimizing the balance between privacy and practicality.

[0116] 3. Residual enhancement module: The residual enhancement module is introduced in the score matching process to significantly improve the reconstruction ability and quality of the generated image by capturing the high-order features of the data.

[0117] 4. Private gradient estimation: Introduce noise in the gradient estimation process to ensure that sensitive information is not leaked during model training and achieve efficient differential privacy protection.

[0118] Through the above innovations, this method generates high-quality high-dimensional data while ensuring data privacy, which is particularly suitable for application fields such as smart vehicle networks that require privacy protection.

[0119] Based on the same inventive concept, the embodiment of the present application also provides a generation model privacy protection device applicable to the intelligent vehicle network for implementing the generation model privacy protection method applicable to the intelligent vehicle network involved above. The implementation scheme for solving the problem provided by the system is similar to the implementation scheme recorded in the above method, so the specific limitations in the embodiment of the generation model privacy protection device applicable to the intelligent vehicle network provided below can be referred to the limitations of the generation model privacy protection method applicable to the intelligent vehicle network above, and will not be repeated here.

[0120] See also Figure 4 , this embodiment provides a privacy protection device for a generation model suitable for an intelligent vehicle network, including:

[0121] The first data processing module is used to generalize the popular features of a given private data set by using an energy model, and map the private data set to a low-dimensional space by using a projection technique to obtain low-dimensional popular features, where the private data set is a data set with privacy protection requirements;

[0122] A residual enhancement module is used to encode the low-dimensional popular features using the pre-trained first generative model, and integrate the encoded masked version into the features of the intermediate layer of the generative model through the residual enhancement network;

[0123] A projection perturbation module is used to train the first generation model based on the gradient estimation process of score matching, and introduce a perturbation projection vector in the training process to obtain a second generation model, where the perturbation projection vector is a vector obtained by perturbing the projection vector obtained by the projection technology;

[0124] The sampling module is used to use Hamiltonian dynamics simulation, combined with the gradient information of the second generation model, to determine the sampling path and perform sampling to obtain sample data that meets the data distribution requirements of the private dataset.

[0125] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0126] Reference Figure 5 The embodiment of the present invention further provides a computer device 1, comprising: a memory 12 and a processor 11 and a computer program 13 stored in the memory 12. When the computer program 13 is executed on the processor 11, a method as described in any one of the above methods is implemented.

[0127] The computer device 1 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The computer device 1 may include, but is not limited to, a processor 11 and a memory 12. Those skilled in the art will appreciate that Figure 5 It is only an example of the computer device 1 and does not constitute a limitation on the computer device 1. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, it may also include input and output devices, network access devices, etc.

[0128] The processor 11 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0129] In some embodiments, the memory 12 may be an internal storage unit of the computer device 1, such as a hard disk or memory of the computer device 1. In other embodiments, the memory 12 may also be an external storage device of the computer device 1, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the computer device 1. Further, the memory 12 may also include both an internal storage unit of the computer device 1 and an external storage device. The memory 12 is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program, etc. The memory 12 may also be used to temporarily store data that has been output or is to be output.

[0130] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the method described in any one of the above methods is implemented.

[0131] In this embodiment, if the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium can at least include: any entity or device that can carry the computer program code to the camera / terminal device, recording medium, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. For example, a USB flash drive, a mobile hard disk, a disk or an optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.

[0132] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0133] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0134] In the embodiments disclosed in the present application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are only schematic. For example, the division of the modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0135] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A privacy protection method for a generative model suitable for intelligent vehicle networking, characterized in that: The steps include: The energy model is used to generalize the popular features of a given private data set, and the projection technology is used to map the private data set to a low-dimensional space to obtain low-dimensional popular features, wherein the private data set is a data set with privacy protection requirements; Encoding the low-dimensional popular features using a pre-trained first generative model, and integrating the encoded masked version into the features of the intermediate layer of the generative model through a residual enhancement network; The first generative model is trained based on a gradient estimation process of score matching, and a perturbation projection vector is introduced during the training process to obtain a second generative model, where the perturbation projection vector is a vector obtained by perturbing a projection vector obtained by the projection technique; By utilizing Hamiltonian dynamics simulation and combining the gradient information of the second generation model, a sampling path is determined and sampling is performed to obtain sample data that meets the data distribution requirements of the private data set.

2. The privacy protection method for the generation model applicable to the intelligent vehicle network according to claim 1 is characterized in that: The expression of the low-dimensional popular feature obtained from the private dataset is as follows: In the formula, represents Fisher divergence, E(θ; x) represents energy model, θ and x represent model parameters and model input, Express expectations, represents the logarithmic gradient, p(x) represents the data distribution, U(x) represents the potential energy function, and H(θ; x) represents e E(θ;x) .

3. The privacy protection method for the generation model applicable to the intelligent vehicle network according to claim 2 is characterized in that: The training loss function of the first generative model is as follows: In the formula, represents the loss function, v represents the mapping vector, v T Represents the transpose of the mapped vector.

4. The privacy protection method for the generation model applicable to the intelligent vehicle network according to claim 1 is characterized in that: When the first generation model is trained, a noise vector sampled by Gaussian distribution is also added to the features of the middle layer of the generation model.

5. The privacy protection method for the generation model applicable to the intelligent vehicle network according to claim 3 is characterized in that: The projection vector is perturbed using a random perturbation mechanism, and the expression of the random perturbation mechanism is as follows: Where Pr[·] represents the probability distribution function, v i Represents the elements in the mapped vector, v o represents a random variable, P(·) represents a random perturbation mechanism, v' i Indicates to remove {v i }, v - is a subset of the mapping vector, and ε represents the privacy overhead parameter.

6. The privacy protection method for the generation model applicable to the intelligent vehicle network according to claim 5 is characterized in that: The training loss function of the second generative model is as follows: In the formula, represents the loss function, q θ (x) represents the convergence of the potential energy function U(x) on the model θ.

7. The privacy protection method for the generation model applicable to the intelligent vehicle network according to claim 1 is characterized in that: The sampling process is carried out as follows: Where c(·) represents the sampling function, t represents the sampling element, and λ represents the step size.

8. A privacy protection device for a generative model suitable for intelligent vehicle networking, characterized in that: include: A first data processing module is used to generalize the popular features of a given private data set by using an energy model, and map the private data set to a low-dimensional space by using a projection technique to obtain low-dimensional popular features, wherein the private data set is a data set with privacy protection requirements; A residual enhancement module, configured to encode the low-dimensional popular features using a pre-trained first generative model, and integrate the encoded masked version into the features of the intermediate layer of the generative model through a residual enhancement network; A projection perturbation module, used for training the first generation model based on the gradient estimation process of score matching, and introducing a perturbation projection vector in the training process to obtain a second generation model, wherein the perturbation projection vector is a vector obtained by perturbing the projection vector obtained by the projection technology; The sampling module is used to use Hamiltonian dynamics simulation, combined with the gradient information of the second generation model, to determine the sampling path and perform sampling to obtain sample data that meets the data distribution requirements of the private data set.

9. A computer device, characterized in that: The device comprises a processor and a memory: The memory is used to store a computer program and send instructions of the computer program to the processor; The processor executes a privacy protection method for a generation model applicable to an intelligent vehicle network as described in any one of claims 1 to 7 according to the instructions of the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, it implements a privacy protection method for a generation model applicable to an intelligent vehicle network as described in any one of claims 1 to 7.

Citation Information

Patent Citations

  • Intelligent electric meter fault diagnosis method and device based on federated learning

    CN111537945A

  • Question and answer system model training method and device based on differential privacy technology

    CN114547687A

  • Generative adversarial network-based Internet of Vehicles intrusion detection model training method, detection method and equipment

    CN116232656A

  • Privacy data generation method based on pre-training diffusion model

    CN117852088A

  • Large language model enhanced question and answer generation method

    CN118013051A