Self-adaptive differential privacy method for asset positioning data privacy protection
Through the adaptive differential privacy method, a quad-tree index structure and sensitive asset classifier are constructed, and adaptive mesh division and noise processing are used to solve the shortcomings of positioning data privacy protection in the existing technology, and realize the multi-user scenario Efficient privacy protection.
Patent Information
- Application Number
- CN202411055649.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-02
- Publication Date
- 2024-11-12
AI Technical Summary
The existing privacy protection method for positioning data has not been improved in theory and has failed to effectively protect asset positioning data. Especially in multiple user scenarios, the privacy protection needs cannot be met. The traditional method fails to consider the correlation between positioning data, resulting in privacy Risk of leakage.
Adaptive differential privacy method is adopted to build a quad-tree index structure by acquiring and storing asset positioning data, generate sensitive asset positioning data classifiers, use adaptive mesh division and Laplace noise to calculate unit access probability vectors, and design Constrain the optimization problem, weaken the relevance of asset positioning data, and achieve differential privacy protection.
Effectively protect asset positioning data privacy, improve data utility and privacy, realize reliable privacy protection in multi-user scenarios, and significantly reduce the risk of privacy leakage.
Smart Images

Figure CN118940315A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of network information security, and in particular relates to an adaptive differential privacy method for privacy protection of asset location data. Background Art
[0002] With the popularity of smart devices and social applications, location-based service providers can collect a large amount of positioning data, which is widely used in various fields. When enterprises want to obtain asset positioning data and the information hidden in it, they will install terminals on asset equipment to achieve GPS positioning and send the asset positioning data to trusted third-party research institutions for offline analysis. In fact, third-party research institutions are not always trustworthy and may be a potential adversary. In addition, asset positioning data may expose the spatiotemporal behavior of assets and their hidden corporate relationships. Therefore, business adversaries can infer the hidden relationships between assets through the trajectory correlation between different asset trajectory data, and further infer the important locations of assets or the private information of asset owners, thereby causing serious privacy threats. Therefore, asset positioning data urgently needs to be protected.
[0003] Since the rise of research in the field of privacy protection, positioning data privacy protection has been an important topic in this field. However, the existing positioning data privacy protection methods are not perfect in theory, and there is no method that plays an important role in practical applications. There are limitations in the unrelated positioning data privacy protection methods. They fail to consider the correlation between positioning data, so there is still a risk of privacy leakage; and the methods that rely on positioning data correlation are only applicable to scenarios where two motion trajectories are published, and cannot meet the needs of multiple users at the same time. Therefore, these traditional positioning data privacy protection methods perform poorly in practical applications. With the continuous advancement of deep learning technology, new ideas are provided for positioning data privacy protection. Its powerful data processing capabilities can effectively process multi-asset location data in complex social network environments, providing a more secure and reliable privacy protection solution. Summary of the invention
[0004] In order to solve the above problems, the purpose of the present invention is to provide an adaptive differential privacy method for asset location data privacy protection.
[0005] In order to achieve the above object, the adaptive differential privacy method for asset location data privacy protection provided by the present invention comprises the following steps performed in sequence:
[0006] 1) S1 stage of obtaining and storing asset location data: using the terminal device to read all the original asset location data in the selected space and randomly divide the area, and then calculate the area density of each area; then construct a regional allocation function based on the area density boundary, and use the result of the regional allocation function to determine whether each area needs to be further divided into sub-areas; traverse all areas and recursively perform the above operations, gradually constructing a quadtree index structure in which non-leaf nodes are mixed data of non-asset location data and asset location data, and leaf nodes are asset location data; finally, convert the quadtree index structure into a quadtree matrix storing asset location data and store it in the database;
[0007] 2) Obtaining the S2 stage of the sensitive asset location data classifier: Using the quadtree matrix generated in the above step 1), read the original asset location data from it, and manually classify it, mark the original asset location data with strong correlation and sensitive relationship attributes as sensitive asset location data, generate a sensitive asset location data set, use the unlabeled original asset location data set and the sensitive asset location data set to train the generative adversarial network model, and finally obtain the sensitive asset location data classifier;
[0008] 3) S3 stage of processing asset location data: Use the sensitive asset location data classifier obtained in step 2) above to filter out sensitive asset location data from the database, and then use these sensitive asset location data to generate geographic areas, and then use the adaptive grid division method to divide the above geographic areas into top-level units, and then add Laplace noise, and then decide whether to further divide the top-level units into bottom-level units according to the specific scenario;
[0009] 4) S4 stage of calculating the unit access probability vector: According to the top-level unit and bottom-level unit obtained after the grid division in step 3), the unit access probability vector of the sensitive asset trajectory is calculated, and then the correlation between the sensitive asset trajectories is calculated using these unit access probability vectors, so as to measure the strength of the correlation between the sensitive asset trajectories;
[0010] 5) Design the S5 stage of the constrained optimization problem: If the value of the correlation S between the two sensitive asset trajectories is too large, indicating that the two are too strongly related, the privacy protection technology is implemented, and the privacy protection process is converted into a constrained optimization problem for solution, and the final solution of the updated unit access probability vector of the sensitive asset trajectory to be solved is obtained;
[0011] 6) S6 stage to weaken the correlation of asset location data: obtain the top-level unit set with unit access probability greater than 0 from the final solution of the unit access probability vector after the update of the sensitive asset trajectory obtained in step 5) above, and use this set as the candidate asset trajectory set. The availability of the candidate asset trajectory set is evaluated by the utility scoring function to select the asset trajectory with the highest utility. Finally, from the top-level units in the above asset trajectory, the bottom-level units with the largest access density after adaptive partitioning are selected, and the final asset trajectory that satisfies differential privacy and has high practicality is synthesized based on this.
[0012] In step 1), the specific method of obtaining and storing asset location data is as follows:
[0013] First, the terminal device is used to read all the original asset location data in the selected space, and the original asset location data set is formed by all the original asset location data. Then, the whole space is randomly divided into regions using the Rand function, and the number of original asset location data in each region after division is automatically output as the value of area density. Then, the lower bound L and upper bound H of the overall area density are calculated using equations (1) and (2).
[0014]
[0015]
[0016] Among them, D(x i ) represents the area density of region i; represents the mean area density of all regions; min{D(x i )} represents the minimum area density; max{D(x i )} represents the maximum area density;
[0017] Then, using the region partition function P[x i ] Make the following judgments:
[0018]
[0019] If the area density D(x i ) is greater than or equal to the minimum area density min{D(x i )}, but is less than the lower bound L of the overall area density, then the area partition function P[x i ] is -1, indicating that the asset location data is sparse; if the area density D(x i ) is greater than or equal to the lower bound L of the overall area density and less than or equal to the upper bound H of the overall area density, then the area partition function P[x i] is 0, indicating that the asset location data is concentrated; if the area density D(x i ) is greater than the upper bound H of the overall area density but less than the maximum area density max{D(x i )}, then the region partition function P[x i ] has a value of 1, indicating that the asset location data is dense;
[0020] Afterwards, based on the region partition function P[x i ], using the regional partition function f(P[x i ]) to determine whether each area needs to be further divided:
[0021]
[0022] If the region partition function P[x i ] is -1 or 0, then region i remains the original and is not divided; if the region division function P[x i ] is 1, then the region i is divided into 4 sub-regions of equal size; then the above operation is repeated on these 4 sub-regions to obtain the corresponding region partition function P[x i ] until there is no region partition function P[x i ] stops dividing when the value is 1;
[0023] Thus, a quadtree index structure is gradually constructed, in which non-leaf nodes are mixed data of non-asset location data and asset location data, and leaf nodes are asset location data; finally, the quadtree index structure is converted into a quadtree matrix storing asset location data and stored in the database;
[0024] In the quadtree matrix storing asset location data, the number of rows represents the number of index levels of the quadtree, and the number of columns represents the number of nodes corresponding to the asset in a certain tree level, thereby achieving effective storage and management of asset location data.
[0025] In step 2), the specific method of obtaining the sensitive asset location data classifier is as follows:
[0026] Ⅰ) Using the quadtree matrix stored in the database, query, index and obtain the original asset location data, and then perform manual classification. After that, based on background knowledge, the original asset location data with strong correlation and sensitive relationship attributes are marked as sensitive asset location data t1, and all sensitive asset location data t1 constitute the sensitive asset location data set {T1};
[0027] Ⅱ) Traverse the original asset location data set, and take the other original asset location data except the sensitive asset location data t1 as the non-sensitive asset location data t2, and all the non-sensitive asset location data t2 constitute the non-sensitive asset location data set {T2}; then use the sensitive asset location data set {T1} and the non-sensitive asset location data set {T2} to train the generative adversarial network (GAN) model, so that it has the ability to identify sensitive attributes and associations in the unlabeled original asset location data, and use the trained generative adversarial network model as the sensitive asset location data classifier.
[0028] In step 3), the specific method of processing the asset location data is as follows:
[0029] Ⅰ) Using the sensitive asset location data classifier, the sensitive asset location data is filtered out from the database, and the filtered sensitive location data is used to form multiple sensitive asset location data sets D according to the asset name. n
[0030] ={(longitude, latitude, t1),(longitude, latitude, t2),(longitude, latitude, t3),…,(longitude, latitude, t n )…} Then locate all sensitive assets in dataset D n The least squares method is used to generate the geographic area X=
[0031] {(minimum longitude, maximum longitude), (minimum latitude, maximum latitude)}, and then the adaptive grid division method is used to divide the above geographical area into n 2 The top-level unit will be composed of sensitive asset location dataset D n The unitized dataset is defined as the sensitive asset trajectory. And name all the top-level cells above grid
[0032] II) For n 2 top-level units, and calculate the normalized distinguishable number of accesses q(C i ), the formula is as follows:
[0033]
[0034] Among them, C i represents the i-th top-level unit; represents the jth sensitive asset trajectory of the unit pattern; and Represents the jth sensitive asset trajectory In the top unit C i and the number of different positions in the grid G;
[0035] Adding Laplace noise ξ to the normalized distinguishable access count q(C i ), two thresholds X1 and X2 are set, and the number of visits q(C i ), according to the specific scenario, use the thresholds x1 and X2 to decide whether to add the top unit C i Further divided into bottom-level units: If the normalization after noise addition can distinguish the number of visits q(C i ) is greater than the threshold X1, then the top unit C i Divide into 4 bottom-level units; if it is greater than the threshold X2 but less than the threshold X1, the top-level unit C i Divided into 2 bottom-level units; if it is less than the threshold X2, the top-level unit C i No more division.
[0036] In step 4), the specific method of the computing unit accessing the probability vector is as follows:
[0037] For the i-th sensitive asset track Then its unit access probability vector Then the correlation S between sensitive asset trajectories is calculated to measure the strength of the correlation between sensitive asset trajectories:
[0038]
[0039] in, represents the correlation between the i-th and j-th sensitive asset trajectories, and Represents the i-th sensitive asset track and the jth sensitive asset trajectory The unit access probability vector, sim() represents the function of calculating the correlation, and the cosine similarity function is used here to implement it.
[0040] In step 5), the specific method of the design constraint optimization problem is as follows:
[0041] Ⅰ) In order to weaken the correlation between a sensitive asset trajectory and other sensitive asset trajectories, it is necessary to update its unit access probability vector, transform it into a constrained optimization problem and solve it:
[0042]
[0043] in, represents the constrained optimization problem, P c represents the updated unit access probability vector of the sensitive asset trajectory to be solved, Represents the actual unit access probability vector of the current sensitive asset trajectory, A set of unit access probability vectors representing the actual trajectories of other sensitive assets;
[0044] Ⅱ) To solve the constrained optimization problem, a particle swarm optimization algorithm based on an exponential mechanism is used to perform a heuristic search. * The algorithm calculates and obtains the local optimal particle set and the global optimal particle set and continuously compares the results of the two to seek constrained optimization problems The target value is , and Laplace noise ξ is added when selecting the local optimal particle set and the global optimal particle set to satisfy the differential privacy mechanism. When the above constraint optimization problem When the minimum value is reached, the final solution P of the unit access probability vector after the sensitive asset trajectory is updated is obtained c,udpt .
[0045] In step 6), the specific method of weakening the relevance of asset location data is as follows:
[0046] Ⅰ) The final solution P of the unit access probability vector after the above sensitive asset trajectory is updated c,udpt Traverse in the middle and obtain the top unit set U whose unit access probability is greater than 0 c :
[0047] U c ={C i |P c,updt [i]>0} (8)
[0048] II) Set the top unit set U c As a set of candidate asset trajectories, the following utility scoring function is used to evaluate the candidate asset trajectory set and select the asset trajectory with the highest utility. The trajectory unit access probability and trajectory correlation of each time slot need to be considered. The calculation formula is as follows:
[0049]
[0050] Among them, the utility scoring function f(t j ,C i ) represents the calculation of the jth time slot t j The unit access probability and correlation of sensitive asset trajectories are evaluated. represents the jth time slot t j The i-th top unit C i The unit access probability, γ represents the weight parameter, Rank(.) represents the arrangement in ascending order, and denote the jth time slot t j The unit pattern set of the real asset trajectory and the updated synthetic trajectory in ;
[0051] III) After being processed by the utility scoring function, the asset trajectory with the highest utility score is selected, and its asset trajectory T under the unit mode is obtained. c,syn , traverse the asset trajectory T c,syn The top-level units in the S3 stage are selected from each top-level unit, and the bottom-level unit with the largest access density after adaptive partitioning in the S3 stage is selected. Then, the location with the most frequent access in the bottom-level unit is selected as the new asset location data. All the new asset location data constitute the new asset location data set T. syn , thereby obtaining an asset trajectory that has a differential privacy protection mechanism and also satisfies high efficiency and practicality, that is, the asset location data after privacy protection.
[0052] Compared with the prior art, the adaptive differential privacy method for asset location data privacy protection provided by the present invention has the following effects: the synthesized two-dimensional geographic area is discretized by an adaptive grid partitioning method, a unit access probability vector of each trajectory is generated, and Laplace noise is added to ensure the randomness of each top-level unit partition. Then, the unit access probability vector is used to quantify the correlation of the asset location data trajectory, and the privacy of the asset location data is protected by reducing the similarity of the unit access probability vector. Secondly, a constrained optimization problem is designed and solved by a global particle swarm optimization method to obtain an updated unit access probability vector, which weakens the correlation between asset location data. Finally, according to the updated trajectory unit access probability vector and the unit pattern trajectory corresponding to the given location data, it is converted into a positioning pattern trajectory, thereby effectively protecting the privacy of asset location. The experimental results based on the method of the present invention show that compared with the existing methods, this method is more stable, achieves a better trade-off between data utility and privacy, and can protect asset location data more reliably and efficiently. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] Figure 1 Flowchart of the adaptive differential privacy method for asset location data privacy protection provided by the present invention.
[0054] Figure 2 Schematic diagram of the quadtree structure constructed by the present invention.
[0055] Figure 3 This is a schematic diagram comparing the method of the present invention with other methods under the evaluation of the L-MEAN standard.
[0056] Figure 4 This is a schematic diagram comparing the method of the present invention with other methods under de-anonymization attack means.
[0057] Figure 5 This is a schematic diagram comparing the method of the present invention with other methods under reasoning attack means. DETAILED DESCRIPTION
[0058] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but the following embodiments are by no means intended to limit the present invention in any way.
[0059] like Figure 1 As shown, the adaptive differential privacy method for asset location data privacy protection provided by the present invention includes the following steps performed in sequence:
[0060] 1) S1 stage of obtaining and storing asset location data: using the terminal device to read all the original asset location data in the selected space and randomly divide the area, and then calculate the area density of each area; then construct a regional allocation function based on the area density boundary, and use the result of the regional allocation function to determine whether each area needs to be further divided into sub-areas; traverse all areas and recursively perform the above operations, gradually constructing a quadtree index structure in which non-leaf nodes are mixed data of non-asset location data and asset location data, and leaf nodes are asset location data; finally, convert the quadtree index structure into a quadtree matrix storing asset location data and store it in the database;
[0061] The specific method is as follows:
[0062] First, the terminal device is used to read all the original asset location data in the selected space, and the original asset location data set is formed by all the original asset location data. Then, the whole space is randomly divided into regions using the Rand function, and the number of original asset location data in each region after division is automatically output as the value of area density. Then, the lower bound L and upper bound H of the overall area density are calculated using equations (1) and (2).
[0063]
[0064]
[0065] Among them, D(x i ) represents the area density of region i; represents the mean area density of all regions; min{D(x i )} represents the minimum area density; max{D(x i )} represents the maximum area density;
[0066] Then, using the region partition function P[x i ] Make the following judgments:
[0067]
[0068] If the area density D(x i ) is greater than or equal to the minimum area density min{D(xi )}, but is less than the lower bound L of the overall area density, then the area partition function P[x i ] is -1, indicating that the asset location data is sparse; if the area density D(x i ) is greater than or equal to the lower bound L of the overall area density and less than or equal to the upper bound H of the overall area density, then the area partition function P[x i ] is 0, indicating that the asset location data is concentrated; if the area density D(x i ) is greater than the upper bound H of the overall area density but less than the maximum area density max{D(x i )}, then the region partition function P[x i ] has a value of 1, indicating that the asset location data is dense;
[0069] Afterwards, based on the region partition function P[x i ], using the regional partition function f(P[x i ]) to determine whether each area needs to be further divided:
[0070]
[0071] If the region partition function P[x i ] is -1 or 0, then region i remains the original and is not divided; if the region division function P[x i ] is 1, then the region i is divided into 4 sub-regions of equal size; then the above operation is repeated on these 4 sub-regions to obtain the corresponding region partition function P[x i ] until there is no region partition function P[x i ] stops dividing when the value is 1;
[0072] Thus, a quadtree index structure is gradually constructed, in which non-leaf nodes are mixed data of non-asset location data and asset location data, and leaf nodes are asset location data; finally, the quadtree index structure is converted into a quadtree matrix storing asset location data and stored in the database;
[0073] In the quadtree matrix storing asset location data, the number of rows represents the number of index levels of the quadtree, and the number of columns represents the number of nodes corresponding to the asset in a certain tree level, thereby achieving effective storage and management of asset location data;
[0074] 2) Obtaining the S2 stage of the sensitive asset location data classifier: Using the quadtree matrix generated in the above step 1), read the original asset location data from it, and manually classify it, mark the original asset location data with strong correlation and sensitive relationship attributes as sensitive asset location data, generate a sensitive asset location data set, use the unlabeled original asset location data set and the sensitive asset location data set to train the generative adversarial network model, and finally obtain the sensitive asset location data classifier;
[0075] The specific method is as follows:
[0076] Ⅰ) Using the quadtree matrix stored in the database, query, index and obtain the original asset location data, and then perform manual classification. After that, based on background knowledge, the original asset location data with strong correlation and sensitive relationship attributes are marked as sensitive asset location data t1, and all sensitive asset location data t1 constitute the sensitive asset location data set {T1};
[0077] Ⅱ) Traverse the original asset location data set, and take the other original asset location data except the sensitive asset location data t1 as the non-sensitive asset location data t2, and all the non-sensitive asset location data t2 constitute the non-sensitive asset location data set {T2}; then use the sensitive asset location data set {T1} and the non-sensitive asset location data set {T2} to train the generative adversarial network (GAN) model, so that it has the ability to identify sensitive attributes and associations in the unlabeled original asset location data, and use the trained generative adversarial network model as the sensitive asset location data classifier.
[0078] 3) S3 stage of processing asset location data: Use the sensitive asset location data classifier obtained in step 2) above to filter out sensitive asset location data from the database, and then use these sensitive asset location data to generate geographic areas, and then use the adaptive grid division method to divide the above geographic areas into top-level units, and then add Laplace noise, and then decide whether to further divide the top-level units into bottom-level units according to the specific scenario;
[0079] The specific method is as follows:
[0080] Ⅰ) Using the sensitive asset location data classifier, sensitive asset location data is filtered out from the database, and the filtered sensitive location data is used to form multiple sensitive asset location data sets according to the asset name.
[0081] D n ={(longitude, latitude, t1),(longitude, latitude, t2),(longitude, latitude, t3),…,(longitude, latitude, t n )…},
[0082] Then all sensitive assets are located in the dataset Dn The least square method is used to generate the geographical area X = {) minimum longitude, maximum longitude (, (minimum latitude, maximum latitude)}, and then the adaptive grid division method is used to divide the above geographical area into n 2 The top-level unit will be composed of sensitive asset location dataset D n The unitized dataset is defined as the sensitive asset trajectory. And name all the top-level cells above grid
[0083] II) For n 2 top-level units, and calculate the normalized distinguishable number of accesses q(C i ), the formula is as follows:
[0084]
[0085] Among them, C i represents the i-th top-level unit; represents the jth sensitive asset trajectory of the unit pattern; and Represents the jth sensitive asset trajectory In the top unit C i and the number of different positions in the grid C;
[0086] Adding Laplace noise ξ to the normalized distinguishable access count q(C i ), two thresholds X1 and X2 are set, and the number of visits q(C i ), according to the specific scenario, use the thresholds X1 and X2 to decide whether to add the top unit C i Further divided into bottom-level units: If the normalization after noise addition can distinguish the number of visits q(C i ) is greater than the threshold X1, then the top unit C i Divide into 4 bottom-level units; if it is greater than the threshold X2 but less than the threshold X1, the top-level unit C i Divided into 2 bottom-level units; if it is less than the threshold X2, the top-level unit C i No more divisions;
[0087] 4) S4 stage of calculating the unit access probability vector: According to the top-level unit and bottom-level unit obtained after the grid division in step 3), the unit access probability vector of the sensitive asset trajectory is calculated, and then the correlation between the sensitive asset trajectories is calculated using these unit access probability vectors, so as to measure the strength of the correlation between the sensitive asset trajectories; the specific method is as follows:
[0088] For the i-th sensitive asset track Then its unit access probability vector Then the correlation S between sensitive asset trajectories is calculated to measure the strength of the correlation between sensitive asset trajectories:
[0089]
[0090] in, represents the correlation between the i-th and j-th sensitive asset trajectories, and Represents the i-th sensitive asset track and the jth sensitive asset trajectory The unit access probability vector, sim() represents the function of calculating the correlation, and the cosine similarity function is used here to implement it.
[0091] 5) Design the S5 stage of the constrained optimization problem: If the value of the correlation S between the two sensitive asset trajectories is too large, indicating that the two are too strongly related, the privacy protection technology is implemented, and the privacy protection process is converted into a constrained optimization problem for solution, and the final solution of the updated unit access probability vector of the sensitive asset trajectory to be solved is obtained;
[0092] The specific method is as follows:
[0093] Ⅰ) In order to weaken the correlation between a sensitive asset trajectory and other sensitive asset trajectories, it is necessary to update its unit access probability vector, transform it into a constrained optimization problem and solve it:
[0094]
[0095] in, represents the constrained optimization problem, P c represents the updated unit access probability vector of the sensitive asset trajectory to be solved, Represents the actual unit access probability vector of the current sensitive asset trajectory, A set of unit access probability vectors representing the actual trajectories of other sensitive assets;
[0096] Ⅱ) To solve the constrained optimization problem, a particle swarm optimization algorithm based on an exponential mechanism is used to perform a heuristic search. * The algorithm calculates and obtains the local optimal particle set and the global optimal particle set and continuously compares the results of the two to seek constrained optimization problems The target value is set, and Laplace noise ξ (i.e., exponential mechanism) is added when selecting the local optimal particle set and the global optimal particle set to satisfy the differential privacy mechanism. When the above constraint optimization problem When the minimum value is reached, the final solution P of the unit access probability vector after the sensitive asset trajectory is updated is obtained c,udpt .
[0097] 6) S6 stage to weaken the correlation of asset location data: obtain the top-level unit set with unit access probability greater than 0 from the final solution of the unit access probability vector after the update of the sensitive asset trajectory obtained in step 5) above, and use this set as the candidate asset trajectory set. The availability of the candidate asset trajectory set is evaluated by the utility scoring function to select the asset trajectory with the highest utility. Finally, from the top-level units in the above asset trajectory, the bottom-level units with the largest access density after adaptive partitioning are selected, and the final asset trajectory that satisfies differential privacy and has high practicality is synthesized based on this.
[0098] The specific method is as follows:
[0099] Ⅰ) The final solution P of the unit access probability vector after the above sensitive asset trajectory is updated c,udpt Traverse in the middle and obtain the top unit set U whose unit access probability is greater than 0 c :
[0100]
[0101] II) Set the top unit set U c As a set of candidate asset trajectories, the following utility scoring function is used to evaluate the candidate asset trajectory set and select the asset trajectory with the highest utility. The trajectory unit access probability and trajectory correlation of each time slot need to be considered. The calculation formula is as follows:
[0102]
[0103] Among them, the utility scoring function f(t j ,C i ) represents the calculation of the jth time slot t j The unit access probability and correlation of sensitive asset trajectories are evaluated. represents the jth time slot t j The i-th top unit C i The unit access probability, γ represents the weight parameter, Rank(.) represents the arrangement in ascending order, and denote the jth time slot t j The unit pattern set of the real asset trajectory and the updated synthetic trajectory in ;
[0104] III) After being processed by the utility scoring function, the asset trajectory with the highest utility score is selected, and its asset trajectory T under the unit mode is obtained. c,syn , traverse the asset trajectory T s,synFor the top-level units, select the bottom-level units with the highest access density after adaptive partitioning in the S3 stage from each top-level unit. Then, select the position with the most frequent access in the bottom-level units as the new asset positioning data. The new asset positioning data set T is composed of all the new asset positioning data syn , thereby obtaining an asset trajectory with a differential privacy protection mechanism and also meeting high efficiency and practicality, that is, the asset positioning data after privacy protection.
[0105] Figure 2 is a schematic diagram of the quadtree structure constructed in the present invention. The quadtree matrix is constructed according to the result of the above regional partitioning function P[X i . In the figure, the node with the letter S represents the asset positioning data in the interval where the area density of the region is min{D(x i )}<D(x i )<L; the node with the letter M is the asset positioning data in the interval where the area density of the region is L≤D(x i )≤H; the node with the letter F is the data mixed with non-asset positioning data and asset positioning data in the interval where the area density of the region is H≤D(x i )≤max{Dx i}, and it is re-partitioned until all nodes after partitioning are asset positioning data.
[0106] Figure 3 is an experimental comparison chart using the L-MEAN standard in the present invention.
[0107] The present invention uses the L-MEAN standard for evaluation:
[0108] The improved particle swarm optimization algorithm PSO-DP adopted in the present invention, by considering the uncertainty introduced by random perturbation, has proven to have good usability and robustness in different data scenarios in the clustering task. By introducing the matrix dissimilarity distance measure as the evaluation index of the distance loss in the iteration, the L-MEAN value of the distance loss between adjacent iteration rounds is calculated. The smaller the L-MEAN value, the better the clustering effect generated by the iteration. The calculation method of L-MEAN is as follows:
[0109]
[0110] The maximum number of iterations of clustering is represented by the symbol I, and the Hamming metric function is represented by the symbol N. The Hamming metric function N is used to calculate the number of non-zero element indicators in the distance calculation, k∈(0,K].
[0111]
[0112]
[0113] j∈[1,J] represents the number of digital attributes, μ is the weight of the classification feature, l is the feature dimension of the user, x is the user dataset, and g is the cluster center point set.
[0114] Figure 4 (a) and (b) are schematic diagrams showing the comparison between the proposed method and INTER (interpolation method), DPT (differential privacy strategy method), and TGM (differential privacy trajectory generation method) under de-anonymization attack when two different data sets are used. Figure 4 It can be seen that the method ADP-LD of the present invention performs well under deanonymization attacks compared with other methods. The INTER method still ensures data integrity and availability after being attacked by anonymization, but the privacy protection is insufficient and it is easily identified by anonymization attacks. Although both TGM and DPT adopt differential privacy strategies, the computational complexity of DPT is relatively high and requires a large amount of resources for calculation; although the trajectory generated by TGM meets the requirements, it is too different from actual applications. The method of the present invention has obvious advantages over other methods in preventing attack pattern recognition, data practicality, high precision, and privacy protection. It has the lowest success rate under deanonymization attacks and the best effect.
[0115] Figure 5 (a) and (b) are schematic diagrams showing the comparison between the proposed method and INTER (interpolation method), DPT (differential privacy strategy method), and TGM (differential privacy trajectory generation method) under inference attack when two different data sets are used. Figure 5 It can be seen that the method ADP-LD of the present invention performs well under inference attack means compared with other methods. The INTER method is simple and easy to use, but it shows the problems of insufficient privacy protection and weak resistance under inference attack means. The TGM method adopts a differential privacy strategy, and its privacy protection ability has been improved, but the data intensive reading loss is large and the complexity is large. The DPT method focuses on the combination of privacy protection and practicality, but the generated trajectory is still too far away from the actual requirements, resulting in it being unable to be applied in practice. The method of the present invention has obvious advantages over other methods in terms of privacy protection strength, data practicality, high resistance, and strong scalability. It has the lowest success rate under inference attack means and the best effect.
Claims
1. An adaptive differential privacy method for asset location data privacy protection, Features: The adaptive differential privacy method for asset location data privacy protection includes the following steps performed in sequence: 1) S1 stage of obtaining and storing asset location data: using the terminal device to read all the original asset location data in the selected space and randomly divide the area, and then calculate the area density of each area; then construct a regional allocation function based on the area density boundary, and use the result of the regional allocation function to determine whether each area needs to be further divided into sub-areas; traverse all areas and recursively perform the above operations, gradually constructing a quadtree index structure in which non-leaf nodes are mixed data of non-asset location data and asset location data, and leaf nodes are asset location data; finally, convert the quadtree index structure into a quadtree matrix storing asset location data and store it in the database; 2) Obtaining the S2 stage of the sensitive asset location data classifier: Using the quadtree matrix generated in the above step 1), read the original asset location data from it, and manually classify it, mark the original asset location data with strong correlation and sensitive relationship attributes as sensitive asset location data, generate a sensitive asset location data set, use the unlabeled original asset location data set and the sensitive asset location data set to train the generative adversarial network model, and finally obtain the sensitive asset location data classifier; 3) S3 stage of processing asset location data: Use the sensitive asset location data classifier obtained in step 2) above to filter out sensitive asset location data from the database, and then use these sensitive asset location data to generate geographic areas, and then use the adaptive grid division method to divide the above geographic areas into top-level units, and then add Laplace noise, and then decide whether to further divide the top-level units into bottom-level units according to the specific scenario; 4) S4 stage of calculating the unit access probability vector: According to the top-level unit and bottom-level unit obtained after the grid division in step 3), the unit access probability vector of the sensitive asset trajectory is calculated, and then the correlation between the sensitive asset trajectories is calculated using these unit access probability vectors, so as to measure the strength of the correlation between the sensitive asset trajectories; 5) Design the S5 stage of the constrained optimization problem: If the value of the correlation S between the two sensitive asset trajectories is too large, indicating that the two are too strongly related, the privacy protection technology is implemented, and the privacy protection process is converted into a constrained optimization problem for solution, and the final solution of the updated unit access probability vector of the sensitive asset trajectory to be solved is obtained; 6) S6 stage to weaken the correlation of asset location data: obtain the top-level unit set with unit access probability greater than 0 from the final solution of the unit access probability vector after the update of the sensitive asset trajectory obtained in step 5) above, and use this set as the candidate asset trajectory set. The availability of the candidate asset trajectory set is evaluated by the utility scoring function to select the asset trajectory with the highest utility. Finally, from the top-level units in the above asset trajectory, the bottom-level units with the largest access density after adaptive partitioning are selected, and the final asset trajectory that satisfies differential privacy and has high practicality is synthesized based on this.
2. The adaptive differential privacy method for asset location data privacy protection according to claim 1, Features: In step 1), the specific method of obtaining and storing asset location data is as follows: First, the terminal device is used to read all the original asset location data in the selected space, and the original asset location data set is formed by all the original asset location data. Then, the whole space is randomly divided into regions using the Rand function, and the number of original asset location data in each region after division is automatically output as the value of area density. Then, the lower bound L and upper bound H of the overall area density are calculated using equations (1) and (2). Among them, D(x i ) represents the area density of region i; represents the mean area density of all regions; min{D(x i )} represents the minimum area density; max{D(x i )} represents the maximum area density; Then, using the region partition function P[x i ] Make the following judgments: If the area density D(x i ) is greater than or equal to the minimum area density min{D(x i )}, but is less than the lower bound L of the overall area density, then the area partition function P[x i ] is -1, indicating that the asset location data is sparse; if the area density D(x i ) is greater than or equal to the lower bound L of the overall area density and less than or equal to the upper bound H of the overall area density, then the area partition function P[x i ] is 0, indicating that the asset location data is concentrated; if the area density D(x i ) is greater than the upper bound H of the overall area density but less than the maximum area density max{D(x i )}, then the region partition function P[x i ] has a value of 1, indicating that the asset location data is dense; Afterwards, based on the region partition function P[x i ], using the regional partition function f(P[x i ]) to determine whether each area needs to be further divided: If the region partition function P[x i ] is -1 or 0, then region i remains the original and is not divided; if the region division function P[x i ] is 1, then the region i is divided into 4 sub-regions of equal size; then the above operation is repeated on these 4 sub-regions to obtain the corresponding region partition function P[x i ] until there is no region partition function P[x i ] stops dividing when the value is 1; Thus, a quadtree index structure is gradually constructed, in which non-leaf nodes are mixed data of non-asset location data and asset location data, and leaf nodes are asset location data; finally, the quadtree index structure is converted into a quadtree matrix storing asset location data and stored in the database; In the quadtree matrix storing asset location data, the number of rows represents the number of index levels of the quadtree, and the number of columns represents the number of nodes corresponding to the asset in a certain tree level, thereby achieving effective storage and management of asset location data.
3. The adaptive differential privacy method for asset location data privacy protection according to claim 1, Features: In step 2), the specific method of obtaining the sensitive asset location data classifier is as follows: Ⅰ) Using the quadtree matrix stored in the database, we query, index and obtain the original asset location data, and then perform manual classification. After that, based on the background knowledge, we label the original asset location data with strong correlation and sensitive relationship attributes as sensitive asset location data. 1 , located by all sensitive assets data t 1 Constructing sensitive asset location dataset {T 1 }; II) Traverse the original asset location data set and remove sensitive asset location data t 1 Other original asset location data other than the above is regarded as non-sensitive asset location data. 2 , located by all non-sensitive assets 2 Constructing a non-sensitive asset location dataset {T 2 }; Then use sensitive assets to locate the dataset {T 1 } and non-sensitive asset location datasets {T 2 The generative adversarial network model is trained to enable it to identify sensitive attributes and associations in unlabeled raw asset location data, and the trained generative adversarial network model is used as a classifier for sensitive asset location data.
4. The adaptive differential privacy method for asset location data privacy protection according to claim 1, Features: In step 3), the specific method of processing the asset location data is as follows: Ⅰ) Using the sensitive asset location data classifier, sensitive asset location data is filtered out from the database, and the filtered sensitive location data is used to form multiple sensitive asset location data sets according to the asset name. D n ={(longitude, latitude, t 1 ),(longitude, latitude, t 2 ),(longitude, latitude, t 3 ),…,(longitude, latitude, t n )…}, Then all sensitive assets are located in the dataset D n The least square method is used to generate the geographical area X = {(minimum longitude, maximum longitude), (minimum latitude, maximum latitude)}, and then the adaptive grid division method is used to divide the above geographical area into n 2 The top-level unit will be composed of sensitive asset location dataset D n The unitized dataset is defined as the sensitive asset trajectory. And name all the top-level cells above grid II) For n 2 top-level units, and calculate the normalized distinguishable number of accesses q(C i ), the formula is as follows: Among them, C i represents the i-th top-level unit; represents the jth sensitive asset trajectory of the unit pattern; and Represents the jth sensitive asset trajectory In the top unit C i and the number of different positions in the grid G; Adding Laplace noise ξ to the normalized distinguishable access count q(C i ), set two thresholds X 1 and X 2 , based on the normalized distinguishable access times q(C i ), according to the specific scenario, using the threshold X 1 and X 2 To decide whether to add the top cell C i Further divided into bottom-level units: If the normalization after noise addition can distinguish the number of visits q(C i ) is greater than the threshold X 1 , then the top unit C i Divided into 4 bottom units; if it is greater than the threshold X 2 But less than the threshold X 1 , then the top unit C i Divided into 2 bottom units; if less than the threshold X 2 Then the top unit C i No more division.
5. The adaptive differential privacy method for asset location data privacy protection according to claim 1, Features: In step 4), the specific method of the computing unit accessing the probability vector is as follows: For the i-th sensitive asset track Then its unit access probability vector Then the correlation S between sensitive asset trajectories is calculated to measure the strength of the correlation between sensitive asset trajectories: in, represents the correlation between the i-th and j-th sensitive asset trajectories, and Represents the i-th sensitive asset track and the jth sensitive asset trajectory The unit access probability vector, sim() represents the function of calculating the correlation, and the cosine similarity function is used here to implement it.
6. The adaptive differential privacy method for asset location data privacy protection according to claim 1, Features: In step 5), the specific method of the design constraint optimization problem is as follows: Ⅰ) In order to weaken the correlation between a sensitive asset trajectory and other sensitive asset trajectories, it is necessary to update its unit access probability vector, transform it into a constrained optimization problem and solve it: in, represents the constrained optimization problem, P c represents the updated unit access probability vector of the sensitive asset trajectory to be solved, Represents the actual unit access probability vector of the current sensitive asset trajectory, A set of unit access probability vectors representing the actual trajectories of other sensitive assets; Ⅱ) To solve the constrained optimization problem, a particle swarm optimization algorithm based on an exponential mechanism is used to perform a heuristic search. * The algorithm calculates and obtains the local optimal particle set and the global optimal particle set and continuously compares the results of the two to seek constrained optimization problems The target value is , and Laplace noise ξ is added when selecting the local optimal particle set and the global optimal particle set to satisfy the differential privacy mechanism. When the above constraint optimization problem When the minimum value is reached, the final solution P of the unit access probability vector after the sensitive asset trajectory is updated is obtained c,udpt .
7. The adaptive differential privacy method for asset location data privacy protection according to claim 1, Features: In step 6), the specific method of weakening the relevance of asset location data is as follows: Ⅰ) The final solution P of the unit access probability vector after the above sensitive asset trajectory is updated c,udpt Traverse in the middle and obtain the top unit set U whose unit access probability is greater than 0 c : U c ={C i |P c,updt [i]>0} (8) II) Set the top unit set U c As a set of candidate asset trajectories, the following utility scoring function is used to evaluate the candidate asset trajectory set and select the asset trajectory with the highest utility. The trajectory unit access probability and trajectory correlation of each time slot need to be considered. The calculation formula is as follows: Among them, the utility scoring function f(t j ,C i ) represents the calculation of the jth time slot t j The unit access probability and correlation of sensitive asset trajectories are evaluated. represents the jth time slot t j The i-th top unit C i The unit access probability, γ represents the weight parameter, Rank(.) represents the arrangement in ascending order, and denote the jth time slot t j The unit pattern set of the real asset trajectory and the updated synthetic trajectory in ; III) After being processed by the utility scoring function, the asset trajectory with the highest utility score is selected, and its asset trajectory T under the unit mode is obtained. c,syn , traverse the asset trajectory T c,syn The top-level units in the S3 stage are selected from each top-level unit, and the bottom-level unit with the largest access density after adaptive partitioning in the S3 stage is selected. Then, the most frequently accessed location in the bottom-level unit is selected as the new asset location data. All new asset location data constitute the new asset location data set n. syn , thereby obtaining an asset trajectory that has a differential privacy protection mechanism and also satisfies high efficiency and practicality, that is, the asset location data after privacy protection.
Citation Information
Cited By
Sensitive frequent information hiding method, device and system and medium
CN120744987A
A sensitive frequent information hiding method, device, system and medium
CN120744987B
