A multi-feature cross fusion-based differential privacy location protection method and system

Abstract

The application provides a differential privacy location protection method and system based on multi-feature cross fusion. The method comprises the following steps: constructing a semantic frequency sample set based on a semantic stay point set; identifying low-frequency high-sensitivity abnormal semantics through JS divergence based on the semantic frequency sample set, and correcting the initial semantic sensitivity to obtain a semantic sensitivity set; matching the semantic sensitivity of the semantic stay point based on the semantic stay point set and the corrected semantic sensitivity set, and combining the semantic stay point space-time density and stay time to perform second-order cross fusion to construct three groups of second-order cross correlation features; adopting PPO proximal policy optimization to adaptively optimize the fusion weights of the three groups of second-order cross correlation features, and calculating the privacy level corresponding to the semantic stay point; dynamically allocating a differential privacy budget to the semantic stay point based on the privacy level, performing differential noise injection and location generalization based on the privacy budget, and outputting the protected location data. The problems of rigid privacy budget allocation and large data utility loss in the traditional method are solved.

Description

Technical Field

[0001] This invention relates to the fields of network security and location-based services, and in particular to a differential privacy location protection method and system. Background Technology

[0002] With the rapid development of mobile internet, the Internet of Things (IoT), and positioning technology, location-based services (LBS) have been widely applied in navigation, social networking, connected vehicles, smart cities, and other scenarios, becoming an important support for people's daily lives and social operations. While users obtain convenient services, they need to continuously upload trajectory data, which includes highly sensitive information such as home, workplace, medical institutions, and government locations. Once illegally obtained, analyzed, and used for reasoning, this can easily lead to user privacy leaks, tracking of behavioral trajectories, and even risks to personal safety and property security. Therefore, implementing secure, efficient, and personalized privacy protection for location data while ensuring the availability of location services is of significant practical importance.

[0003] Currently, location privacy protection technologies mainly include four categories: k-anonymity and spatial anonymity, location obfuscation, cryptographic encryption, and differential privacy. k-anonymity achieves hiding by generalizing the user's location to an anonymous region, but it relies on the attacker's background knowledge assumptions, making it vulnerable to semantic reasoning attacks and ineffective in sparsely populated areas. Location obfuscation achieves protection by adding noise or offsetting coordinates; it is simple to implement and has low overhead, but lacks rigorous security proofs, and the noise intensity is difficult to adaptively control, easily leading to insufficient protection or data distortion. Encryption methods offer high security, but their computational and communication overhead is high, making them difficult to meet the needs of real-time location services and large-scale data processing on mobile devices. Differential privacy, with its strict mathematical security guarantees and independence from the attacker's background knowledge, has become the mainstream technology for current location privacy protection.

[0004] Despite the significant advantages of differential privacy, existing location protection schemes still have obvious shortcomings: First, privacy budgets are mostly allocated in a fixed manner, unable to be dynamically adjusted according to location sensitivity; second, they rely solely on single features to assess privacy risks, ignoring multi-dimensional correlations such as semantic sensitivity, spatiotemporal density, and dwell time, resulting in insufficiently accurate risk characterization; third, they cannot identify infrequently visited but highly sensitive locations, offering insufficient protection for occasionally visited but highly private venues; and fourth, they struggle to flexibly adjust privacy protection strength according to the scenario, making it difficult to dynamically balance privacy protection strength and data availability, potentially leading to high privacy protection strength but significant location information perturbation and poor usability. Faced with the trend of highly dynamic, fine-grained, and personalized location services, existing technologies struggle to simultaneously meet the requirements of security, usability, and adaptability. Summary of the Invention

[0005] To address the technical problems of existing differential privacy location protection methods, such as rigid privacy budget allocation, single risk assessment, easy leakage of low-frequency high-sensitivity locations, and difficulty in dynamically balancing privacy protection strength and data usability, this invention proposes a differential privacy location protection method and system based on multi-feature cross-fusion. By constructing a multi-feature cross-fusion differential privacy location protection framework, it first quantifies location semantic sensitivity and combines JS divergence to identify low-frequency high-sensitivity semantics. Then, it performs a second-order cross-fusion of semantic sensitivity, spatiotemporal density, and dwell time, and uses reinforcement learning to adaptively optimize feature weights. Finally, it implements differentiated noise injection and location generalization based on privacy levels, achieving refined, adaptive, and hierarchical privacy protection of user location data. This solves the problems of rigid privacy budget allocation, single risk assessment dimensions, easy leakage of low-frequency high-sensitivity locations, and difficulty in balancing privacy strength and data utility in traditional location privacy protection methods, significantly improving privacy protection capabilities while effectively reducing location perturbation errors.

[0006] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0007] A differential privacy location protection method based on multi-feature cross-fusion includes:

[0008] S1: By detecting stop points, obtain the stop points in the user's original GPS trajectory, and combine them with POI data to match semantic types for the stop points, thus constructing a semantic stop point set;

[0009] S2: Construct a semantic frequency sample set based on the semantic dwell point set, quantify the initial semantic sensitivity by clustering the access frequency based on the semantic frequency sample set, identify low-frequency high-sensitivity abnormal semantics based on the semantic frequency sample set by JS divergence, and correct the initial semantic sensitivity based on the low-frequency high-sensitivity abnormal semantics to obtain the corrected semantic sensitivity set.

[0010] S3: Based on the semantic dwell point set and the modified semantic sensitivity set, match the semantic sensitivity of the semantic dwell points, and extract the spatiotemporal density and dwell time of the semantic dwell points. Perform second-order cross-fusion on the semantic sensitivity, spatiotemporal density and dwell time of the semantic dwell points to construct three sets of second-order cross-association features.

[0011] S4: The PPO proximal strategy is used to optimize the fusion weights of three sets of second-order cross-association features adaptively, and the privacy level corresponding to the semantic dwell point is calculated by weighted fusion.

[0012] S5: Based on the privacy level, dynamically allocate differential privacy budgets for semantic dwell points, perform differential noise injection and location generalization based on the privacy budget, and output protected location data.

[0013] Furthermore, by detecting stop points, stop points in the user's original GPS trajectory are obtained, including: traversing the original GPS trajectory points, calculating the Euclidean distance and time interval between any two points, filtering out continuous point regions where the distance is less than a distance threshold and the stop time is greater than a time threshold, and using the center trajectory point of the region as the stop point to obtain the set of stop points. , This is the c-th stop point;

[0014] By combining POI data to match semantic types for stop points, a semantic stop point set is constructed, including:

[0015] Combine POI data with the set of stops Perform semantic type matching, and record the arrival time and departure time of the first trajectory point in the area where the stop point is located as the arrival time and departure time, respectively, to obtain a set of semantic stop points: A single semantic dwell point is defined as a quintuple: ,in, , Longitude and latitude, respectively Let i be the semantic type of the i-th semantic dwell point. For arrival time, This refers to the departure time.

[0016] Furthermore, a semantic frequency sample set is constructed based on the semantic dwell point set, including:

[0017] Extract the access counts of semantic types from the set of semantic dwell points, and perform cumulative calculation on the access counts of the same semantic type to construct a semantic count sample set.

[0018] Furthermore, the initial semantic sensitivity is quantified by clustering access counts based on the semantic frequency sample set, including:

[0019] use Clustering divides the semantic frequency samples into K clusters, sorts them in descending order of the cluster median, and assigns weights as semantic sensitivity: , This represents the sensitivity of the k-th cluster.

[0020] Semantic sensitivity satisfies normalization constraints: , Let be the number of cluster samples in the k-th cluster.

[0021] Furthermore, based on the semantic frequency sample set, low-frequency, high-sensitivity anomaly semantics are identified using JS divergence, including:

[0022] An n×m user-semantic access matrix is ​​constructed based on the semantic frequency sample set, where n is the number of users and m is the number of semantic types. The JS divergence of each semantic type is calculated to obtain the JS divergence set of semantic types. The mean of the JS divergence of all semantic types plus one standard deviation is used as the anomaly threshold. Semantic types that exceed the anomaly threshold are marked as low-frequency, high-sensitivity anomalous semantics.

[0023] Based on the low-order, high-sensitivity anomaly semantic correction, the initial semantic sensitivity is corrected, resulting in a corrected semantic sensitivity set, including:

[0024] For non-abnormal semantics, the initial semantic sensitivity is retained; and the semantic sensitivity of low-frequency, high-sensitivity abnormal semantics is increased to the highest level. Then, all semantic sensitivity values ​​are normalized to complete the semantic sensitivity correction, resulting in a corrected set of semantic sensitivity values. .

[0025] Furthermore, constructing three sets of second-order cross-association features includes: calculating the spatiotemporal density of semantic dwell points as follows: The calculated stay duration is , The preset statistical radius, radius The number of dwell points within the area; the dwell time and spatiotemporal density are normalized respectively to obtain the dwell time feature set. and spatiotemporal density sets ; Calculate the cross-features of spatiotemporal density and semantic sensitivity ; Calculate the cross-features of spatiotemporal density and dwell time: ; Calculate the cross-features of semantic sensitivity and dwell time ;

[0026] The method for calculating the privacy level of semantic dwell points is as follows:

[0027] ;

[0028] in, , , For feature fusion weights, satisfying .

[0029] Furthermore, a proximal strategy is employed to adaptively optimize the fusion weights of the three sets of second-order cross-association features, and to calculate the privacy level corresponding to the semantic dwell point. This includes: using the three sets of second-order cross-association features as state input, using the feature fusion weights as action output, and defining a reward function based on privacy protection strength and data availability. The PPO proximal strategy is then used to optimize and obtain the optimal feature fusion weights. The reward function based on privacy protection strength and data availability is as follows:

[0030] ;

[0031] Among them, Current state Next action The instant rewards obtained To score the strength of privacy protection, Based on the baseline privacy budget, Budget allocated for the current step; To score data availability, The preset root mean square error is the maximum value. , This refers to the set of semantic dwell points used in the optimization of the policy network and the evaluation network in PPO proximal policy optimization. Original position Location protected for privacy; To constrain the indicator function, a constraint penalty term is used in the reinforcement learning reward function to satisfy the constraint. hour If the value is 1, then the condition is not met. -1, These are the preset weighting coefficients.

[0032] Furthermore, differential privacy budgets are dynamically allocated to semantic dwell points based on privacy levels, including:

[0033] Based on the privacy level of each semantic dwell point, a differential privacy budget is adaptively allocated using a preset dynamic privacy budget allocation function; the dynamic privacy budget allocation function is as follows: ,in For the first Differential privacy budget for semantic dwell point allocation For the preset total differential privacy budget, For the first Privacy levels for each semantic dwell point This represents the total number of semantic dwell points.

[0034] Furthermore, differential noise injection and location generalization are performed based on a privacy budget, including:

[0035] Based on a preset basic privacy radius using a privacy budget, the principle is: the higher the privacy level, the larger the allocated differential privacy budget. The smaller the value, the lower the corresponding basic privacy radius. The larger;

[0036] Adding Laplace noise to the basic privacy radius yields the final privacy-preserving radius. , It is Laplace noise. For location query sensitivity, For the set of semantic dwell points A set that differs by only one stop This is a location lookup function;

[0037] Based on privacy level thresholds, semantic dwell points are categorized into different risk levels, and location generalization is performed according to these different risk levels:

[0038] Privacy level High-risk privacy-protected stop points: Using the privacy protection radius as the boundary, search for stop points outside the radius that have the same semantic type as the current stop point, and select one to replace the original stop point coordinates;

[0039] Privacy level Medium-risk stop points: Using a privacy protection radius as the boundary, search for stop points within the radius that have the same semantic type as the current stop point, and select one to replace the original stop point coordinates. , Privacy level threshold;

[0040] Privacy level Low-risk stop points: No location replacement operation is applied; the original stop point's true coordinates are directly preserved.

[0041] A differential privacy location protection system based on multi-feature cross-fusion comprises, in sequence, an end-user layer, a privacy processing layer, and a service application layer; the privacy processing layer includes:

[0042] Semantic Stop Point Extraction Unit: Used to obtain stop points in the user's original GPS trajectory through stop point detection, and to match semantic types for stop points with POI data to construct a semantic stop point set;

[0043] Semantic sensitivity quantification unit: used to construct a semantic frequency sample set based on the set of semantic dwell points, quantify the initial semantic sensitivity by clustering the access frequency based on the semantic frequency sample set, identify low-frequency high-sensitivity abnormal semantics by JS divergence based on the semantic frequency sample set, and correct the initial semantic sensitivity based on the low-frequency high-sensitivity abnormal semantics to obtain the corrected semantic sensitivity set;

[0044] Multi-feature cross-fusion unit: used to match the semantic sensitivity of semantic dwell points based on the set of semantic dwell points and the modified set of semantic sensitivity, and extract the spatiotemporal density and dwell time of semantic dwell points. It performs second-order cross-fusion on the semantic sensitivity, spatiotemporal density and dwell time of semantic dwell points to construct three sets of second-order cross-association features.

[0045] Reinforcement learning weight optimization unit: used to optimize the fusion weights of three sets of second-order cross-association features adaptively using the PPO proximal strategy, and calculate the privacy level corresponding to the semantic dwell point through weighted fusion;

[0046] Differential privacy protection execution unit: used to dynamically allocate differential privacy budgets for semantic dwell points based on privacy levels, perform differential noise injection and location generalization based on the privacy budget, and output protected location data.

[0047] The beneficial effects of this invention are as follows:

[0048] By fusing multi-dimensional features such as semantic sensitivity, spatiotemporal density, and dwell time, and constructing second-order cross-correlation features, a multi-dimensional and accurate characterization of location privacy risks is achieved, solving the problems of single risk assessment and inaccurate risk characterization in existing technologies. By quantifying semantic sensitivity and combining it with JS divergence to identify low-frequency, high-sensitivity anomalous semantics and forcibly increasing their sensitivity weight, key protection of low-frequency, high-sensitivity locations is strengthened, solving the problem of easy leakage of low-frequency, high-sensitivity locations. By adaptively optimizing feature fusion weights through reinforcement learning and dynamically calculating the privacy level of each dwell point, a differentiated and dynamic allocation of differential privacy budget is achieved, solving the problem of rigid privacy budget allocation and inability to adapt to location sensitivity. By implementing a differential privacy protection strategy of hierarchical noise injection and location generalization based on privacy level, a dynamic adaptation between privacy protection strength and data availability is achieved, solving the problem of difficulty in dynamically balancing privacy protection strength and data availability. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a flowchart of the differential privacy location protection method based on multi-feature cross-fusion of the present invention.

[0051] Figure 2 This is a schematic diagram of the semantic sensitivity quantification and abnormal semantic recognition process of the present invention.

[0052] Figure 3 This is a schematic diagram of the differential privacy location protection system based on multi-feature cross-fusion according to the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] This invention proposes a differential privacy location protection method and system based on multi-feature cross-fusion, targeting dynamic location service scenarios. It addresses problems in existing location privacy protection technologies, such as fixed privacy budget allocation, single risk assessment dimensions, insufficient protection of low-frequency, highly sensitive locations, and the difficulty in balancing privacy strength and data utility. The invention employs a three-level distributed architecture, such as... Figure 3 As shown, the system is divided into an end-user layer, a privacy processing layer, and a service application layer. In the end-user layer, mobile terminals use GPS positioning technology to collect user location coordinates in real time. The privacy processing layer exists as a trusted third party and integrates five core processing units: a semantic dwell point extraction unit that combines spatial distance thresholds and time duration thresholds to clean the raw location data and semantically classify dwell points; a semantic sensitivity quantification unit that quantifies semantic sensitivity and identifies abnormal semantics based on the number of visits to dwell points; a multi-feature cross-fusion unit that cross-fusions semantic sensitivity, spatiotemporal density, and dwell time; a reinforcement learning weight optimization unit that optimizes weights through reinforcement learning; and a differential privacy protection execution unit that uses differential privacy algorithms to anonymize sensitive location information. The service application layer consists of various location service providers. Through a hierarchical processing strategy, it ensures user location privacy and security while also considering service application quality, achieving refined, adaptive, and highly available privacy protection of user location trajectories, significantly improving privacy and security levels while maintaining location service quality.

[0055] A differential privacy location protection method based on multi-feature cross-fusion, such as Figure 1 As shown, the steps include:

[0056] S1: By detecting stop points, obtain the stop points in the user's original GPS trajectory, and combine them with POI data to match semantic types for the stop points, thus constructing a semantic stop point set.

[0057] Each user's original GPS track includes a series of track points. j is the user index. This represents the a-th trajectory point, a single trajectory point. , Let a be the longitudes of the trajectory points. Let a be the latitude of the trajectory points. The time for 'a' trajectory points; POI data includes , This represents the b-th POI data point, a single POI data point. , The semantic type for POI data points.

[0058] In this embodiment of the application, the stop points in the user's original GPS trajectory are obtained through stop point detection, including:

[0059] Traverse the original GPS track points, calculate the Euclidean distance and time interval between any two points, filter out continuous point regions where the distance is less than a distance threshold and the dwell time is greater than a time threshold, and use the center track point of the region as the dwell point to obtain the set of dwell points. , Let c be the c-th stop.

[0060] In this embodiment of the application, semantic types are matched for dwell points based on POI data to construct a semantic dwell point set, including:

[0061] POI data Set of stops Perform semantic type matching, and record the arrival time and departure time of the first trajectory point in the area where the stop point is located as the arrival time and departure time, respectively, to obtain a set of semantic stop points: A single semantic dwell point is defined as a quintuple: ,in Let i be the semantic type of the i-th semantic dwell point. For arrival time, For the time of departure, for the user All semantic types it accesses constitute a set: .

[0062] S2: Construct a semantic frequency sample set based on the semantic dwell point set. Quantify the initial semantic sensitivity by clustering access counts based on this sample set. Identify low-frequency, high-sensitivity abnormal semantics using JS divergence based on this sample set. Correct the initial semantic sensitivity based on these low-frequency, high-sensitivity abnormal semantics, resulting in a corrected semantic sensitivity set, such as... Figure 2 As shown.

[0063] In this embodiment of the application, a semantic frequency sample set is constructed based on the semantic dwell point set, including:

[0064] Extract the access counts of semantic types from the set of semantic dwell points, and perform cumulative calculation on the access counts of the same semantic type to construct a semantic count sample set.

[0065] In this embodiment of the application, the initial semantic sensitivity is quantified by clustering access counts based on the semantic frequency sample set, including:

[0066] use Clustering divides the semantic frequency samples into K clusters, sorts them in descending order of the cluster median, and assigns weights as semantic sensitivity: , This represents the sensitivity of the k-th cluster.

[0067] Semantic sensitivity satisfies normalization constraints: , Let be the number of cluster samples in the k-th cluster.

[0068] In this embodiment of the application, low-frequency, highly sensitive anomalous semantics are identified using JS divergence based on a semantic frequency sample set, including:

[0069] Construct an n×m user semantic access matrix based on the semantic frequency sample set, where n is the number of users and m is the number of semantic types. Calculate the JS divergence for each semantic type to obtain the JS divergence set for each semantic type. The mean of the JS divergence of all semantic types plus one standard deviation is used as the anomaly threshold. Semantic types that exceed the anomaly threshold are marked as low-frequency, high-sensitivity anomaly semantics.

[0070] In this embodiment of the application, the initial semantic sensitivity is corrected based on low-frequency, high-sensitivity anomaly semantics, resulting in a corrected set of semantic sensitivity, including:

[0071] For non-abnormal semantics, the initial semantic sensitivity is retained; and the semantic sensitivity of low-frequency, high-sensitivity abnormal semantics is directly increased to the highest level. Then, a normalization process is performed on all semantic sensitivity values ​​to complete the semantic sensitivity correction, resulting in a corrected set of semantic sensitivity values. .

[0072] S3: Based on the semantic dwell point set and the modified semantic sensitivity set, match the semantic sensitivity of the semantic dwell points, and extract the spatiotemporal density and dwell time of the semantic dwell points. Perform second-order cross-fusion on the semantic sensitivity, spatiotemporal density and dwell time of the semantic dwell points to construct three sets of second-order cross-association features.

[0073] In this embodiment of the application, the semantic sensitivity of the semantic dwell points is matched based on the set of semantic dwell points and the modified set of semantic sensitivity, including matching the modified set of semantic sensitivity. With semantic dwell point set Matching is performed to obtain the semantic sensitivity of semantic dwell points, and then normalization is performed to obtain a normalized set of semantic sensitivity of semantic dwell points. .

[0074] In this embodiment of the application, the extraction of the spatiotemporal density and dwell time of semantic dwell points includes:

[0075] Duration of stay: ;

[0076] Spatiotemporal density: ;

[0077] in, The preset statistical radius, radius The number of stops within the area.

[0078] The dwell time and spatiotemporal density are normalized respectively to obtain the dwell time feature set. and spatiotemporal density sets .

[0079] In this embodiment of the application, the semantic sensitivity, spatiotemporal density, and dwell time of the semantic dwell point are subjected to second-order cross-fusion to construct three sets of second-order cross-association features, including:

[0080] Construct three sets of second-order cross features:

[0081] The cross-feature of spatiotemporal density and semantic sensitivity is: ;

[0082] The cross-feature of spatiotemporal density and dwell time is as follows: ;

[0083] The cross-feature of semantic sensitivity and dwell time is: .

[0084] Semantic dwell point The privacy level is defined as:

[0085] ;

[0086] in, , , The feature fusion weights satisfy the condition that sum to one.

[0087] S4: The fusion weights of the three sets of second-order cross-association features are adaptively optimized using Proximal Policy Optimization (PPO) reinforcement learning, and the privacy level corresponding to the semantic dwell point is calculated by weighted fusion.

[0088] The three sets of second-order cross-correlation features obtained in step S3 are used as the state input of the policy network, and the feature fusion weights are used as the action output of the policy network. The optimal feature fusion weights are obtained by using the PPO proximal policy optimization.

[0089] Both the policy network and the judge network employ fully connected neural networks;

[0090] Using the pruning loss optimized by the near-end strategy as the policy loss:

[0091]

[0092] in, These are the weight parameters of the policy network. The importance sampling ratio, For the dominant function, For the trimming operator, To trim hyperparameters, This represents the expected value of the samples across all trajectory samples collected in this round.

[0093] The mean squared error between the value score and the reward function is used as the value loss.

[0094]

[0095] in, For the weight parameters of the value network, For value network prediction, The true cumulative reward for the state at time t. This is the discount factor.

[0096] The reward function is a weighted sum of privacy protection strength and data availability, mathematically defined as follows:

[0097] ;

[0098] in, Current state Next action The instant rewards obtained To score the strength of privacy protection, ,in Based on the baseline privacy budget, The budget allocated to the current step; the larger this value, the stronger the privacy protection. The data availability score is calculated using the root mean square error (RMSE). )measure:

[0099] ;

[0100] in, The preset root mean square error is the maximum value. The calculation formula is: ,in This refers to the set of semantic dwell points used in the optimization of the policy network and evaluation network during PPO proximal policy optimization, representing the underlying data of the PPO agent's interaction environment. Original position Location protected for privacy; To constrain the indicator function, a constraint penalty term is used in the reinforcement learning reward function to satisfy the constraint. hour If the value is 1, then the condition is not met. =-1; These are preset weighting coefficients used to balance the proportions of privacy, utility, and constraint penalties in the reinforcement learning reward function.

[0101] The policy network and the evaluation network are jointly optimized based on policy loss and value loss to obtain a trained policy network. The trained policy network is then used to predict the three sets of second-order cross-correlation features obtained in step S4 to obtain the optimal feature fusion weights.

[0102] Furthermore, the privacy level of semantic dwell points is calculated based on the optimal feature fusion weights. .

[0103] S5: Based on the privacy level, dynamically allocate differential privacy budgets for semantic dwell points, perform differential noise injection and location generalization based on the privacy budget, and output protected location data.

[0104] In this embodiment, the differential privacy budget is dynamically allocated based on the privacy level of each semantic dwell point, including: adaptively allocating the differential privacy budget according to the privacy level of each semantic dwell point using a preset dynamic privacy budget allocation function, wherein the dynamic privacy budget allocation function is: ,in For the first Differential privacy budget for semantic dwell point allocation For the preset total differential privacy budget, For the first Privacy levels for each semantic dwell point This represents the total number of semantic dwell points.

[0105] In this embodiment of the application, the basic privacy radius is set based on the privacy budget, including:

[0106] Based on the allocated privacy budget Determine the basic privacy radius corresponding to this stop point. Among them, the higher the privacy level, the larger the allocated privacy budget. The smaller the value, the lower the corresponding basic privacy radius. The larger the area, the stronger the privacy protection for high-risk locations.

[0107] In this embodiment, Laplace noise is added to the basic privacy radius to obtain the final privacy-protected radius. This includes: sensitivity to location queries. This generates noise that follows a Laplace distribution. The location query function is defined as follows: Input semantic dwell point dataset Output the set of two-dimensional coordinates of all dwell points. In the differential privacy framework, global sensitivity... The definition is: for any two adjacent datasets that differ by only one stop point and The query function outputs the maximum change in value. Its mathematical expression is: This value represents the maximum change in the coordinate set caused by adding or deleting a stop point, and determines the intensity of Laplace noise required to mask the impact of a single record. The noise generation formula is: in: Let be the noise value that follows a Laplace distribution, and its scaling parameter is . The probability density function of the Laplace distribution is: The noise is superimposed on the basic privacy radius to generate the final privacy protection radius centered on the actual point of stay. ;

[0108] In this embodiment, semantic dwell points are classified into different risk levels based on privacy level thresholds, and location generalization is performed based on different risk levels:

[0109] Privacy level High-risk privacy-protected stop points: Using the privacy protection radius as the boundary, search for stop points outside the radius that have the same semantic type as the current stop point, and select one to replace the original stop point coordinates;

[0110] Privacy level Medium-risk stop points: Using a privacy protection radius as the boundary, search for stop points within the radius that have the same semantic type as the current stop point, and select one to replace the original stop point coordinates. , Privacy level threshold;

[0111] Privacy level Low-risk stop points: No location replacement operation is applied; the original stop point's true coordinates are directly retained as the output.

[0112] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A differential privacy location protection method based on multi-feature cross-fusion, characterized in that, include: S1: By detecting stop points, obtain the stop points in the user's original GPS trajectory, and combine them with POI data to match semantic types for the stop points, thus constructing a semantic stop point set; S2: Construct a semantic frequency sample set based on the semantic dwell point set, quantify the initial semantic sensitivity by clustering the access frequency based on the semantic frequency sample set, identify low-frequency high-sensitivity abnormal semantics based on the semantic frequency sample set by JS divergence, and correct the initial semantic sensitivity based on the low-frequency high-sensitivity abnormal semantics to obtain the corrected semantic sensitivity set. S3: Based on the semantic dwell point set and the modified semantic sensitivity set, match the semantic sensitivity of the semantic dwell points, and extract the spatiotemporal density and dwell time of the semantic dwell points. Perform second-order cross-fusion on the semantic sensitivity, spatiotemporal density and dwell time of the semantic dwell points to construct three sets of second-order cross-association features. S4: The PPO proximal strategy is used to optimize the fusion weights of three sets of second-order cross-association features adaptively, and the privacy level corresponding to the semantic dwell point is calculated by weighted fusion. S5: Based on the privacy level, dynamically allocate differential privacy budgets for semantic dwell points, perform differential noise injection and location generalization based on the privacy budget, and output protected location data.

2. The differential privacy location protection method based on multi-feature cross-fusion according to claim 1, characterized in that, By detecting stop points, the stop points in the user's original GPS trajectory are obtained. This includes: traversing the original GPS trajectory points, calculating the Euclidean distance and time interval between any two points, filtering out continuous point regions where the distance is less than a distance threshold and the stop time is greater than a time threshold, and using the center trajectory point of the region as the stop point to obtain a set of stop points. , This is the c-th stop point; By combining POI data to match semantic types for stop points, a semantic stop point set is constructed, including: Combine POI data with the set of stops Perform semantic type matching, and record the arrival and departure times of the first trajectory point and the last trajectory point in the area where the stop point is located, respectively, to obtain a set of semantic stop points: A single semantic dwell point is defined as a quintuple: ,in, , They are longitude and latitude, respectively. Let i be the semantic type of the i-th semantic dwell point. For arrival time, This refers to the departure time.

3. The differential privacy location protection method based on multi-feature cross-fusion according to claim 2, characterized in that, A semantic frequency sample set is constructed based on the set of semantic dwell points, including: Extract the access counts of semantic types from the set of semantic dwell points, and perform cumulative calculation on the access counts of the same semantic type to construct a semantic count sample set.

4. The differential privacy location protection method based on multi-feature cross-fusion according to claim 3, characterized in that, Initial semantic sensitivity is quantified by clustering access counts based on a semantic frequency sample set, including: use Clustering divides the semantic frequency samples into K clusters, sorts them in descending order of the cluster median, and assigns weights as semantic sensitivity: , This represents the sensitivity of the k-th cluster; Semantic sensitivity satisfies normalization constraints: , Let be the number of cluster samples in the k-th cluster.

5. The differential privacy location protection method based on multi-feature cross-fusion according to any one of claims 1-5, characterized in that, Based on a semantic frequency sample set, low-frequency, high-sensitivity anomaly semantics are identified using JS divergence, including: An n×m user semantic access matrix is ​​constructed based on the semantic frequency sample set, where n is the number of users and m is the number of semantic types. The JS divergence of each semantic type is calculated to obtain the JS divergence set of semantic types. The mean of the JS divergence of all semantic types plus one standard deviation is used as the anomaly threshold. Semantic types that exceed the anomaly threshold are marked as low-frequency, high-sensitivity anomalous semantics. Based on the low-order, high-sensitivity anomaly semantic correction, the initial semantic sensitivity is corrected, resulting in a corrected semantic sensitivity set, including: For non-abnormal semantics, the initial semantic sensitivity is retained; and the semantic sensitivity of low-frequency, high-sensitivity abnormal semantics is increased to the highest level. Then, all semantic sensitivity values ​​are normalized to complete the semantic sensitivity correction, resulting in a corrected set of semantic sensitivity values. .

6. The differential privacy location protection method based on multi-feature cross-fusion according to claim 5, characterized in that, Constructing three sets of second-order cross-association features includes: calculating the spatiotemporal density of semantic dwell points. The calculated stay duration is , The preset statistical radius, radius The number of dwell points within the area; the dwell time and spatiotemporal density are normalized respectively to obtain the dwell time feature set. and spatiotemporal density sets ; Calculate the cross-features of spatiotemporal density and semantic sensitivity ; Calculate the cross-features of spatiotemporal density and dwell time: ; Calculate the cross-features of semantic sensitivity and dwell time ; The method for calculating the privacy level of semantic dwell points is as follows: ； in, , , For feature fusion weights, satisfying .

7. The differential privacy location protection method based on multi-feature cross-fusion according to claim 6, characterized in that, A proximal optimization strategy is employed to adaptively optimize the fusion weights of three sets of second-order cross-association features, calculating the privacy level corresponding to the semantic dwell point. This includes: using the three sets of second-order cross-association features as state input, using the feature fusion weights as action output, and defining a reward function based on privacy protection strength and data availability. The PPO proximal optimization strategy is used to obtain the optimal feature fusion weights. The reward function based on privacy protection strength and data availability is as follows: ； Among them, Current state Next action The instant rewards obtained To score the strength of privacy protection, Based on the baseline privacy budget, Budget allocated for the current step; To score data availability, The preset root mean square error is the maximum value. , This refers to the set of semantic dwell points used in the optimization of the policy network and the evaluation network in PPO proximal policy optimization. Original position Location protected for privacy; To constrain the indicator function, a constraint penalty term is used in the reinforcement learning reward function to satisfy the constraint. hour If the value is 1, then the condition is not met. -1, These are the preset weighting coefficients.

8. The differential privacy location protection method based on multi-feature cross-fusion according to claim 7, characterized in that, Based on privacy levels, differential privacy budgets are dynamically allocated to semantic dwell points, including: Based on the privacy level of each semantic dwell point, a differential privacy budget is adaptively allocated using a preset dynamic privacy budget allocation function; the dynamic privacy budget allocation function is as follows: ,in For the first Differential privacy budget allocated to each stop point. For the preset total differential privacy budget, For the first Privacy levels for each semantic dwell point. This represents the total number of semantic dwell points.

9. The differential privacy location protection method based on multi-feature cross-fusion according to claim 8, characterized in that, Differentiated noise injection and location generalization are performed based on a privacy budget, including: Based on a preset basic privacy radius using a privacy budget, the principle is: the higher the privacy level, the larger the allocated differential privacy budget. The smaller the value, the lower the corresponding basic privacy radius. The larger; Adding Laplace noise to the basic privacy radius yields the final privacy-preserving radius. , It is Laplace noise. For location query sensitivity, For the set of semantic dwell points A set that differs by only one stop This is a location lookup function; Based on privacy level thresholds, semantic dwell points are categorized into different risk levels, and location generalization is performed according to these different risk levels: Privacy level High-risk privacy-protected stop points: Using the privacy protection radius as the boundary, search for stop points outside the radius that have the same semantic type as the current stop point, and select one to replace the original stop point coordinates; Privacy level Medium-risk stop points: Using a privacy protection radius as the boundary, search for stop points within the radius that have the same semantic type as the current stop point, and select one to replace the original stop point coordinates. , Privacy level threshold; Privacy level Low-risk stop points: No location replacement operation is applied; the original stop point's true coordinates are directly preserved.

10. A differential privacy location protection system based on multi-feature cross-fusion, characterized in that, It consists of, in order, the end-user layer, the privacy processing layer, and the service application layer; the privacy processing layer includes: Semantic Stop Point Extraction Unit: Used to obtain stop points in the user's original GPS trajectory through stop point detection, and to match semantic types for stop points with POI data to construct a semantic stop point set; Semantic sensitivity quantification unit: used to construct a semantic frequency sample set based on the set of semantic dwell points, quantify the initial semantic sensitivity by clustering the access frequency based on the semantic frequency sample set, identify low-frequency high-sensitivity abnormal semantics by JS divergence based on the semantic frequency sample set, and correct the initial semantic sensitivity based on the low-frequency high-sensitivity abnormal semantics to obtain the corrected semantic sensitivity set; Multi-feature cross-fusion unit: used to match the semantic sensitivity of semantic dwell points based on the set of semantic dwell points and the modified set of semantic sensitivity, and extract the spatiotemporal density and dwell time of semantic dwell points. It performs second-order cross-fusion on the semantic sensitivity, spatiotemporal density and dwell time of semantic dwell points to construct three sets of second-order cross-association features. Reinforcement learning weight optimization unit: used to optimize the fusion weights of three sets of second-order cross-association features adaptively using the PPO proximal strategy, and calculate the privacy level corresponding to the semantic dwell point through weighted fusion; Differential privacy protection execution unit: used to dynamically allocate differential privacy budgets for semantic dwell points based on privacy levels, perform differential noise injection and location generalization based on the privacy budget, and output protected location data.

Images