Location privacy protection method based on virtual track generation

Abstract

The invention relates to a position privacy protection method based on virtual track generation, and belongs to the technical field of position privacy protection. According to the method, two generative adversarial networks are constructed, travel habits of users and track information of real users are learned respectively, and virtual user LBS query service and corresponding track data are generated. By comprehensively considering the influence of the user behavior pattern and the geographic topological structure, the generated virtual track is more real and credible, and a better privacy protection effect is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of location privacy protection, and in particular relates to a location privacy protection method based on virtual trajectory generation. Background Art

[0002] With the continuous development of information technology, location-based services (LBS) have become more and more important and popular. However, many LBS may collect users' location information, which may lead to the leakage of location privacy. Attackers can obtain users' driving trajectories through the correlation of time-correlated locations. In order to solve this problem, it is very necessary to study a technology that can effectively protect users' location privacy. Summary of the invention

[0003] 1. Technical issues to be resolved

[0004] The technical problem to be solved by the present invention is to design a location privacy protection method to achieve the goal of protecting the user's trajectory privacy.

[0005] (II) Technical solution

[0006] In order to solve the above technical problems, the present invention provides a location privacy protection method based on virtual trajectory generation, comprising the following steps:

[0007] Step 1: Data preparation

[0008] In the data preparation stage, complete the data collection and preprocessing work:

[0009] The data includes two types: LBS service request data and user driving trajectory data;

[0010] The LBS service request data includes the following key data: first location information, first time information, user information, service type information, additional parameters and weather information;

[0011] The first location information is the core data of LBS service, including the user's latitude and longitude coordinates; the first time information refers to the time of LBS service request; time information is used to analyze user behavior and provide customized services; user information refers to the user's identity information; service type information refers to the specific service content requested by the user, including the content of the point of interest query and navigation requirements; additional parameters refer to the user's requirements for service quality, including the user's requirements for search distance in the point of interest query service; weather information refers to the weather conditions on the day of the service request;

[0012] The user driving trajectory data includes the following key data: second location information, second time information, user information, and vehicle information;

[0013] The second location information refers to the user's trajectory data, including a location coordinate sequence of length n; the second time information refers to the time corresponding to the user's driving trajectory; the user information refers to the user's identity information; the vehicle information refers to the type of vehicle the user takes;

[0014] In the data preprocessing stage, data encoding and data filling are completed;

[0015] Step 2: Model building phase

[0016] Construct two learning models to learn the LBS service request obtained in step 1 and the real user driving trajectory respectively, and generate false LBS service requests and corresponding false trajectories;

[0017] In step 2, a GAN model based on a convolutional neural network is first built. In the GAN model, the generator G and the discriminator D use convolutional neural networks, and the structure includes: convolution layer, pooling layer, and fully connected layer in sequence; the generator G learns the real LBS service request data to generate a fake LBS service request, and the discriminator D distinguishes whether the LBS service request sample comes from the real service request data set or the generated service request data through learning;

[0018] Secondly, we build a GAN network based on transformer-LSTM. In the GAN network, the generator G is designed based on the transformer model, and the discriminator D uses the LSTM network structure. The generator G learns the distribution characteristics and behavior patterns of the real trajectory data, and synthesizes false trajectory data based on the input real trajectory data and random noise. The discriminator D distinguishes whether the trajectory sample comes from the real trajectory data or the synthesized false trajectory data through learning.

[0019] Step 3: Model training phase

[0020] For the two learning models constructed in step 2, perform the following model training process respectively:

[0021] (1) Initialize the initial parameters of the model;

[0022] (2) Train the discriminator. During training, fix the parameters of the generator. Generate false data through the generator and input the false data into the discriminator together with the real data. Calculate the loss function and complete the parameter update.

[0023] (3) Train the generator. During training, the parameters of the discriminator need to be fixed. The generator generates false data, inputs it into the discriminator, calculates the loss function, and completes the parameter update.

[0024] (4) Repeat the process of (2) and (3) until the network model converges;

[0025] Step 4: Trajectory Generation Phase

[0026] According to the user's privacy protection requirements, k-1 false LBS service requests and corresponding false trajectories need to be generated to meet the requirements of k-anonymity. The meaning of k-anonymity is that the data set contains one real data and k-1 false data, and the false data and the real data have a high similarity, so the attacker cannot identify which data is real and which data is false. The total number of real data and false data is k, which is called k-anonymity.

[0027] According to the time information and weather information in the user's LBS service request, a trained generator is used to generate k-1 false LBS service requests; the false LBS service request data includes the user's starting location and end location; further, a false trajectory from the user's starting location to the end location is generated using the trained generator.

[0028] Preferably, the first time information and the second time information both include a day of the week and a time point.

[0029] Preferably, in the data preprocessing stage, the method for completing the data encoding work is specifically as follows: for time attributes and category attribute information, one-hot encoding is used for encoding; the week attributes in the first and second time information are encoded as 7-dimensional binary vectors, and the point of interest attribute information is encoded as a 21-dimensional binary vector, and the point of interest attribute information involves service type information and additional parameters; for location attribute information, longitude and latitude coordinates are used to represent it; user information is not encoded and is only used to identify the ownership of the data.

[0030] Preferably, in the data preprocessing stage, the method for completing the data filling work is specifically: for different user driving trajectory data lengths, the trajectory filling technology is applied, and the empty trajectory points are filled into each trajectory using zero pre-filling, so that all trajectories have the same length.

[0031] The present invention also provides a system for implementing the method.

[0032] (III) Beneficial effects

[0033] This paper proposes a location privacy protection method based on virtual trajectory generation. Two generative adversarial networks are constructed to learn the user's travel habits and the real user's trajectory information respectively, and generate a virtual user LBS query service and corresponding trajectory data. By comprehensively considering the influence of user behavior patterns and geographic topology, the generated virtual trajectory is more authentic and has a better privacy protection effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flow chart of the method of the present invention;

[0035] Figure 2 It is a schematic diagram of the structure of a CNN-based generative adversarial network of the present invention;

[0036] Figure 3 It is a schematic diagram of the structure of the generative adversarial network based on transformer-LSTM of the present invention. DETAILED DESCRIPTION

[0037] In order to make the purpose, content and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below in conjunction with the drawings and examples.

[0038] The present invention proposes a location privacy protection method based on virtual trajectory generation. By collecting real LBS service requests and vehicle driving trajectories, a generative adversarial network (GAN) is used to generate virtual trajectory samples to meet the requirements of k-anonymity. In this method, two generative adversarial networks are constructed to learn the user's travel habits and the real user's trajectory information respectively, and generate virtual user LBS query services and corresponding trajectory data. The virtual trajectory sample distribution generated by this method is close to the real sample distribution, and the attacker cannot distinguish between the real trajectory and the virtual trajectory, thereby achieving the goal of protecting the user's trajectory privacy.

[0039] The overall process of the present invention is as follows Figure 1 As shown, the following steps are included:

[0040] Step 1: Data preparation

[0041] In the data preparation stage, the main work is to complete the data collection and preprocessing;

[0042] The data includes two types: LBS service request data and user driving trajectory data;

[0043] The LBS service request data includes the following key data: first location information, first time information, user information, service type information, additional parameters and weather information.

[0044] (1) The first location information is the core data of the LBS service, including the user's latitude and longitude coordinates;

[0045] (2) The first time information refers to the time of LBS service request; time information is important for analyzing user behavior and providing customized services, including the day of the week and time point.

[0046] (3) User information refers to the user’s identity information, such as device ID;

[0047] (4) Service type information refers to the specific service content requested by the user, such as the content of the point of interest query, navigation requirements, etc.;

[0048] (5) Additional parameters refer to the user's requirements for service quality; for example, the user's requirements for search distance in the point of interest query service;

[0049] (6) Weather information refers to the weather conditions on the day the service is requested. This information will affect the user's travel preferences.

[0050] The user driving trajectory data includes the following key data: second location information, second time information, user information, and vehicle information.

[0051] (1) The second location information refers to the user's trajectory data, including a location coordinate sequence with a length of n;

[0052] (2) The second time information refers to the time corresponding to the user's driving trajectory, including the day of the week and the time point;

[0053] (3) User information refers to the user’s identity information, such as device ID;

[0054] (4) Vehicle information refers to the type of vehicle the user takes, such as bus, taxi, or private car.

[0055] Data preprocessing mainly completes data encoding and data filling.

[0056] For time attributes and category attribute information, one-hot encoding is used for encoding. For example, the day of the week attribute in the first and second time information is encoded as a 7-dimensional binary vector, and the POI attribute information (involving service type information and additional parameters) is encoded as a 21-dimensional binary vector; for location attribute information (including the first and second location information), longitude and latitude coordinates are used for representation. User information is not encoded and is only used to identify the ownership of the data.

[0057] Since the length of each user's driving trajectory data is different, the trajectory filling technology is applied to fill the empty trajectory points into each trajectory using zero pre-filling to ensure that all trajectories have the same length, which is convenient for speeding up the training process.

[0058] Step 2: Model building phase

[0059] The present invention constructs two learning models to respectively learn the LBS service request obtained in step 1 and the real user driving trajectory, and generates a false LBS service request and a corresponding false trajectory.

[0060] First, build a GAN model based on convolutional neural network, see Figure 2In the GAN model, the generator G and the discriminator D use convolutional neural networks, and the structures both include: convolution layer, pooling layer, and fully connected layer in sequence; the generator G learns the real LBS service request data to generate a false LBS service request, and the discriminator D distinguishes whether the LBS service request sample comes from the real service request data set or the generated service request data through learning. In this embodiment, the number of convolution layers and pooling layers can be set according to actual conditions.

[0061] Secondly, build a GAN network based on transformer-LSTM, such as Figure 3 As shown in the figure, in the GAN network, the generator G is designed based on the transformer model, and the discriminator D uses the LSTM network structure. The generator G learns the distribution characteristics and behavior patterns of the real trajectory data, and synthesizes false trajectory data based on the input real trajectory data and random noise. The discriminator D distinguishes whether the trajectory samples come from real trajectory data or synthesized false trajectory data through learning.

[0062] The design principles of the above two learning models are as follows:

[0063] In the process of generating virtual service requests, the format and size of the data set are relatively simple, so the GAN network model chooses the basic neural network, which has a simple model and short training time.

[0064] In the process of generating virtual trajectory services, the data set is more complex, in which the trajectory data is two-dimensional and composed of time and space. The GAN network model must be able to recognize the time series relationship therein, and the transformer and LSTM models are learning models specially designed for sequence data, which can meet the requirements of learning trajectory data, which is a time series data. In the scenario of the present invention, it is necessary to ensure that the generated false trajectory is more realistic, and the requirements for the generator G are higher, and better performance is required. The discriminator D is used to assist in optimizing the generator G, and the performance requirements are lower than those of the generator G. Therefore, the generator G selects a relatively more complex and more powerful transformer model, which can guarantee better performance requirements, and the discriminator D selects a more basic and simpler LSTM model, which can reduce the training time while meeting the functional requirements.

[0065] Step 3: Model training phase

[0066] For the two learning models constructed in step 2, perform the following model training process respectively:

[0067] (1) Initialize the initial parameters of the model;

[0068] (2) Train the discriminator. During training, the parameters of the generator need to be fixed. Generate false data through the generator and input the false data into the discriminator together with the real data. Calculate the loss function and complete the parameter update.

[0069] (3) Train the generator. During training, the parameters of the discriminator need to be fixed. The generator generates false data, inputs it into the discriminator, calculates the loss function, and completes the parameter update.

[0070] (4) Repeat the process of (2) and (3) until the network model converges.

[0071] Step 4: Trajectory Generation Phase

[0072] According to the user's privacy protection requirements, k-1 false LBS service requests and corresponding false trajectories need to be generated to meet the requirements of k-anonymity. The meaning of k-anonymity is that the data set contains one real data and k-1 false data, and the false data and the real data have a high similarity, so that the attacker cannot identify which data is real and which data is false. The total number of real data and false data is k, which is called k-anonymity. The larger k is, the more false data is generated, and the better the privacy protection effect is.

[0073] According to the time information, weather information, etc. in the user's LBS service request, use the trained Figure 2 The generator in generates k-1 false LBS service requests; the false LBS service request data includes the user's starting location and end location (point of interest location); further, using the trained Figure 3 The generator in generates a fake trajectory from the user's starting position to the end position.

[0074] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A location privacy protection method based on virtual trajectory generation, characterized in that: The following steps are involved: Step 1: Data preparation In the data preparation stage, complete the data collection and preprocessing work: The data includes two types: LBS service request data and user driving trajectory data; The LBS service request data includes the following key data: first location information, first time information, user information, service type information, additional parameters and weather information; The first location information is the core data of LBS service, including the user's latitude and longitude coordinates; the first time information refers to the time of LBS service request; time information is used to analyze user behavior and provide customized services; user information refers to the user's identity information; service type information refers to the specific service content requested by the user, including the content of the point of interest query and navigation requirements; additional parameters refer to the user's requirements for service quality, including the user's requirements for search distance in the point of interest query service; weather information refers to the weather conditions on the day of the service request; The user driving trajectory data includes the following key data: second location information, second time information, user information, and vehicle information; The second location information refers to the user's trajectory data, including a location coordinate sequence of length n; the second time information refers to the time corresponding to the user's driving trajectory; User information refers to the user's identity information; vehicle information refers to the type of vehicle the user takes; In the data preprocessing stage, data encoding and data filling are completed; Step 2: Model building phase Construct two learning models to learn the LBS service request obtained in step 1 and the real user driving trajectory respectively, and generate false LBS service requests and corresponding false trajectories; In step 2, we first build a GAN model based on a convolutional neural network. In the GAN model, the generator G and the discriminator D use convolutional neural networks, and their structures include: convolutional layer, pooling layer, and fully connected layer in sequence; The generator G learns from real LBS service request data to generate fake LBS service requests, and the discriminator D learns to distinguish whether the LBS service request samples come from the real service request dataset or the generated service request data; Secondly, we build a GAN network based on transformer-LSTM. In the GAN network, the generator G is designed based on the transformer model, and the discriminator D uses the LSTM network structure. The generator G learns the distribution characteristics and behavior patterns of the real trajectory data, and synthesizes false trajectory data based on the input real trajectory data and random noise. The discriminator D distinguishes whether the trajectory sample comes from the real trajectory data or the synthesized false trajectory data through learning. Step 3: Model training phase For the two learning models constructed in step 2, perform the following model training process respectively: (1) Initialize the initial parameters of the model; (2) Train the discriminator. During training, fix the parameters of the generator. Generate false data through the generator and input the false data into the discriminator together with the real data. Calculate the loss function and complete the parameter update. (3) Training the generator; During training, the parameters of the discriminator need to be fixed; The generator generates false data, inputs it into the discriminator, calculates the loss function, and completes the parameter update; (4) Repeat the process of (2) and (3) until the network model converges; Step 4, trajectory generation phase According to the user's privacy protection requirements, k-1 false LBS service requests and corresponding false trajectories need to be generated to meet the requirements of k-anonymity. The meaning of k-anonymity is that the data set contains one real data and k-1 false data, and the false data and the real data have a high similarity, so the attacker cannot identify which data is real and which data is false. The total number of real data and false data is k, which is called k-anonymity. According to the time information and weather information in the user's LBS service request, a trained generator is used to generate k-1 false LBS service requests; the false LBS service request data includes the user's starting location and end location; further, a false trajectory from the user's starting location to the end location is generated using the trained generator.

2. The method according to claim 1, characterized in that The first time information and the second time information both include a day of the week and a time point.

3. The method according to claim 2, characterized in that In the data preprocessing stage, the method for completing data encoding is as follows: for time attributes and category attribute information, one-hot encoding is used for encoding; the week attributes in the first and second time information are encoded as 7-dimensional binary vectors, and the point of interest attribute information is encoded as a 21-dimensional binary vector. The point of interest attribute information involves service type information and additional parameters; for location attribute information, longitude and latitude coordinates are used to represent it; user information is not encoded and is only used to identify the ownership of the data.

4. The method according to claim 3, characterized in that In the data preprocessing stage, the method for completing the data filling work is specifically as follows: for different user driving trajectory data lengths, the trajectory filling technology is applied, and the empty trajectory points are filled into each trajectory using zero pre-filling so that all trajectories have the same length.

5. The method according to claim 1, characterized in that The number of convolutional layers and pooling layers is set according to actual conditions.

6. The method according to claim 1, characterized in that The larger k is, the more false data is generated and the better the privacy protection effect is.

7. The method according to claim 1, characterized in that User information refers to the device ID.

8. The method according to claim 1, characterized in that The destination location is determined based on the location of the point of interest.

9. The method according to any one of claims 1 to 8, characterized in that The method is applied in the field of network security technology.

10. A system for implementing the method according to any one of claims 1 to 9.

Images