Self-adaptive aggregation privacy protection method and device, electronic equipment and storage medium

Abstract

The invention provides an adaptive aggregation privacy protection method, which is applied to a vehicle client of a user, and comprises the following steps: carrying out privacy budget allocation on trajectory data of a vehicle of the user to obtain a privacy budget corresponding to the user; according to the privacy budget corresponding to the user, performing Laplace noise addition on the trajectory data to obtain noise addition trajectory data of the vehicle of the user; and sending the noise adding trajectory data to a road side unit (RSU). By designing a personalized privacy budget allocation scheme, customized privacy protection measures are provided for the user, and better use experience is provided for the user while the privacy protection requirement of the user is met.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to an adaptive aggregation privacy protection method, device, electronic device and storage medium. Background Art

[0002] With the rapid development of Internet of Things (IoT) and Artificial Intelligence (AI) technologies and the widespread application of location-based services (LBS), the collection and use of user location information has increased. How to better protect user privacy and security has become an urgent issue to be solved. LBS cloud servers obtain user location information to provide users with real-time navigation, route planning and other functions, and collect user location information to optimize user experience and improve efficiency. At the same time, when processing user location information, LBS service providers may intentionally or unintentionally cause user information leakage. In existing user location privacy protection schemes, privacy protection measures that combine encryption, pseudonymization, anonymization and other technical means are often used to protect user location information. However, due to the use of unified privacy protection measures, it may lead to over-protection or under-protection of the location privacy of some users, which fails to meet the personalized needs of users and thus affects the user experience. Summary of the invention

[0003] 1. Technical issues to be resolved The present invention provides an adaptive aggregation privacy protection method to solve the technical problem that existing privacy protection measures cannot meet the personalized needs of users. By designing a personalized privacy budget allocation scheme, customized privacy protection measures are provided for users, thereby providing users with a better user experience while meeting their privacy protection needs.

[0004] (II) Technical solution In a first aspect, the present invention provides an adaptively aggregated privacy protection method, which is applied to a vehicle client of a user, and the method comprises: Allocating a privacy budget to the trajectory data of the vehicle of the user to obtain a privacy budget corresponding to the user; According to the privacy budget corresponding to the user, Laplace noise is performed on the trajectory data to obtain noisy trajectory data of the user's vehicle; and The noisy trajectory data is sent to the roadside unit RSU, the noisy trajectory data is used for user clustering, and is used as a training sample for the local model corresponding to the vehicle client to which the user belongs, the local model corresponding to the vehicle client to which the user belongs is used to obtain the inter-cluster local model corresponding to the cluster to which the user belongs, and the inter-cluster local model corresponding to the cluster to which the user belongs is used to obtain the global model, wherein the inter-cluster local model is obtained by model aggregation of the inter-cluster local models corresponding to all clusters respectively, the global model is obtained by model aggregation of the inter-cluster local models corresponding to all clusters respectively, and the global model is used to provide LBS services for users.

[0005] Furthermore, the allocating a privacy budget to the trajectory data of the vehicle of the user includes: Determine polar coordinates corresponding to a plurality of location point data in the trajectory data, and analyze the polar coordinates corresponding to the plurality of location point data by a probability density function to obtain a distribution of location points corresponding to the trajectory data; Determining a sensitive area of ​​the user based on the sensitive position of the user and the radius of the sensitive area; and A privacy budget corresponding to the user is determined according to a relationship between the distribution of the location points and the sensitive area.

[0006] Furthermore, the radius of the sensitive area It is expressed as: ; in, represents the sum of the sensitive distances respectively corresponding to the LBS positions of at least one service request point of the user in the trajectory corresponding to the trajectory data, where the sensitive distance corresponding to the LBS position is the distance between the LBS position and the sensitive position corresponding to the LBS position, is the Lambert w function, with parameters Used to normalize the radius of the sensitive area, is the distance between the LBS position of the user without noise and the LBS position with noise that the user can accept, that is, equivalent to the error that the user can tolerate. is the privacy budget corresponding to the user.

[0007] Furthermore, the privacy budget includes the privacy budget of the user outside the sensitive area. and the user's privacy budget in sensitive areas ,in: ; in, Indicates the user The LBS location The sensitive position corresponding to the LBS position The sensitive distance between represents the number of at least one service request point, represents the distance between the position point of the user in the trajectory corresponding to the trajectory data and the sensitive position, represents the radius of the sensitive area, A weighted function representing the Euclidean distance between the position point of the user in the trajectory corresponding to the trajectory data and the sensitive position is expressed as follows: .

[0008] Furthermore, the scale of the noise corresponding to the Laplace noising is determined by the following expression: ; in, represents the sensitivity, which means that for the query function For the data set The maximum change in the output result is represents the privacy budget of the user, Indicates the scale of the noise.

[0009] In a second aspect, the present invention provides an adaptive aggregation privacy protection method, which is applied to RSU, and the method includes: Acquire noisy trajectory data of vehicles of multiple users from vehicle clients of the multiple users, wherein the noisy trajectory data is obtained by performing Laplace noise on the trajectory data of the vehicles of the users by the corresponding vehicle clients according to the privacy budgets corresponding to the users, and the privacy budgets corresponding to the users are obtained by performing privacy budget allocation on the trajectory data of the vehicles of the users by the corresponding vehicle clients; Clustering the multiple users based on the noisy trajectory data of the vehicles of the multiple users to obtain multiple clusters; Aggregating the local models uploaded by all vehicle clients in the corresponding cluster to obtain inter-cluster local models corresponding to all clusters, and sending the inter-cluster local models to the vehicle clients in the corresponding cluster for training to update the inter-cluster local models; and The inter-cluster local models corresponding to all the clusters are sent to the cloud server, and the inter-cluster local models corresponding to all the clusters are used for model aggregation to obtain a global model, and the global model is used to provide LBS services for users.

[0010] In a third aspect, the present invention provides an adaptive aggregation privacy protection method, which is applied to a cloud server, and the method includes: Sending an initial global model, where the initial global model is used for each user's vehicle client to train and obtain a local model; Receive local models corresponding to all clusters sent by the RSU, wherein the local models corresponding to all clusters are trained by the RSU based on the noisy trajectory data of the vehicles of the users in each cluster, the users in each cluster are determined by the RSU based on the clustering of the noisy trajectory data of the vehicles of all users, the noisy trajectory data are obtained by the corresponding vehicle client performing Laplace noise on the trajectory data of the vehicle of the user according to the privacy budget corresponding to the user, and the privacy budget corresponding to the user is obtained by the corresponding vehicle client performing privacy budget allocation on the trajectory data of the vehicle of the user; and The local models corresponding to all clusters are aggregated to obtain a global model, and the global model is used to provide LBS services for users.

[0011] In a fourth aspect, the present invention provides an adaptive aggregation privacy protection device, which is applied to a user's vehicle client, comprising: A privacy budget allocation unit, configured to allocate a privacy budget to the trajectory data of the vehicle of the user to obtain a privacy budget corresponding to the user; a trajectory noisy unit, configured to perform Laplace noise on the trajectory data according to a privacy budget corresponding to the user, to obtain noisy trajectory data of the user's vehicle; and A sending unit is used to send the noisy trajectory data to a roadside unit RSU, the noisy trajectory data is used for clustering users and used as a training sample for a local model corresponding to the vehicle client to which the user belongs, the local model corresponding to the vehicle client to which the user belongs is used to obtain an inter-cluster local model corresponding to the cluster to which the user belongs, the inter-cluster local model corresponding to the cluster to which the user belongs is used to obtain a global model, wherein the inter-cluster local model is obtained by model aggregation of inter-cluster local models corresponding to all clusters respectively, the global model is obtained by model aggregation of inter-cluster local models corresponding to all clusters respectively, and the global model is used to provide LBS services for users.

[0012] In a fifth aspect, the present invention provides an electronic device, applied to a user's vehicle client, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the adaptive aggregation privacy protection method as described in the first aspect is implemented.

[0013] In a sixth aspect, the present invention provides a computer-readable storage medium, which is applied to a user's vehicle client and has a computer program stored thereon. When the computer program is executed by a processor, the privacy protection method of adaptive aggregation as described in the first aspect is implemented.

[0014] (III) Beneficial effects The privacy protection method of adaptive aggregation provided by the present invention provides customized privacy protection measures for users by designing a personalized privacy budget allocation scheme, thereby providing users with a better user experience while meeting their privacy protection needs. In addition, by designing a hierarchical federated learning algorithm, the data heterogeneity problem caused by the differences in users' privacy protection needs is effectively handled, communication resources are saved, and upload time is shortened. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are 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.

[0016] Figure 1 It is a flowchart of the privacy protection method of adaptive aggregation provided by the present invention; Figure 2 It is a flowchart of the hierarchical federated learning method provided by the present invention; Figure 3 is a schematic diagram of an adaptive aggregation privacy protection device provided by the present invention; Figure 4 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are 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.

[0018] At present, with the rapid development of Internet of Things (IoT) and artificial intelligence (AI) technologies and the widespread application of location-based services (LBS), the collection and use of user location information is increasing, which directly challenges the privacy security and data sensitivity of users. LBS cloud servers obtain location information through the vehicle's location module and provide vehicle users with real-time navigation and route planning functions, thereby optimizing driving experience and efficiency. However, when processing user location information, LBS service providers may intentionally or unintentionally cause information leakage. Traditional location privacy protection schemes often use technical means such as encryption, pseudonymization, and anonymization to achieve the purpose of protecting user location information by adopting privacy protection measures that combine a variety of different privacy protection technical means. However, this location privacy protection method has obvious disadvantages: a unified privacy protection scheme may lead to over-protection or under-protection of user location privacy, which will affect the user's service experience and make it difficult to fully meet the user's personalized needs.

[0019] In addition, federated learning (FL) currently provides a distributed learning paradigm. In traditional centralized learning, cloud servers obtain and store raw data from multiple user clients. In contrast, FL learning allows user clients to keep data on local devices and only share intermediate model parameters, thereby effectively protecting the data privacy of user clients. However, the user client training data in FL learning is usually non-independent and identically distributed (Non-IID). Studies have shown that Non-IID data may lead to a decrease in FL learning accuracy compared to independent and identically distributed (IID) data. In many existing studies, FL learning models usually rely on cloud servers for global model aggregation, and user clients upload locally trained model parameters to cloud servers for parameter aggregation. However, the heterogeneity of clients leads to excessively large uploaded parameters, which consumes a lot of communication resources and significantly prolongs the upload time, especially when training complex models in multiple rounds. This problem is particularly prominent.

[0020] In order to solve the above technical problems, the present invention provides a privacy protection method for adaptively aggregating heterogeneous data in a connected vehicle environment, so as to achieve a balance between the user's personalized needs for privacy protection and service quality, efficiently integrate and utilize heterogeneous data on the basis of protecting the user's trajectory privacy, provide users with personalized models, and improve user experience and service quality.

[0021] Specifically, the present invention designs a personalized location privacy budget allocation scheme for adaptive privacy protection of vehicle trajectories, evaluates the aggregation accuracy of vehicle trajectories, and balances the relationship between privacy and data availability. At the same time, a hierarchical federated learning algorithm is designed to effectively deal with the data heterogeneity problem caused by differences in users' privacy protection requirements.

[0022] Figure 1 FIG. 1 is a flow chart of a privacy protection method for adaptive aggregation according to an embodiment of the present invention. Figure 1 As shown, the method includes: Step 101: Allocate a privacy budget for the trajectory data of the user's vehicle to obtain a privacy budget corresponding to the user; Step 102: performing Laplace noise addition on the trajectory data according to the privacy budget corresponding to the user to obtain the noisy trajectory data of the user's vehicle; Step 103: Send the noisy trajectory data to the roadside unit RSU, the noisy trajectory data is used for user clustering and used as a training sample for the local model corresponding to the vehicle client to which the user belongs, the local model corresponding to the vehicle client to which the user belongs is used to obtain the inter-cluster local model corresponding to the cluster to which the user belongs, the inter-cluster local model corresponding to the cluster to which the user belongs is used to obtain the global model, wherein the inter-cluster local model is obtained by model aggregation of the inter-cluster local models corresponding to all clusters respectively, the global model is obtained by model aggregation of the inter-cluster local models corresponding to all clusters respectively, and the global model is used to provide LBS services for users.

[0023] Specifically, in step 101, the vehicle client establishes a privacy allocation budget plan based on user needs, designs a differential privacy model and analyzes the probability distribution of random variable values, thereby allocating a privacy budget to the trajectory data of the user's vehicle to obtain the privacy budget corresponding to the user. On the one hand, it can protect data privacy. On the other hand, due to different user needs, the added noise is uneven. By analyzing the probability distribution of random variable values, the noise data near the probability distribution is determined, and the noise with large deviations is excluded. The noise data set that conforms to the probability distribution is determined for subsequent training.

[0024] The trajectory data of the user's vehicle includes multiple location data of the user's vehicle in a preset time series. It should be noted that, considering that different users have different sensitivities to different places and different privacy requirements, differential privacy protection is provided for the user in combination with the user's sensitive location, and then the trajectory data of the user's vehicle is generalized according to the Laplace noise in step 102 to improve privacy security.

[0025] The differential privacy model is set up as follows. In order to offset the influence of individual data records on the query results and ensure the privacy of the output, noise is added to any query function from the Laplace distribution. In the above example, the noise scale is given by Based on this, the following algorithm can be defined , for the dataset Perform the following processing: , in, represents the sensitivity, that is, for the query function , for the dataset , the maximum change in the output result is defined as follows: in, represents the first standardized distance, and Represents two adjacent data sets with only one piece of data different between them. Differential privacy, based on which a personalized location privacy budget is allocated for user privacy protection needs.

[0026] In addition, in order to describe the probability distribution of the values ​​of continuous random variables, the Laplace probability density function is defined as: in represents the location of the distribution center, represents the mean of the Laplace distribution, is the scale parameter that determines the dispersion of the distribution.

[0027] Furthermore, a privacy budget is allocated to the trajectory data of the vehicle of the user, including: Determine polar coordinates corresponding to a plurality of location point data in the trajectory data, and analyze the polar coordinates corresponding to the plurality of location point data by a probability density function to obtain a distribution of location points corresponding to the trajectory data; Determining a sensitive area of ​​the user based on the sensitive position of the user and the radius of the sensitive area; and A privacy budget corresponding to the user is determined according to a relationship between the distribution of the location points and the sensitive area.

[0028] Specifically, based on geographic indistinguishability, the polar coordinates of multiple location point data in the trajectory data of the user's vehicle are determined, and the edge distribution of specific location points is analyzed through the probability density function, so as to analyze the distance between the user's unnoised LBS position and the noisy LBS position that the user can accept, which is equivalent to the error that the user can tolerate, that is, the value customized by the user according to different privacy protection requirements. Specifically, according to different user privacy protection requirements, the radius of the privacy position is set. The larger the radius, the higher the inaccuracy of the service obtained by the user. The size of the radius determines the quality of service obtained by the user.

[0029] Considering the location privacy security, we first add noise to the location point data of the user’s vehicle in two-dimensional space. In time The current position is defined as , the position information after the above Laplace noise addition is ,in represents the offset radius, is the offset angle. The error distance between the two is defined as . , is the location information after the above Laplace noise addition. According to geographic indistinguishability, the actual location Converted to polar coordinates, the generated position with added noise Can be represented as a point . Define the parameters , used to normalize the radius of the sensitive area, where is a standard parameter factor, then the resulting probability density function is: According to the probability density function, we can see that this is actually the product of two marginal distributions. The two random variables are represented by (radius) and (angle) represents different possibilities of displacement, and the corresponding marginal distribution is: Therefore, it can be seen that ,and ,for Consider the cumulative distribution function: according to Defining random numbers is generated uniformly on the interval [0,1), so , which is expressed as follows: is the Lambert w function, is the distance between the user's LBS position without noise and the LBS position with noise that the user can accept, that is, the error that the user can tolerate. In order for the generated location point data after noise to meet the user's needs, it must meet , from which we can deduce: In addition, users The LBS location of the service request point on the trajectory path is defined as , for example, hospitals, shopping malls, cinemas, etc., because users The trajectory location information sent when requesting a service may contain sensitive location information, such as hospitals, etc. Define User All sensitive positions in the trajectory are . It follows that: Therefore, the users on the entire trajectory path The service request points are comprehensively evaluated and the radius R of the sensitive area is derived on the basis of satisfying the above formula.

[0030] in, represents the sum of the sensitive distances respectively corresponding to the LBS positions of at least one service request point of the user in the trajectory corresponding to the trajectory data, where the sensitive distance corresponding to the LBS position is the distance between the LBS position and the sensitive position corresponding to the LBS position, is the Lambert w function, with parameters Used to normalize the radius of the sensitive area, is the distance between the LBS position of the user without noise and the LBS position with noise that the user can accept, that is, equivalent to the error that the user can tolerate. is the privacy budget corresponding to the user.

[0031] Optionally, It may represent the sum of the sensitive distances corresponding to the LBS positions of all service request points of the user in the trajectory corresponding to the trajectory data.

[0032] Optionally, the sensitive position corresponding to the LBS position may be a sensitive position that is closest to the LBS position.

[0033] Optionally, the sensitive position corresponding to the LBS position may also be all or part of the sensitive positions whose distance to the LBS position is less than a specific distance threshold. In this case, if there are multiple sensitive positions corresponding to the LBS position, the average value of the distances between the LBS position and the multiple corresponding sensitive positions may be determined first, and then the average value may be used as the sensitive distance corresponding to the LBS position.

[0034] As mentioned above, the radius of the sensitive area can be deduced It is expressed as: ; Based on the sensitive position set by the user and the radius of the above sensitive area, the sensitive area of ​​the user is determined.

[0035] Next, the personalized location privacy budget corresponding to the user is determined according to the relationship between the distribution of the location points and the sensitive area.

[0036] Specifically, a personalized location privacy budget allocation scheme (PLPB) is proposed, in which the location of the sensitive area is determined according to the radius of the sensitive area. After determining the sensitive areas, based on the user The distance between the position point in the trajectory corresponding to the trajectory data and the sensitive position , adaptively adjust the allocation of privacy budget, thereby providing users with personalized location privacy budget allocation, and then providing users with customized privacy protection measures.

[0037] in, Indicates that the user is assigned The privacy budget outside the sensitive area, where the risk is lower, is mainly determined by the Euclidean distance. Indicates that the user is assigned The privacy budget for locations within sensitive areas, which are high risk, is the total budget minus the budget allocated to all locations outside the sensitive area, which is evenly distributed.

[0038] in, Indicates the user The LBS location The sensitive position corresponding to the LBS position The sensitive distance between represents the number of at least one service request point, Indicates the user the distance between the position point in the trajectory corresponding to the trajectory data and the sensitive position, represents the radius of the sensitive area, Indicates the user The weighted function of the Euclidean distance between the position point in the trajectory corresponding to the trajectory data and the sensitive position is expressed as follows: .

[0039] Specifically, represents the privacy budget weight that is dynamically adjusted as the distance between the user's location point in the trajectory corresponding to the trajectory data and the sensitive location changes. It should be noted that exist When it is 1, it means that the privacy budget allocation outside the sensitive area is completely based on the user LBS location; when , that is, user When in sensitive areas, It gradually decreases as the distance decreases, reflecting the increase in privacy risk as the distance decreases. For requests outside of sensitive areas, follow The privacy budget is allocated, that is, the allocated privacy budget is: , which indicates that the greater the distance, the lower the risk and the more budget allocated; while the user When in sensitive areas, due to the relatively high risk, the corresponding privacy budget allocation will be reduced accordingly, thereby protecting user privacy.

[0040] In step 102, the vehicle client performs Laplace noise addition on the trajectory data according to the privacy budget corresponding to the user to obtain the noisy trajectory data of the user's vehicle.

[0041] Specifically, as mentioned above, for the dataset Perform the following Laplace noise processing: , The scale of the noise corresponding to the Laplace noise addition is determined by the following expression: ; in, represents the sensitivity, which means that for the query function For the data set The maximum change in the output result is represents the privacy budget of the user, Indicates the scale of the noise.

[0042] In step 103, the vehicle client sends the noisy trajectory data to the RSU, the noisy trajectory data is used for user clustering and used as a training sample for the local model corresponding to the vehicle client to which the user belongs, the local model corresponding to the vehicle client to which the user belongs is used to obtain the inter-cluster local model corresponding to the cluster to which the user belongs, the inter-cluster local model corresponding to the cluster to which the user belongs is used to obtain the global model, wherein the inter-cluster local model is obtained by model aggregation of the inter-cluster local models corresponding to all clusters respectively, the global model is obtained by model aggregation of the inter-cluster local models corresponding to all clusters respectively, and the global model is used to provide LBS services for users.

[0043] Specifically, the vehicle client adds noise to the user's trajectory data locally and sends the noisy trajectory data to the RSU. Therefore, there is no worry about user privacy leakage. In addition, since the vehicle client designs a personalized privacy budget allocation plan, it provides users with customized privacy protection measures, which meets the user's privacy protection needs while providing users with a better user experience.

[0044] The present invention also provides a privacy protection method for adaptive aggregation, which is applied to RSU, and the method includes: Acquire noisy trajectory data of vehicles of multiple users from vehicle clients of the multiple users, wherein the noisy trajectory data is obtained by performing Laplace noise on the trajectory data of the vehicles of the users by the corresponding vehicle clients according to the privacy budgets corresponding to the users, and the privacy budgets corresponding to the users are obtained by performing privacy budget allocation on the trajectory data of the vehicles of the users by the corresponding vehicle clients; Clustering the multiple users based on the noisy trajectory data of the vehicles of the multiple users to obtain multiple clusters; Aggregating the local models uploaded by all vehicle clients in the corresponding cluster to obtain the inter-cluster local models corresponding to all clusters, and sending the inter-cluster local models to the vehicle clients in the corresponding cluster for training to update the inter-cluster local models; The inter-cluster local models corresponding to all the clusters are sent to the cloud server, and the inter-cluster local models corresponding to all the clusters are used for model aggregation to obtain a global model, and the global model is used to provide LBS services for users. Specifically, the present invention proposes a hierarchical federated learning algorithm.

[0045] After the vehicle client uses Laplace noising to perform personalized noise processing on the trajectory data of the user's vehicle to meet the local differential privacy requirements, the RSU receives the noisy trajectory data sent by the vehicle client and performs density clustering based on the similarity of vehicle trajectories.

[0046] Defining Users Location By latitude ,longitude and time Composition, that is , then the user The trajectory of The original trajectory data is composed of the position points passed within the time period, and is expressed as follows: In order to prevent the leakage of the user's vehicle trajectory information, we The vehicle trajectory data is added with Laplace noise, and the trajectory data after adding noise is expressed as follows: Next, the RSU analyzes the noisy trajectory data of multiple vehicles sent by the multiple vehicle clients and constructs a Hausdorff distance matrix, and performs trajectory clustering based on the matrix to obtain a trajectory clustering result.

[0047] Specifically, define the user Vehicles The trajectory sequence is expressed as .user vehicle The trajectory sequence is expressed as , then the Hausdorff distance between two trajectories is defined as: in, and From the trajectory To track and from the trajectory To track The one-way Hausdorff distance. For the trajectory Up a little; For the trajectory Up a little; is the distance between two points. Representation trajectory Each point in The closest trajectory to this point middle The distance between points, It means to sort these distances and then take the maximum value among these minimum distances as The value of The smaller the Hausdorff distance, the more similar the two trajectories are.

[0048] Based on the above operations, by calculating the Hausdorff distance between each pair of trajectories, we construct a The Hausdorff distance matrix , which is used as the input of the DBSCAN density clustering algorithm. The matrix is ​​defined as ,in , It should be noted that the obtained trajectory clustering results reflect the similarity of trajectory data of vehicles of different users, and further reflect the behaviors and preferences of different users.

[0049] For example, there are Clients and RSU, according to the trajectory clustering results The cluster is Each client There is a local dataset , each RSU Corresponding to multiple clients in a cluster, the number of clusters The corresponding set is , The total number of clients contained in the cluster is Therefore, the The maximum number of clients that an RSU can communicate with is , the client's local data set is represented as , all local data sets of the entire training user .

[0050] After clustering is completed, the RSU aggregates the local models uploaded by all vehicle clients in the corresponding cluster to obtain the inter-cluster local models corresponding to all clusters, and sends the inter-cluster local models to the vehicle clients in the corresponding cluster for training to update the inter-cluster local models. The vehicle clients perform local training based on the initial global model sent by the cloud server to obtain the local model.

[0051] The following first explains the training of the local model by the vehicle client.

[0052] Specifically, the vehicle client clips the threshold value according to the gradient Clip the gradient so that ,in For the The training parameters of the client. The loss function of the client is defined as: in, The dataset for model training, for Middle Therefore, the client Sensitivity It can be expressed as: From the above formula, we can see that the sensitivity of the client in each round of training is , is the amplitude of the added Gaussian noise, represents the failure probability of privacy protection, so the standard deviation of Gaussian noise to be added in each round of training can be calculated as: Generate model weights using stochastic gradient descent ,in, is the learning rate, is the batch size, randomly select a sample set , the sampling probability is , for each Compute the gradient: First, the gradient is normalized to obtain: Then add Gaussian noise to the gradient and get: The model parameters are updated to ,by As an indicator of the update step size, The model parameters are updated in the round iteration as follows: Each vehicle client performs stochastic gradient descent locally, and the number of epochs of local training is set to The vehicle client completes each After the local training epoch, the updated weights are obtained and sent to the RSU. In the round training iteration, each RSU passes After round aggregation, the weights obtained are sent to the cloud server. In the In the round of local aggregation, the client The local model after the local training epoch update is: The client is completing After a round of local training, Sent to RSU. RSU updates the inter-cluster local model by aggregating the update weights of all clients: After completing Before round aggregation, RSU will obtain the local model between clusters Distributed to all clients in clusters for training and updating.

[0053] Finish After round aggregation, RSU will train the local model between clusters Sent to the cloud server. The cloud server updates the global model by aggregating the inter-cluster local models from all RSUs: Based on the global model obtained above If convergence is achieved, the training is completed, otherwise the iteration continues.

[0054] Through the above method, RSU clusters vehicle clients to ensure that subsequent training can obtain models that better meet user needs. After training the model in the same cluster, aggregating it to obtain a local model between clusters can reduce the impact of heterogeneous data and speed up model training, thereby saving communication resources and shortening upload time.

[0055] The present invention also provides an adaptive aggregation privacy protection method, which is applied to a cloud server, and the method comprises: Sending an initial global model, where the initial global model is used for each user's vehicle client to train and obtain a local model; Receive local models corresponding to all clusters sent by the RSU, wherein the local models corresponding to all clusters are trained by the RSU based on the noisy trajectory data of the vehicles of the users in each cluster, the users in each cluster are determined by the RSU based on the clustering of the noisy trajectory data of the vehicles of all users, the noisy trajectory data are obtained by the corresponding vehicle client performing Laplace noise on the trajectory data of the vehicle of the user according to the privacy budget corresponding to the user, and the privacy budget corresponding to the user is obtained by the corresponding vehicle client performing privacy budget allocation on the trajectory data of the vehicle of the user; The local models corresponding to all clusters are aggregated to obtain a global model, and the global model is used to provide LBS services for users.

[0056] Specifically, the cloud server sends an initial global model, which is used for each user's vehicle client to train and obtain a local model. Optionally, the cloud server can send the initial global model to the RSU, and the RSU sends a global training model to each cluster and sets a gradient clipping threshold according to the number of vehicles in each cluster after clustering.

[0057] like Figure 2As shown in the figure, the user's vehicle client (Vehicle User) collects the user's trajectory data and performs differential privacy noise addition on it to achieve personalized privacy protection (Personalized Privacy Protection). After data generalization (Trajectory Generalization) is performed at the vehicle layer, the road side unit (RSU) as the edge server layer (Edge Server Layer) receives the noisy trajectory data from the vehicle layer, constructs the Hausdorff distance matrix (Hausdorff Distance Matrix), and clusters the pairs through the density clustering (DBSCAN Clustering) algorithm. The vehicle client performs local training (Local Training) on ​​the local model, in which Gaussian noise (Gradient Add gaussian Noise) is added in each round of training. After the training is completed, the RSU downloads the local model trained by the vehicle client corresponding to each cluster (Upload Updated Model) and aggregates (Local Model Aggregation) to obtain the inter-cluster local model, and sends the inter-cluster local model to the vehicle client for training and updating. After the training is completed, the inter-cluster local model is uploaded to the central server (CentralServer) as the cloud server layer (Cloud Server Layer) for global model aggregation (Global Aggregation). In addition, the cloud server also sends the initial global model to each vehicle client for downloading the global model (Download Global Model) for training and updating.

[0058] Through the privacy protection method of adaptive aggregation of heterogeneous data, the balance between the personalized needs of users for privacy protection and service quality is achieved. Only the original data of the user's vehicle client is stored and used locally. After the vehicle client is noisy, the noisy trajectory data is uploaded to the RSU side for subsequent processing, thus ensuring the user's location privacy security. Due to the customized privacy and security protection for users, the vehicle client will upload a large amount of heterogeneous data. Using FL hierarchical learning, the heterogeneous data is efficiently integrated and utilized on the basis of protecting the privacy of user trajectories. All noisy trajectory data are clustered on the RSU side. Each cluster corresponds to an RSU to aggregate the local model to obtain the inter-cluster local model. Finally, the cloud server obtains all the inter-cluster local models and aggregates them to obtain the global model. In this way, while providing personalized services to users and improving user experience and service quality, it can effectively deal with the data heterogeneity problem caused by the difference in users' privacy protection needs, save communication resources, and shorten upload time.

[0059] The present invention also provides an adaptive aggregation privacy protection device, such as Figure 3 As shown, the vehicle client applied to the user includes: A privacy budget allocation unit 201 is used to allocate a privacy budget to the trajectory data of the vehicle of the user to obtain a privacy budget corresponding to the user; A trajectory noisy unit 202, configured to perform Laplace noise on the trajectory data according to the privacy budget corresponding to the user, to obtain noisy trajectory data of the user's vehicle; and The sending unit 203 is used to send the noisy trajectory data to the roadside unit RSU, the noisy trajectory data is used for user clustering, and is used as a training sample for the local model corresponding to the vehicle client to which the user belongs, the local model corresponding to the vehicle client to which the user belongs is used to obtain the inter-cluster local model corresponding to the cluster to which the user belongs, and the inter-cluster local model corresponding to the cluster to which the user belongs is used to obtain the global model, wherein the inter-cluster local model is obtained by model aggregation of the inter-cluster local models corresponding to all clusters respectively, and the global model is obtained by model aggregation of the inter-cluster local models corresponding to all clusters respectively, and the global model is used to provide LBS services for users.

[0060] The present invention also provides an electronic device, such as Figure 4 As shown, a vehicle client applied to a user includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the privacy protection method for adaptive aggregation as described in the first aspect is implemented.

[0061] The electronic device may include: a processor 410, a communications interface 420, a memory 430 and a communication bus 440, wherein the processor 410, the communications interface 420 and the memory 430 communicate with each other via the communication bus 440. The processor 410 may call the logic instructions in the memory 430 to execute the privacy protection method of adaptive aggregation, which is applied to the vehicle client of the user, and the method includes: Allocating a privacy budget to the trajectory data of the vehicle of the user to obtain a privacy budget corresponding to the user; According to the privacy budget corresponding to the user, Laplace noise is performed on the trajectory data to obtain noisy trajectory data of the user's vehicle; and The noisy trajectory data is sent to the roadside unit RSU, the noisy trajectory data is used for user clustering, and is used as a training sample for the local model corresponding to the vehicle client to which the user belongs, the local model corresponding to the vehicle client to which the user belongs is used to obtain the inter-cluster local model corresponding to the cluster to which the user belongs, and the inter-cluster local model corresponding to the cluster to which the user belongs is used to obtain the global model, wherein the inter-cluster local model is obtained by model aggregation of the inter-cluster local models corresponding to all clusters respectively, the global model is obtained by model aggregation of the inter-cluster local models corresponding to all clusters respectively, and the global model is used to provide LBS services for users.

[0062] In addition, the logic instructions in the above-mentioned memory 430 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM: Read-Only Memory), random access memory (RAM: Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0063] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to perform a privacy protection method for adaptive aggregation, which is applied to a vehicle client of a user, and the method comprises: Allocating a privacy budget to the trajectory data of the vehicle of the user to obtain a privacy budget corresponding to the user; According to the privacy budget corresponding to the user, Laplace noise is performed on the trajectory data to obtain noisy trajectory data of the user's vehicle; and The noisy trajectory data is sent to the roadside unit RSU, the noisy trajectory data is used for user clustering, and is used as a training sample for the local model corresponding to the vehicle client to which the user belongs, the local model corresponding to the vehicle client to which the user belongs is used to obtain the inter-cluster local model corresponding to the cluster to which the user belongs, and the inter-cluster local model corresponding to the cluster to which the user belongs is used to obtain the global model, wherein the inter-cluster local model is obtained by model aggregation of the inter-cluster local models corresponding to all clusters respectively, the global model is obtained by model aggregation of the inter-cluster local models corresponding to all clusters respectively, and the global model is used to provide LBS services for users.

[0064] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the computer can execute the privacy protection method of adaptive aggregation provided by the above methods, which is applied to a vehicle client of a user. The method includes: Allocating a privacy budget to the trajectory data of the vehicle of the user to obtain a privacy budget corresponding to the user; According to the privacy budget corresponding to the user, Laplace noise is performed on the trajectory data to obtain noisy trajectory data of the user's vehicle; and The noisy trajectory data is sent to the roadside unit RSU, the noisy trajectory data is used for user clustering, and is used as a training sample for the local model corresponding to the vehicle client to which the user belongs, the local model corresponding to the vehicle client to which the user belongs is used to obtain the inter-cluster local model corresponding to the cluster to which the user belongs, and the inter-cluster local model corresponding to the cluster to which the user belongs is used to obtain the global model, wherein the inter-cluster local model is obtained by model aggregation of the inter-cluster local models corresponding to all clusters respectively, the global model is obtained by model aggregation of the inter-cluster local models corresponding to all clusters respectively, and the global model is used to provide LBS services for users.

[0065] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0066] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. 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 adaptive aggregation, characterized in that: Applied to a vehicle client of a user, the method comprises: Allocating a privacy budget to the trajectory data of the vehicle of the user to obtain a privacy budget corresponding to the user; According to the privacy budget corresponding to the user, Laplace noise is performed on the trajectory data to obtain noisy trajectory data of the user's vehicle; and The noisy trajectory data is sent to the roadside unit RSU, the noisy trajectory data is used for user clustering, and is used as a training sample for the local model corresponding to the vehicle client to which the user belongs, the local model corresponding to the vehicle client to which the user belongs is used to obtain the inter-cluster local model corresponding to the cluster to which the user belongs, and the inter-cluster local model corresponding to the cluster to which the user belongs is used to obtain the global model, wherein the inter-cluster local model is obtained by model aggregation of the inter-cluster local models corresponding to all clusters respectively, the global model is obtained by model aggregation of the inter-cluster local models corresponding to all clusters respectively, and the global model is used to provide LBS services for users.

2. The privacy protection method of adaptive aggregation according to claim 1, characterized in that: The allocating a privacy budget to the trajectory data of the vehicle of the user includes: Determine polar coordinates corresponding to a plurality of location point data in the trajectory data, and analyze the polar coordinates corresponding to the plurality of location point data by a probability density function to obtain a distribution of location points corresponding to the trajectory data; Determining a sensitive area of ​​the user based on the sensitive position of the user and the radius of the sensitive area; and A privacy budget corresponding to the user is determined according to a relationship between the distribution of the location points and the sensitive area.

3. The privacy protection method of adaptive aggregation according to claim 2, characterized in that: The radius of the sensitive area It is expressed as: ； in, represents the sum of the sensitive distances corresponding to the LBS positions of at least one service request point of the user in the trajectory corresponding to the trajectory data, where the sensitive distance corresponding to the LBS position is the distance between the LBS position and the sensitive position corresponding to the LBS position, is the Lambert w function, with parameters Used to normalize the radius of the sensitive area, is the distance between the LBS position of the user without noise and the LBS position with noise that the user can accept, that is, equivalent to the error that the user can tolerate. is the privacy budget corresponding to the user.

4. The privacy protection method of adaptive aggregation according to claim 2, characterized in that: The privacy budget includes the user's privacy budget outside the sensitive area and the user's privacy budget in sensitive areas ,in: ； in, Indicates the user The LBS location The sensitive position corresponding to the LBS position The sensitive distance between represents the number of at least one service request point, represents the distance between the position point of the user in the trajectory corresponding to the trajectory data and the sensitive position, represents the radius of the sensitive area, A weighted function representing the Euclidean distance between the position point of the user in the trajectory corresponding to the trajectory data and the sensitive position is expressed as follows: .

5. The privacy protection method of adaptive aggregation according to claim 1, characterized in that: The scale of the noise corresponding to the Laplace noise addition is determined by the following expression: ； in, represents the sensitivity, which means that for the query function For the data set The maximum change in the output result is represents the privacy budget of the user, Indicates the scale of the noise.

6. A privacy protection method for adaptive aggregation, characterized in that: Applied to RSU, the method comprises: Acquire noisy trajectory data of vehicles of multiple users from vehicle clients of the multiple users, wherein the noisy trajectory data is obtained by performing Laplace noise on the trajectory data of the vehicles of the users by the corresponding vehicle clients according to the privacy budgets corresponding to the users, and the privacy budgets corresponding to the users are obtained by performing privacy budget allocation on the trajectory data of the vehicles of the users by the corresponding vehicle clients; Clustering the multiple users based on the noisy trajectory data of the vehicles of the multiple users to obtain multiple clusters; Aggregating the local models uploaded by all vehicle clients in the corresponding cluster to obtain inter-cluster local models corresponding to all clusters, and sending the inter-cluster local models to the vehicle clients in the corresponding cluster for training to update the inter-cluster local models; and The inter-cluster local models corresponding to all the clusters are sent to the cloud server, and the inter-cluster local models corresponding to all the clusters are used for model aggregation to obtain a global model, and the global model is used to provide LBS services for users.

7. A privacy protection method for adaptive aggregation, characterized in that: Applied to a cloud server, the method comprises: Sending an initial global model, where the initial global model is used for each user's vehicle client to train and obtain a local model; Receive local models corresponding to all clusters sent by the RSU, wherein the local models corresponding to all clusters are trained by the RSU based on the noisy trajectory data of the vehicles of the users in each cluster, the users in each cluster are determined by the RSU based on the clustering of the noisy trajectory data of the vehicles of all users, the noisy trajectory data are obtained by the corresponding vehicle client performing Laplace noise on the trajectory data of the vehicle of the user according to the privacy budget corresponding to the user, and the privacy budget corresponding to the user is obtained by the corresponding vehicle client performing privacy budget allocation on the trajectory data of the vehicle of the user; and The local models corresponding to all clusters are aggregated to obtain a global model, and the global model is used to provide LBS services for users.

8. An adaptive aggregation privacy protection device, applied to a user's vehicle client, characterized in that: include: A privacy budget allocation unit, configured to allocate a privacy budget to the trajectory data of the vehicle of the user to obtain a privacy budget corresponding to the user; A trajectory noise adding unit, configured to perform Laplace noise adding on the trajectory data according to the privacy budget corresponding to the user, so as to obtain the noisy trajectory data of the vehicle of the user; as well as A sending unit is used to send the noisy trajectory data to a roadside unit RSU, the noisy trajectory data is used for clustering users and used as a training sample for a local model corresponding to the vehicle client to which the user belongs, the local model corresponding to the vehicle client to which the user belongs is used to obtain an inter-cluster local model corresponding to the cluster to which the user belongs, the inter-cluster local model corresponding to the cluster to which the user belongs is used to obtain a global model, wherein the inter-cluster local model is obtained by model aggregation of inter-cluster local models corresponding to all clusters respectively, the global model is obtained by model aggregation of inter-cluster local models corresponding to all clusters respectively, and the global model is used to provide LBS services for users.

9. An electronic device, applied to a user's vehicle client, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the privacy protection method for adaptive aggregation according to any one of claims 1 to 5 is implemented.

10. A computer-readable storage medium, applied to a user's vehicle client, having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the privacy protection method for adaptive aggregation as claimed in any one of claims 1 to 5 is implemented.

Images