CN115168900A - Track data privacy protection method and system for intelligent traffic system - Google Patents

Abstract

The invention discloses a track data privacy protection method and a track data privacy protection system for an intelligent traffic system, wherein the method comprises the following steps: acquiring a real-time real track data set of a user; loading the real track data set into a pre-constructed and trained end-to-end deep learning model to generate a synthetic track data set; clustering the track points of the synthetic track data set under each timestamp by using a k-means clustering algorithm based on Euclidean distance, and generating generalized tracks by randomly combining track cluster centers under different timestamps after clustering; and adding Laplace noise and consistency constraint to the counting matrix of the generalized track to obtain a difference privacy counting matrix with limited noise quantity and issuing the difference privacy counting matrix. The synthetic track data set can ensure balance between data availability and privacy, track release data meet the differential privacy requirement, and a superior privacy guarantee is provided.

Description

A kind of trajectory data privacy protection method and system for intelligent transportation system

technical field

The invention relates to the technical field of data security protection, in particular to a method and system for privacy protection of trajectory data used in an intelligent transportation system.

Background technique

Intelligent transportation system is the effective and comprehensive application of advanced science and technology such as information technology, computer technology, data communication technology, sensor technology, electronic control technology, artificial intelligence, etc. in transportation, service control and vehicle manufacturing, strengthening vehicles, roads, users The connection between the three forms a comprehensive transportation system that guarantees safety, improves efficiency, improves the environment, and saves energy; trajectory data is an important privacy data in the smart transportation system. For example, in the urban transportation system, it can be Through statistical analysis of the moving trajectory of the vehicle, a more reasonable urban traffic plan can be formulated to avoid urban congestion. The navigation software can infer the driving preference of the owner by analyzing the trajectory of the vehicle in order to recommend the optimal driving route.

However, while the trajectory data provides convenience, it is also accompanied by corresponding privacy and security issues. After the trajectory data is collected, it is usually open to the public for researchers to mine and analyze the data; however, during the collection and use of trajectory data, attacks The attacker may track the user's movement, or further infer the user's workplace, home address, social relations, physical condition, hobbies and other private information, and sell it to a third party for economic benefits, so that the user's privacy is leaked, so how? The problem of making trajectory data work while protecting user privacy is particularly important.

At present, the anonymity algorithm k-anonymity is usually used to protect the privacy of data. k-anonymity requires each user to be indistinguishable from at least k-1 other users within a certain time and space, so that attackers cannot obtain data from at least k users. Identify the attack target and then infer its exact location; however, due to the relative discreteness of the trajectory coordinate points, it is difficult to meet the use conditions of k-anonymity, and some studies have shown that k-anonymity uses privacy protection algorithms on real data, but it still cannot resist combined attacks. , exact attack, or privacy stealing means such as background knowledge attack; and this type of algorithm needs to use clustering algorithm frequently in the process of use, which will have a relatively large impact on the trajectory accuracy; and when the data set changes, such The model needs to be rebuilt so it does not have the characteristics of reuse, which will greatly increase the extra time overhead.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the deficiencies in the prior art, provide a trajectory data privacy protection method and system for an intelligent transportation system, and solve the problems in the prior art due to frequent use of clustering algorithms and real data sets on real data sets. The non-reusability brings about the technical problems of insufficient accuracy and efficiency of trajectory data set publishing methods, and insufficient protection of trajectory data privacy.

In order to solve the above-mentioned technical problems, the present invention adopts the following technical solutions to realize:

In a first aspect, the present invention provides a trajectory data privacy protection method for an intelligent transportation system, the method comprising:

Obtain real-time real-time trajectory data sets of users;

Load the real trajectory dataset into a pre-built and trained end-to-end deep learning model to generate a synthetic trajectory dataset;

Use the k-means clustering algorithm to cluster the trajectory points of the synthetic trajectory data set under each timestamp based on the Euclidean distance, and generate the cluster centers of the trajectory clusters under different timestamps after the clustering by random combination. trajectories;

After adding Laplace noise and consistency constraints to the counting matrix of the generalized trajectory, a differentially private counting matrix with limited noise is obtained and published.

In combination with the first aspect, preferably, the training process of the end-to-end deep learning model includes the following steps:

Collect the historical real trajectory data of different users as the original trajectory data of the training model;

The original trajectory data is processed by centroid standardization to obtain the centroid deviation coordinates of each trajectory point, and the encoded standard original trajectory is obtained;

Each of the centroid deviation coordinates is processed into a 64-dimensional vector by a linear rectification function;

Use 64 LSTM Cell long short-term memory network units to perform time series prediction on the 64-dimensional vector to obtain synthetic trajectory data, and use the tanh hyperbolic tangent function to decode the centroid deviation coordinates of the synthetic trajectory data and obtain the decoded synthetic trajectory;

Calculate the trajectory similarity loss value of the synthetic trajectory through a trajectory loss function in combination with the standard original trajectory;

Perform binary classification processing on the trajectory similarity loss value through the sigmod activation function to obtain the discrimination result of the synthetic trajectory;

If the judgment result is false, then re-input the original trajectory data into the model for repeated training, and use the backpropagation algorithm to solve the model optimization problem to update the network parameters of the model, until the judgment result is true, stop training, and set the The synthetic trajectory output of the discriminant result is true to form a synthetic trajectory dataset.

With reference to the first aspect, preferably, after processing each of the centroid deviation coordinates into a 64-dimensional vector by using a linear rectification function, the method further includes: adding random noise to the vector, so that each set of vectors of the original trajectory is maintained. the same length.

With reference to the first aspect, preferably, the step of calculating the trajectory similarity loss of the synthetic trajectory by using the trajectory loss function in combination with the standard original trajectory includes:

Performing centroid normalization processing on the synthetic trajectory to obtain the corresponding centroid deviation coordinates of each trajectory point, and obtaining an encoded standard synthetic trajectory;

The centroid deviation coordinates of each of the standard original trajectory and the standard synthetic trajectory are respectively processed into corresponding 64-dimensional vectors by a linear rectification function;

Combined with the corresponding 64-dimensional vectors and the trajectory loss function of the two groups, the trajectory similarity loss value TLoss of the synthetic trajectory is obtained through the calculation and training of formula (1):

TLoss=αL _BCE (l ^t ,l ^p )+βL _GPS (t ^t ,t ^p ) (1)

where l ^t and l ^p represent the discriminative labels of the standard original trajectory and standard cooperative trajectory, respectively; t ^t and t ^p represent the 64-dimensional vector of the standard original trajectory and the corresponding standard cooperative trajectory, respectively; L _BCE represents the binary crossover Entropy loss function, L _GPS represents a loss function that uses least squares error to measure the similarity between two trajectories; α and β represent the weights of L _BCE and L _GPS , respectively.

In combination with the first aspect, preferably, the calculation formula for processing each of the centroid deviation coordinates into a 64-dimensional vector through a linear rectification function is as follows:

表示编号为i的轨迹点的64维向量，△lat_i,和△lon_i分别表示编号为i的轨迹点的经度偏差和纬度偏差，f^GPS表示线性整流函数，W_GPS表示轨迹点i的质心偏差坐标(△lat_i,△lon_i)的向量权重。In the formula,

Represents the 64-dimensional vector of the trajectory point numbered i, △lat _i , and △lon _i represent the longitude deviation and latitude deviation of the trajectory point numbered i, respectively, f ^GPS represents the linear rectification function, W _GPS represents the centroid of the trajectory point i Vector weights for the deviation coordinates (Δlat _i , Δlon _i ).

In combination with the first aspect, preferably, the calculation formula for obtaining synthetic trajectory data by using 64 LSTM Cell long short-term memory network units to perform time series prediction on the 64-dimensional vector is as follows:

O=LSTM(T,W _lstm ) (3)

In the formula: T={t ₁ ,t ₂ ,…,t _i ,…,t _maxlength }, T represents the coordinate features of all trajectory points in the original trajectory data, and t _i represents the 64-dimensional vector of the trajectory point numbered i , maxlength represents the longest length of a single trajectory in the original trajectory data; W _lstm represents the weight matrix of the input data; O represents the synthetic trajectory data, where the data contained in O is represented as O={o ₁ ,o ₂ ,...,o _i ,...,o _maxlength }, o _i represents the coordinate composite value of the output of t _i after LSTM processing.

In combination with the first aspect, preferably, the calculation formula for decoding the centroid deviation coordinates of the synthetic trajectory data using the tanh hyperbolic tangent function is as follows:

表示合成轨迹数据中轨迹点i的质心偏差坐标的解码坐标，

和

分别表示轨迹点i的经度偏差和纬度偏差；W_dGPS表示坐标向量的解码矩阵权重；D^GPS为tanh双曲正切函数。In the formula,

is the decoded coordinate representing the centroid deviation coordinate of the trajectory point i in the synthetic trajectory data,

and

respectively represent the longitude deviation and latitude deviation of the track point i; W _dGPS represents the decoding matrix weight of the coordinate vector; D ^GPS is the tanh hyperbolic tangent function.

In a second aspect, the present invention provides a trajectory data privacy protection system for an intelligent transportation system, the system comprising:

The acquisition module is used to acquire the real-time real trajectory data set of the user;

a synthetic trajectory module, used to load the real trajectory data set into a pre-built and trained end-to-end deep learning model to generate a synthetic trajectory data set;

The clustering module is configured to use the k-means clustering algorithm to cluster the trajectory points of the synthetic trajectory data set under each timestamp based on the Euclidean distance, and to pass the clustered trajectory cluster centers under different timestamps after the clustering. Generate generalization trajectories in a random combination;

A noise-adding publishing module is used to add Laplace noise and consistency constraints to the counting matrix of the generalized trajectory to obtain a differentially private counting matrix with limited noise and publish it.

With reference to the second aspect, preferably, the synthetic trajectory module includes an acquisition unit; the end-to-end deep learning model includes a trajectory generator and a trajectory discriminator; the trajectory generator includes a first input layer, a first embedding layer, a first LSTM modeling layer, a first output layer; the trajectory discriminator includes a second input layer, a second embedding layer, a second LSTM modeling layer, and a second output layer; wherein:

The acquisition unit is used to collect the historical real trajectory data of different users as the original trajectory data of the training model;

The first input layer is used to obtain the centroid deviation coordinates of each trajectory point by performing the centroid normalization process on the original trajectory data, and obtain the encoded standard original trajectory;

The first embedding layer is used to use the multilayer perceptron MLP to process each of the centroid deviation coordinates into a 64-dimensional vector through a linear rectification function; and add random noise to the vector, so that each The group vectors all keep the same length as the longest trajectory;

The first LSTM modeling layer is used to perform time series prediction processing on the 64-dimensional vector through 64 LSTM Cell long-term and short-term memory network units to obtain synthetic trajectory data;

The first output layer is used to decode the longitude and latitude deviation of the synthetic trajectory data by using the two dense layers Den through the tanh hyperbolic tangent function to obtain the decoded synthetic trajectory;

The second input layer is configured to use the standard original trajectory and the synthesized trajectory as the input data of the trajectory discriminator, and perform centroid normalization on the synthesized trajectory to obtain the corresponding centroid deviation coordinates of each trajectory point, and Obtain the encoded standard synthetic trajectory;

The second embedding layer is used for using the multi-layer perceptron MLP to process the centroid deviation coordinates of each of the standard original trajectory and the standard synthetic trajectory into a corresponding 64-dimensional vector respectively through a linear rectification function;

The second LSTM modeling layer is used to calculate the trajectory similarity loss value through the trajectory loss function using the corresponding 64-dimensional vectors of the two groups through 64 LSTM Cells;

The second output layer is used to use a dense layer Den to perform binary classification processing on the trajectory similarity loss value through the sigmod activation function to obtain the discrimination result of the synthetic trajectory; if the discrimination result is false, the original trajectory The data is re-entered into the trajectory generator for training, and the back-propagation algorithm is used to solve the model optimization problem to update the network parameters in the first LSTM modeling layer in the trajectory generator. The training is stopped until the judgment result is true, and the judgment result is The real synthetic trajectory output forms the synthetic trajectory dataset.

In a third aspect, the present invention provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, realizes the trajectory data for an intelligent transportation system according to any one of the first aspects The steps of the privacy protection method.

Compared with the prior art, the beneficial effects achieved by the present invention:

The end-to-end deep learning model provided by the present invention uses the trajectory loss function to determine that the trajectory similarity loss meets the standard to generate a synthetic trajectory to replace the real trajectory, so as to avoid privacy leakage when the real trajectory data set is encrypted and released, and in the real trajectory When the trajectory data set is updated, the model does not need to be retrained, and the synthetic trajectory can be obtained only by changing the input, which greatly reduces the time in the data protection process; in addition, random noise is added to the vector during the model training process, so that all Each set of vectors in the original trajectory maintains the same length, which speeds up the training process and improves the computational efficiency. Finally, after obtaining the synthetic trajectory from a trusted third party, the k-means algorithm is used to cluster the trajectory points per unit time based on the Euclidean distance. The generalized trajectories are generated by random combination of the trajectory cluster centers under different timestamps after clustering, and the final trajectory number plus Laplace noise and consistency constraints are published to meet the differential privacy requirements, and the present invention can ensure the trajectory Data privacy while improving the usefulness of trajectory data release.

Description of drawings

1 is a flowchart of a method for protecting trajectory data privacy for an intelligent transportation system provided by an embodiment of the present invention;

2 is a flowchart of differential privacy publishing in the trajectory data privacy protection method for an intelligent transportation system provided by an embodiment of the present invention;

3 is a structural principle block diagram of a trajectory data privacy protection system for an intelligent transportation system provided by an embodiment of the present invention;

FIG. 4 is a structural principle block diagram of an end-to-end deep learning model training provided by an embodiment of the present invention.

Detailed ways

The technical solutions of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. If there is no conflict, the embodiments of the present application and the technical features in the embodiments may be combined with each other.

The term "and/or" in this article is only an association relationship to describe the associated objects, indicating that there can be three kinds of relationships, for example, A and/or B, it can mean that A exists alone, A and B exist at the same time, and A and B exist independently B these three cases. In addition, the character "/" in this article generally indicates that the related objects before and after are an "or" relationship.

Example 1:

As shown in FIG. 1 , this embodiment introduces a trajectory data privacy protection method for an intelligent transportation system, which specifically includes the following steps:

Step 1: Obtain the real-time real trajectory data set of the user;

Step 2: Load the real trajectory data set into a pre-built and trained end-to-end deep learning model to generate a synthetic trajectory data set;

Step 3: Use the k-means clustering algorithm to cluster the trajectory points of the synthetic trajectory data set under each timestamp based on the Euclidean distance, and cluster the trajectory cluster centers under different timestamps after the clustering through random combinations. way to generate generalization trajectories;

Step 4: After adding Laplace noise and consistency constraints to the counting matrix of the generalized trajectory, a differential privacy counting matrix with limited noise is obtained and published.

The training process of the end-to-end deep learning model involved in step 2 provided by the embodiment of the present invention includes the following steps:

Step 1: Collect the historical real trajectory data of different users as the original trajectory data of the training model;

Step 2: standardize the original trajectory data through the centroid to obtain the centroid deviation coordinates of each trajectory point, and obtain the encoded standard original trajectory;

Step 3: Process each of the centroid deviation coordinates into a 64-dimensional vector through a linear rectification function; the calculation formula is as follows:

表示编号为i的轨迹点的64维向量，△lat_i,和△lon_i分别表示编号为i的轨迹点的经度偏差和纬度偏差，f^GPS表示线性整流函数，W_GPS表示轨迹点i的质心偏差坐标(△lat_i,△lon_i)的向量权重。In the formula,

Represents the 64-dimensional vector of the trajectory point numbered i, △lat _i , and △lon _i represent the longitude deviation and latitude deviation of the trajectory point numbered i, respectively, f ^GPS represents the linear rectification function, W _GPS represents the centroid of the trajectory point i Vector weights for the deviation coordinates (Δlat _i , Δlon _i ).

As an embodiment of the present invention, random noise is added to the vector in this step, that is, empty track points are filled into each track, so that each set of vectors in the original track has the same length as the longest track ; It can speed up the model training process and improve the computing efficiency;

Step 4: Use 64 LSTM Cell long short-term memory network units to perform time series prediction on the 64-dimensional vector to obtain synthetic trajectory data, and use the tanh hyperbolic tangent function to decode the centroid deviation coordinates of the synthetic trajectory data and obtain. Decoded synthetic trajectory;

Wherein, the calculation formula for obtaining synthetic trajectory data by using 64 LSTM Cell long short-term memory network units to perform time series prediction on the 64-dimensional vector is as follows:

O=LSTM(T,W _lstm ) (2)

In the formula: T={t ₁ ,t ₂ ,…,t _i ,…,t _maxlength }, T represents the coordinate features of all trajectory points in the original trajectory data, and t _i represents the 64-dimensional vector of the trajectory point numbered i , maxlength represents the longest length of a single trajectory in the original trajectory data; W _lstm represents the weight matrix of the input data; O represents the synthetic trajectory data, where the data contained in O is represented as O={o ₁ ,o ₂ ,...,o _i ,...,o _maxlength }, o _i represents the coordinate composite value of the output of t _i after LSTM processing.

The calculation formula for decoding the centroid deviation coordinates of the synthetic trajectory data using the tanh hyperbolic tangent function is as follows:

表示合成轨迹数据中轨迹点i的质心偏差坐标的解码坐标，

和

分别表示轨迹点i的经度偏差和纬度偏差；W_dGPS表示坐标向量的解码矩阵权重；D^GPS为tanh双曲正切函数。In the formula,

is the decoded coordinate representing the centroid deviation coordinate of the trajectory point i in the synthetic trajectory data,

and

respectively represent the longitude deviation and latitude deviation of the track point i; W _dGPS represents the decoding matrix weight of the coordinate vector; D ^GPS is the tanh hyperbolic tangent function.

Step 5: Calculate the trajectory similarity loss value of the synthetic trajectory through the trajectory loss function in combination with the standard original trajectory;

As an embodiment of the present invention, the step of calculating the trajectory similarity loss of the synthetic trajectory by using the trajectory loss function in combination with the standard original trajectory in step 5 includes:

Step 5.1: perform centroid normalization processing on the synthetic trajectory to obtain the corresponding centroid deviation coordinates of each trajectory point, and obtain an encoded standard synthetic trajectory;

Step 5.2: The centroid deviation coordinates of each of the standard original trajectory and the standard synthetic trajectory are respectively processed into corresponding 64-dimensional vectors through a linear rectification function;

Step 5.3: Combine the two sets of corresponding 64-dimensional vectors and the trajectory loss function, and calculate and train through formula (4) to obtain the trajectory similarity loss value TLoss of the synthetic trajectory:

TLoss=αL _BCE (l ^t ,l ^p )+βL _GPS (t ^t ,t ^p ) (4)

where l ^t and l ^p represent the discriminative labels of the standard original trajectory and standard cooperative trajectory, respectively; t ^t and t ^p represent the 64-dimensional vector of the standard original trajectory and the corresponding standard cooperative trajectory, respectively; L _BCE represents the binary crossover Entropy loss function, L _GPS represents a loss function that uses least squares error to measure the similarity between two trajectories; α and β represent the weights of L _BCE and L _GPS , respectively;

It is further explained that the present invention uses the trajectory loss function to judge the trajectory similarity loss value of the synthetic trajectory obtained by training the original trajectory data proposed in this paper through the end-to-end deep learning model, and different weights can be set according to different scenarios. parameters to change the judgment standard of the similarity loss value, and the scope of application has been enlarged;

Step 6: Perform a binary classification process on the trajectory similarity loss value through a sigmod activation function to obtain a discrimination result of the synthetic trajectory;

If the judgment result is false, then re-input the original trajectory data into the model for repeated training, and use the backpropagation algorithm to solve the model optimization problem to update the network parameters of the model, until the judgment result is true, stop training, and set the The synthetic trajectory output of the discriminant result is true to form a synthetic trajectory dataset.

The method provided by the present invention is further clarified by conducting experiments on several groups of data. First, the centroid deviation coordinates (Δlat _i , Δlon _i ) of each trajectory point are obtained by standardizing the original trajectory data through the centroid, where Δlat _i , and Δ lon _i represents the longitude deviation and latitude deviation of the track point numbered i respectively, and the encoded standard original track is obtained, and the data is as follows:

The centroid deviation coordinates of the three trajectory points in the standard original trajectory T1` are: {(0.07202433, 0.02669937), (0.07295694, 0.02329249), (0.07202433, 0.02669937)};

The centroid deviation coordinates of the three trajectory points in the standard original trajectory T2` are: {(-0.07308236,-0.01193083),(-0.01476175,-0.01333096),(-0.00440719,0.01410267)},

The centroid deviation coordinates of the three trajectory points in the standard original trajectory T3` are: {(0.07202433, 0.02669937), (0.06020328, 0.0145211), (-0.04868138, 0.18410108)},

The centroid deviation coordinates of the three trajectory points in the standard original trajectory T4` are: {(-0.07105364,-0.01321791),(-0.05255251,-0.02247193),(0.06672003,0.04290826)},

The centroid deviation coordinates of the three trajectory points in the standard original trajectory T5` are: {(0.03229337, 0.0198541), (0.03229337, 0.0198541), (0.01348513, 0.00932372);

The centroid deviation coordinates of the three trajectory points in the standard original trajectory T6` are: {(0.04715935, 0.02345322), (0.029829, 0.02150825), (0.03253291, 0.01968764).

构成解码后的合成轨迹，得到结果如下：Next, each centroid deviation coordinate of the above six standard trajectories is processed into a 64-dimensional vector by a linear rectification function; then 64 LSTM Cell long short-term memory network units are used to perform time series prediction on the 64-dimensional vector to obtain synthetic trajectory data. , and use the tanh hyperbolic tangent function to decode the centroid deviation coordinates of the synthetic trajectory data to obtain the corresponding decoded coordinates

The decoded synthetic trajectory is formed, and the results are as follows:

The corresponding decoding coordinates of the three trajectory points in the synthetic trajectory T1 are: {(0.07200033, 0.02660037), (0.07295114, 0.02322249), (0.07201233, 0.02661237)},

The corresponding decoding coordinates of the three trajectory points in the synthetic trajectory T2 are: {(-0.07308326, -0.01193803), (-0.01471675, -0.01330396), (-0.00447019, 0.01412067)},

The corresponding decoding coordinates of the three trajectory points in the synthetic trajectory T3 are: {(0.07204233, 0.02666937), (0.06023028, 0.0142511), (-0.04886138, 0.18411008)},

The corresponding decoding coordinates of the three trajectory points in the synthetic trajectory T4 are: {(-0.07103564, -0.01312791), (-0.05252551, -0.02274193), (0.06670203, 0.04209826)},

The corresponding decoding coordinates of the three trajectory points in the synthetic trajectory T5 are: {(0.03223937, 0.0195841), (0.03223937, 0.0195841), (0.01345813, 0.00933272),

The corresponding decoding coordinates of the three track points in the synthetic track T6 are: {(0.04719535, 0.02354322), (0.029289, 0.02158025), (0.03235291, 0.01986764).

Then, the corresponding similarity loss value is obtained after the operation of step 5 in the method provided by the present invention, and step 6 performs binary classification processing on the similarity loss values of the above six trajectories through the sigmod activation function, and obtains the discrimination results of the above six synthetic trajectories; Among them, when the label value of the discrimination result is 0, it is a false trajectory, and the model needs to be retrained and predicted; when the value of the discrimination result is 1, it is the real trajectory; the synthetic trajectory data set with the value of 1 in the discrimination result is output.

Next, the six synthetic trajectory data sets output above are used to use the k-means clustering algorithm to cluster the trajectory points of the synthetic trajectory data set under each timestamp based on the Euclidean distance, and cluster the different time points after the clustering. The stamped trajectory cluster centers are randomly combined to generate generalized trajectories; by randomly combining the cluster centers, a new trajectory is generated to complement the number of trajectories in the data set n, the number of generalized trajectories is counted, and the count matrix of generalized trajectories is generated C = {c _k |c ₁ ,c ₂ ,…c _n }, where c _k represents the statistics of the k-th generalization trajectory in the corresponding count matrix; the count matrix C is shown in the following table:

generalization trajectory synthetic track true count T11->T21->T31 T1,T2 3 T12->T21->T31 null 2 T11->T21->T32 T3 1 T12->T21->T32 T6 2 T11->T22->T31 null 0 T12->T22->T31 T5 1 T11->T22->T32 null 0 T12->T22->T32 T4 1

并对

添加一致性约束，其

表示对应计数矩阵中第k条泛化轨迹的统计数据。将差分隐私计数矩阵

进行排序得到排序后的计数矩阵S_k＝{s_k|s₁,s₂,…s_n}，结合S_k计算中间变量L_k和Q_k获取轨迹数据结果集

Add Laplace noise with parameter ε to the count matrix C to get the differentially private count matrix

and to

Add consistency constraints, which

Represents statistics corresponding to the k-th generalization trajectory in the count matrix. The differential privacy count matrix

Perform sorting to obtain a sorted count matrix S _k ={s _k |s ₁ ,s ₂ ,...s _n }, and combine _Sk to calculate intermediate variables L _k and Q _k to obtain a result set of trajectory data

In the formula: k∈[1,n], algebraic symbols m, j, z are all natural numbers;

令矩阵{L_k|L₁,L₂,…,L_n}与矩阵

内的元素对应s_k的顺序保持一致获得轨迹数据结果集

其中

表示噪声量受限的差分隐私计数矩阵中第k条泛化轨迹的统计数据；轨迹数据结果集

如下表所示：Sort the elements of the intermediate variable matrix {Q _k |Q ₁ ,Q ₂ ,…,Q _n } to get

Let the matrix {L _k |L ₁ ,L ₂ ,…,L _n } be the same as the matrix

The order of the elements in the corresponding _sk is consistent to obtain the result set of trajectory data

in

Represents the statistics of the k-th generalized trajectory in the differentially private count matrix with limited amount of noise; the result set of trajectory data

As shown in the table below:

generalization trajectory synthetic track Differential Privacy Count T11->T21->T31 T1,T2 3 T12->T21->T31 null 1 T11->T21->T32 T3 2 T12->T21->T32 T6 1 T11->T22->T31 null 1 T12->T22->T31 T5 0 T11->T22->T32 null 0 T12->T22->T32 T4 2

The differential privacy publishing process in the method provided by the present invention is shown in Figure 2. Taking the generalized trajectory T11->T21->T31 as an example, it is obtained by clustering the synthetic trajectories T1 and T2, and the end-to-end deep learning model outputs The synthetic trajectory of T1 and T2 can ensure that the specific coordinate information in T1 and T2 will not be reversely cracked, and then perform noise-limited differential privacy protection on the statistical count of the generalized trajectory T11->T21->T31 through the trajectory publishing mechanism, and further The geostatistical privacy of the trajectory data is protected, and the privacy cracking attack on the statistical information of the trajectory data set can be better resisted.

To sum up, the trajectory data privacy protection method for an intelligent transportation system provided by the embodiments of the present invention uses synthetic trajectories output by an end-to-end deep learning model to replace real trajectories, and these synthetic trajectories can be used for privacy protection by trusted third parties. The substitute of the real trajectory required for processing is used for data sharing and data release; and the method of the present invention utilizes the black box property of machine learning to solve the disadvantage that the existing technology can be reversely cracked to a certain extent; in addition, the present invention Method When the trajectory data set is finally released, a clustering algorithm on the trajectory data can better ensure the privacy of the trajectory data and ensure the high availability of the trajectory data set, and the model is also reusable, when the trajectory data set is updated. When the model is not retrained, the synthetic trajectory data can be obtained only by changing the input, which greatly reduces the time consumed by data protection and improves the calculation efficiency; It can be released according to the sexual constraints to meet the differential privacy requirements, ensure the privacy of trajectory data and improve the usefulness of trajectory data release.

Embodiment 2:

Referring to FIG. 3 and FIG. 4 , an embodiment of the present invention provides a trajectory data privacy protection system for an intelligent transportation system, which can be used to implement the method described in Embodiment 1, specifically including:

The acquisition module is used to acquire the real-time real trajectory data set of the user;

a synthetic trajectory module, used to load the real trajectory data set into a pre-built and trained end-to-end deep learning model to generate a synthetic trajectory data set;

The clustering module is configured to use the k-means clustering algorithm to cluster the trajectory points of the synthetic trajectory data set under each timestamp based on the Euclidean distance, and to pass the clustered trajectory cluster centers under different timestamps after the clustering. Generate generalization trajectories in a random combination;

A noise-adding publishing module is used to add Laplace noise and consistency constraints to the counting matrix of the generalized trajectory to obtain a differentially private counting matrix with limited noise and publish it.

As an embodiment of the present invention, the synthetic trajectory module includes a collection unit; as shown in FIG. 4 , the end-to-end deep learning model includes a trajectory generator and a trajectory discriminator; the trajectory generator includes a first input layer, a first embedding layer, a first LSTM modeling layer, and a first output layer; the trajectory discriminator includes a second input layer, a second embedding layer, a second LSTM modeling layer, and a second output layer; wherein:

The acquisition unit is used to collect the historical real trajectory data of different users as the original trajectory data of the training model;

The first input layer is used to obtain the centroid deviation coordinates of each trajectory point by performing the centroid normalization process on the original trajectory data, and obtain the encoded standard original trajectory;

The first embedding layer is used to use the multilayer perceptron MLP to process each of the centroid deviation coordinates into a 64-dimensional vector through a linear rectification function; and add random noise to the vector, so that each The group vectors all keep the same length as the longest trajectory;

The first LSTM modeling layer is used to perform time series prediction processing on the 64-dimensional vector through 64 LSTM Cell long-term and short-term memory network units to obtain synthetic trajectory data;

The first output layer is used to decode the latitude and longitude deviation of the synthetic trajectory data by using two dense layers Den through the tanh hyperbolic tangent function to obtain the decoded synthetic trajectory;

The second input layer is configured to use the standard original trajectory and the synthesized trajectory as the input data of the trajectory discriminator, and perform centroid normalization on the synthesized trajectory to obtain the corresponding centroid deviation coordinates of each trajectory point, and Obtain the encoded standard synthetic trajectory;

The second embedding layer is used for using the multi-layer perceptron MLP to process the centroid deviation coordinates of each of the standard original trajectory and the standard synthetic trajectory into a corresponding 64-dimensional vector respectively through a linear rectification function;

The second LSTM modeling layer is used to calculate the trajectory similarity loss value through the trajectory loss function using the corresponding 64-dimensional vectors of the two groups through 64 LSTM Cells;

The second output layer is used to use a dense layer Den to perform binary classification processing on the trajectory similarity loss value through the sigmod activation function to obtain the discrimination result of the synthetic trajectory; if the discrimination result is false, the original trajectory The data is re-entered into the trajectory generator for training, and the back-propagation algorithm is used to solve the model optimization problem to update the network parameters in the first LSTM modeling layer in the trajectory generator. The training is stopped until the judgment result is true, and the judgment result is set. For true synthetic trajectory output, form a synthetic trajectory dataset.

The trajectory data privacy protection system for an intelligent transportation system provided by the embodiment of the present invention and the trajectory data privacy protection method for an intelligent transportation system provided in the first embodiment are based on the same technical concept, and can produce the beneficial effects as described in the first embodiment. For the content that is not described in detail in this embodiment, reference may be made to Embodiment 1.

Embodiment three:

An embodiment of the present invention provides a computer-readable storage medium on which a computer program is stored, characterized in that, when the computer program is executed by a processor, steps such as implementing any one of the methods in the first embodiment are implemented.

As will be appreciated by those skilled in the art, the embodiments of the present application may be provided as a method, a system, or a computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It will be understood that each process and/or block in the flowchart illustrations and/or block diagrams, and combinations of processes and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to the processor of a general purpose computer, special purpose computer, embedded processor or other programmable data processing device to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing device produce Means for implementing the functions specified in a flow or flow of a flowchart and/or a block or blocks of a block diagram.

These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory result in an article of manufacture comprising instruction means, the instructions The apparatus implements the functions specified in the flow or flow of the flowcharts and/or the block or blocks of the block diagrams.

These computer program instructions can also be loaded on a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process such that The instructions provide steps for implementing the functions specified in the flow or blocks of the flowcharts and/or the block or blocks of the block diagrams.

The above are only the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the technical principle of the present invention, several improvements and modifications can also be made. These improvements and modifications It should also be regarded as the protection scope of the present invention.

Claims (10)

1. a trajectory data privacy protection method for an intelligent transportation system, characterized in that the method comprises: Obtain real-time real-time trajectory data sets of users; Load the real trajectory dataset into a pre-built and trained end-to-end deep learning model to generate a synthetic trajectory dataset; Use the k-means clustering algorithm to cluster the trajectory points of the synthetic trajectory data set under each timestamp based on the Euclidean distance, and generate the cluster centers of the trajectory clusters under different timestamps after the clustering by random combination. trajectories; After adding Laplace noise and consistency constraints to the counting matrix of the generalized trajectory, a differentially private counting matrix with limited noise is obtained and published. 2. The trajectory data privacy protection method for an intelligent transportation system according to claim 1, wherein the training process of the end-to-end deep learning model comprises the following steps: Collect the historical real trajectory data of different users as the original trajectory data of the training model; The original trajectory data is processed by centroid standardization to obtain the centroid deviation coordinates of each trajectory point, and the encoded standard original trajectory is obtained; Each of the centroid deviation coordinates is processed into a 64-dimensional vector by a linear rectification function; Use 64 LSTM Cell long short-term memory network units to perform time series prediction on the 64-dimensional vector to obtain synthetic trajectory data, and use the tanh hyperbolic tangent function to decode the centroid deviation coordinates of the synthetic trajectory data and obtain the decoded synthetic trajectory; Calculate the trajectory similarity loss value of the synthetic trajectory through a trajectory loss function in combination with the standard original trajectory; Perform binary classification processing on the trajectory similarity loss value through the sigmod activation function to obtain the discrimination result of the synthetic trajectory; If the judgment result is false, then re-input the original trajectory data into the model for repeated training, and use the backpropagation algorithm to solve the model optimization problem to update the network parameters of the model, until the judgment result is true, stop training, and set the The synthetic trajectory output of the discriminant result is true to form a synthetic trajectory dataset. 3. The trajectory data privacy protection method for an intelligent transportation system according to claim 2, wherein after processing each of the centroid deviation coordinates into a 64-dimensional vector by a linear rectification function, the method further comprises: The vectors add random noise so that each set of vectors in the original trajectory remains the same length as the longest trajectory. 4. The trajectory data privacy protection method for an intelligent transportation system according to claim 2, wherein the step of calculating the trajectory similarity loss of the synthetic trajectory by a trajectory loss function in combination with the standard original trajectory comprises: Performing centroid normalization processing on the synthetic trajectory to obtain the corresponding centroid deviation coordinates of each trajectory point, and obtaining an encoded standard synthetic trajectory; The centroid deviation coordinates of each of the standard original trajectory and the standard synthetic trajectory are respectively processed into corresponding 64-dimensional vectors by a linear rectification function; Combined with the corresponding 64-dimensional vectors and the trajectory loss function of the two groups, the trajectory similarity loss value TLoss of the synthetic trajectory is obtained through the calculation and training of formula (1): TLoss=αL _BCE (l ^t ,l ^p )+βL _GPS (t ^t ,t ^p ) (1) where l ^t and l ^p represent the discriminative labels of the standard original trajectory and standard cooperative trajectory, respectively; t ^t and t ^p represent the 64-dimensional vector of the standard original trajectory and the corresponding standard cooperative trajectory, respectively; L _BCE represents the binary crossover Entropy loss function, L _GPS represents a loss function that uses least squares error to measure the similarity between two trajectories; α and β represent the weights of L _BCE and L _GPS , respectively. 5. The trajectory data privacy protection method for an intelligent transportation system according to claim 2, wherein the calculation formula for processing each of the centroid deviation coordinates into a 64-dimensional vector by a linear rectification function is as follows:

In the formula,

Represents the 64-dimensional vector of the trajectory point numbered i, △lat _i , and △lon _i represent the longitude deviation and latitude deviation of the trajectory point numbered i, respectively, f ^GPS represents the linear rectification function, W _GPS represents the centroid of the trajectory point i Vector weights for the deviation coordinates (Δlat _i , Δlon _i ). 6 . The trajectory data privacy protection method for an intelligent transportation system according to claim 5 , wherein the synthetic trajectory is obtained by performing time series prediction on the 64-dimensional vector using 64 LSTMCell long short-term memory network units. 7 . The formula for calculating the data is as follows: O=LSTM(T,W _lstm ) (3) In the formula: T={t ₁ ,t ₂ ,…,t _i ,…,t _maxlength }, T represents the coordinate features of all trajectory points in the original trajectory data, and t _i represents the 64-dimensional vector of the trajectory point numbered i , maxlength represents the longest length of a single trajectory in the original trajectory data; W _lstm represents the weight matrix of the input data; O represents the synthetic trajectory data, where the data contained in O is represented as O={o ₁ ,o ₂ ,...,o _i ,...,o _maxlength }, o _i represents the coordinate composite value of the output of t _i after LSTM processing. 7. The trajectory data privacy protection method for an intelligent transportation system according to claim 6, wherein the calculation formula for decoding the centroid deviation coordinates of the synthetic trajectory data using the tanh hyperbolic tangent function is as follows:

In the formula,

is the decoded coordinate representing the centroid deviation coordinate of the trajectory point i in the synthetic trajectory data,

and

respectively represent the longitude deviation and latitude deviation of the track point i; W _dGPS represents the decoding matrix weight of the coordinate vector; D ^GPS is the tanh hyperbolic tangent function. 8. A trajectory data privacy protection system for an intelligent transportation system, wherein the system comprises: The acquisition module is used to acquire the real-time real trajectory data set of the user; a synthetic trajectory module, used to load the real trajectory data set into a pre-built and trained end-to-end deep learning model to generate a synthetic trajectory data set; The clustering module is configured to use the k-means clustering algorithm to cluster the trajectory points of the synthetic trajectory data set under each timestamp based on the Euclidean distance, and to pass the clustered trajectory cluster centers under different timestamps after the clustering. Generate generalization trajectories in a random combination; A noise-adding publishing module is used to add Laplace noise and consistency constraints to the counting matrix of the generalized trajectory to obtain a differentially private counting matrix with limited noise and publish it. 9 . The trajectory data privacy protection system for an intelligent transportation system according to claim 8 , wherein the synthetic trajectory module comprises a collection unit; the end-to-end deep learning model comprises a trajectory generator and a trajectory discriminator. 10 . The trajectory generator includes a first input layer, a first embedding layer, a first LSTM modeling layer, and a first output layer; the trajectory discriminator includes a second input layer, a second embedding layer, and a second LSTM modeling layer layer, the second output layer; where: The acquisition unit is used to collect the historical real trajectory data of different users as the original trajectory data of the training model; The first input layer is used to obtain the centroid deviation coordinates of each trajectory point by performing the centroid normalization process on the original trajectory data, and obtain the encoded standard original trajectory; The first embedding layer is used to use the multilayer perceptron MLP to process each of the centroid deviation coordinates into a 64-dimensional vector through a linear rectification function; and add random noise to the vector, so that each The group vectors all keep the same length as the longest trajectory; The first LSTM modeling layer is used to perform time series prediction processing on the 64-dimensional vector through 64 LSTM Cell long-term and short-term memory network units to obtain synthetic trajectory data; The first output layer is used to decode the latitude and longitude deviation of the synthetic trajectory data by using two dense layers Den through the tanh hyperbolic tangent function to obtain the decoded synthetic trajectory; The second input layer is configured to use the standard original trajectory and the synthesized trajectory as the input data of the trajectory discriminator, and perform centroid normalization on the synthesized trajectory to obtain the corresponding centroid deviation coordinates of each trajectory point, and Obtain the encoded standard synthetic trajectory; The second embedding layer is used for using the multi-layer perceptron MLP to process the centroid deviation coordinates of each of the standard original trajectory and the standard synthetic trajectory into a corresponding 64-dimensional vector respectively through a linear rectification function; The second LSTM modeling layer is used to calculate the trajectory similarity loss value through the trajectory loss function using the corresponding 64-dimensional vectors of the two groups through 64 LSTM Cells; The second output layer is used to use a dense layer Den to perform binary classification processing on the trajectory similarity loss value through the sigmod activation function to obtain the discrimination result of the synthetic trajectory; if the discrimination result is false, the original trajectory The data is re-entered into the trajectory generator for training, and the back-propagation algorithm is used to solve the model optimization problem to update the network parameters in the first LSTM modeling layer in the trajectory generator. The training is stopped until the judgment result is true, and the judgment result is The real synthetic trajectory output forms the synthetic trajectory dataset. 10. A computer-readable storage medium on which a computer program is stored, characterized in that, when the computer program is executed by a processor, the trajectory for an intelligent transportation system according to any one of claims 1 to 7 is realized The steps of the data privacy protection method.

Images