A vehicle trajectory privacy protection method based on semantic perception and differential privacy

Abstract

The application discloses a vehicle trajectory privacy protection method based on semantic perception and differential privacy. First, the semantic of interest points of road sections and their areas is extracted, and a road section representation vector is generated by using a graph attention network; second, sensitive positions are identified by combining time and space, semantic weight and access frequency, and differential privacy budget is dynamically allocated; then, a comprehensive utility function is constructed to screen target confusion road sections, and sampling points are generated by using a Laplace mechanism to replace original sensitive positions; finally, local space reconstruction and global time updating are carried out based on path planning. The application can resist background attacks, protect privacy, maximize the continuity of the trajectory of the road network, and improve data availability.

Description

Technical Field

[0001] This application relates to the field of vehicle trajectory privacy protection and data security technology, and in particular to a vehicle trajectory privacy protection method based on semantic awareness and differential privacy. Background Technology

[0002] With the deepening of a new round of technological revolution and industrial transformation, the Internet of Vehicles (IoV), as a product of the deep integration of the automotive industry with fields such as electronic information and transportation, is undergoing rapid evolution. Under the trend of "Software Defined Vehicles (SDV)," the content of in-vehicle services has expanded from basic navigation and positioning to diversified areas such as personalized intelligent recommendations. These location-based services (LBS), while providing convenience to users, have also collected and accumulated an unprecedented scale of vehicle trajectory data. This data contains rich patterns of group movement and individual behavior, demonstrating enormous potential for improvement in alleviating traffic congestion, optimizing urban spatial layout, and providing high-value business insights.

[0003] However, behind the data dividend lies a serious challenge of privacy compliance. In an environment of increasingly stringent global data security legislation, the direct sharing or publication of raw trajectory data is explicitly prohibited by law. Relevant data protection regulations clearly require that personal trajectory information undergo rigorous de-identification processing before being circulated to third parties or used for commercial analysis, ensuring that individual identities cannot be identified. How to legally and compliantly explore and utilize the value of trajectory data has become a common research challenge for both academia and industry.

[0004] The most prominent risk currently facing the release of tracking data is the deanonymization of tracking information, meaning the identity of the person whose tracking data was previously anonymous can be re-identified. Tracking data possesses highly personalized characteristics, uniquely identifying users like fingerprints. Malicious attackers only need minimal background knowledge to accurately pinpoint specific individuals within anonymized datasets. This leakage of tracking location privacy makes it extremely easy to reverse engineer highly sensitive information such as a user's real address and social relationships, potentially triggering serious personal or property security crises.

[0005] Faced with this threat, Differential Privacy (DP), with its rigorous mathematical framework and resistance to background knowledge attacks, has become the mainstream solution for trajectory privacy protection. However, traditional differential privacy mechanisms can significantly reduce data availability in practical applications. Due to the use of indiscriminate global noise injection strategies, existing methods often ignore the objective differences in privacy sensitivity across different geographical locations. This "one-size-fits-all" perturbation not only leads to excessive obfuscation of regular travel routes but also results in insufficient protection for truly high-risk sensitive locations, ultimately causing severe distortion of the trajectory's spatiotemporal geometry and rendering it useless for subsequent analysis.

[0006] To address the aforementioned shortcomings, semantically aware differential privacy schemes offer a novel approach. However, most current strategies remain at a superficial application stage, lacking robust privacy protection due to their fragmented nature and detachment from physical spatial connectivity. This invention focuses on resolving the core contradiction between high privacy protection and high data utility in vehicle trajectory publishing. It aims to explore and construct an adaptive differential privacy protection framework that deeply integrates semantics and road network topology features, maximizing the preservation of the true continuity and deep semantic value of vehicle trajectories. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a vehicle trajectory privacy protection method based on semantic awareness and differential privacy, so as to solve the problem of reduced data availability caused by the application of differential privacy in trajectory data publishing scenarios in existing methods.

[0008] To achieve the above-mentioned objectives, the trajectory privacy protection method based on sensitive semantic location replacement in the Internet of Vehicles of the present invention includes the following steps:

[0009] S1. Data Input: Input the vehicle trajectory dataset and the corresponding city road network data.

[0010] S2. Road Segment Extraction: Based on the road network topology, the input urban road network is divided into regularized sub-segments. Physical nodes such as intersections and road junctions are used as segmentation boundaries, and the sub-segment length is determined according to a preset sub-segment length threshold L and a floating range. Perform length-constrained slicing to generate a set of road network sub-segments with length constraints in the range [L−∆L,L+∆L].

[0011] S3. Semantic Information Mining: For each sub-segment, obtain the point of interest data in its corresponding area through the geographic information API, extract the type and density information of the point of interest, and form a set of semantic features of the road segment with the inherent semantic attributes of the road segment.

[0012] S4. Road segment representation generation: Based on the structural characteristics and semantic features of road segments, a transition probability matrix is ​​first constructed by statistically analyzing the number of transitions of sub-road segments in the trajectory data. Then, this matrix is ​​used as prior information to embed into the attention calculation mechanism of the Graph Attention Network (GAT) to strengthen the expression of relationships between road segments. Finally, through self-supervised training of road segments, a road segment representation vector that integrates physical structure, semantic information and trajectory behavior patterns is generated.

[0013] S5. Sensitive Location Identification: Based on Time Threshold Spatial threshold Potential dwell points are extracted from a continuous sampling sequence. By combining dwell time, semantic weights of points of interest, and access frequency, the sensitivity of each dwell point is calculated, and a sensitivity threshold is set to determine sensitive locations in the trajectory. The sensitivity calculation formula is as follows: ,

[0014] in, For the stop The actual length of stay For the stop The semantic weights are assigned custom values ​​based on the matched interest types. This represents the frequency with which users access this area.

[0015] S6. Privacy Budget Allocation: For the identified sensitive location i, an exponential decay function is used to allocate the privacy budget based on the sensitivity obtained in step S4. Calculate the total privacy budget specific to this location. The formula is:

[0016] Based on the serial combination theorem of differential privacy, the total budget is... According to the preset ratio Divide into road segment privacy budget Location privacy budget ;in A screening mechanism is assigned to target confusing road segments; , and assign Laplace sampling for the dwell points.

[0017] S7. Confusion Segment Filtering: Determine the number of target candidate road segments based on the sensitivity of sensitive locations. Starting from the road segment containing the sensitive location, expand outwards using breadth-first search based on the road network topology. Construct a candidate road segment set by combining semantic similarity and topological reachability. Construct a utility function that combines semantic and spatial structural distance. Based on the differential privacy index mechanism and the allocated road segment privacy budget Calculate the selection probability of each road segment in the candidate road segment set, and select one as the target confusion road segment.

[0018] S8. Sensitive Location Replacement: Locate the position with the highest semantic similarity on the target confusing road segment and use it as the new stop center. Count the original trajectory points. A polar coordinate system is constructed with the new center of stay as the origin. A generalized sampling point satisfying the Laplace distribution is generated in the two-dimensional plane using the gamma distribution equivalent sampling strategy. Coordinates:

[0019] in, The drift distance, Budget for location privacy, Let be a uniformly distributed random variable. The drift direction angle is random.

[0020] S9. Trajectory Spatiotemporal Reconstruction: First, obtain sensitive location points and new stopping centers, and within the midpoint range between them and the preceding and following stopping points, extract the preceding and following splicing points that satisfy the equidistant geometric constraints. Second, perform online local path reconstruction, remove the original trajectory between the preceding and following splicing points, and use the preceding splicing points, new stopping centers, and following splicing points as the start point, waypoint, and end point, respectively. Call the driving route planning API of the online map open platform to obtain the trajectory node sequence that fits the real road topology. Finally, perform global time update, adopting a deduction mechanism based on trajectory departure time, combining the physical speed limit of the road network to calculate the driving time, and inheriting the actual stopping time of the original sensitive points, sequentially accumulating and assigning timestamps to each node. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0022] Figure 1 A flowchart illustrating the vehicle trajectory privacy protection method based on semantic awareness and differential privacy provided in this application embodiment;

[0023] Figure 2 The algorithm flowchart for the specific slicing rules provided in the embodiments of this application;

[0024] Figure 3 A schematic diagram illustrating the process of learning self-supervised road segment representation vectors provided in this application embodiment;

[0025] Figure 4 A schematic diagram illustrating the location of splicing points in trajectory spatiotemporal reconstruction provided in an embodiment of this application;

[0026] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0027] The specific embodiments of the present invention will now be described with reference to the accompanying drawings, so that those skilled in the art can better understand the present invention.

[0028] Figure 1 This is a flowchart illustrating a specific implementation of the vehicle trajectory privacy protection method based on semantic awareness and differential privacy of the present invention. Figure 1 As shown, the specific steps of the present invention include:

[0029] S1. Data Input: Input the vehicle trajectory dataset and obtain the corresponding city road network data from the road network open platform.

[0030] S2. Road Segment Extraction: Based on the road network topology, the input urban road network is divided into regularized sub-segments. Physical nodes such as intersections and road junctions are used as segmentation boundaries. Length-constrained slicing is performed according to a preset sub-segment length threshold L and a floating range ΔL, generating a set of road network sub-segments with length constraints within the range [L-ΔL, L+ΔL]. The specific slicing rule algorithm is as follows: Figure 2 As shown.

[0031] S3. Semantic Information Mining: For each sub-segment, obtain the translation distances to its left and right sides using the geographic information API. The data of points of interest within the region are used to extract the type and density information of the points of interest as external semantic features. The inherent semantic attributes of the road segments (including road type, road segment length, number of lanes, maximum driving speed, and in-degree and out-degree in the road network map) are integrated to finally form a complete set of road segment semantic features describing the attributes of the sub-road segments.

[0032] S4. Road Segment Representation Generation: Based on the structural characteristics and semantic features of road segments, a transition probability matrix is ​​first constructed by statistically analyzing the number of transitions in sub-road segments within the trajectory data. This matrix is ​​then used as prior information to embed into the attention mechanism of a graph attention network to strengthen the expression of relationships between road segments. Finally, through self-supervised training of road segments, a road segment representation vector integrating physical structure, semantic information, and trajectory behavior patterns is generated. Figure 3 The specific implementation example shown has the following steps:

[0033] S41. Input road segment features: Based on the original road network and trajectory data, extract the semantic and structural attributes of each road segment, and count the number of transitions between road segments in the trajectory, and use these as additional feature inputs.

[0034] S42. Graph Structure Construction: Based on the road network topology and road segment transfer relationships, construct a road segment-level graph structure to form nodes and their adjacency relationships. Nodes represent road segments, and edges represent the connection relationships between road segments, including physical connections and trajectory transfer probabilities.

[0035] S43. Attention Weight Calculation: Input the graph structure from S42 into the graph attention network, and calculate the attention weights between nodes by introducing structural relationships and transition priors.

[0036] S44. Masking Recovery Task: Under the self-supervised training framework, random masking is performed on some road segment features, enabling the model to recover the masked features using neighborhood information, thereby enhancing the road segment representation's ability to encode structural semantics.

[0037] S45. Contrastive Learning Task: Construct positive and negative sample pairs based on road segment similarity, and guide the model to narrow the representation distance of similar road segments and distinguish dissimilar road segments through contrastive loss, thereby further improving the discriminativeness of the representation space.

[0038] S46. Loss Fusion Optimization: The recovery loss generated by the mask recovery task and the contrastive loss generated by the contrastive learning task are weighted and fused, and the model parameters are updated through backpropagation to achieve joint optimization of road segment representation.

[0039] S47. Output road segment representation: After training and optimization, output a road segment representation vector that integrates structural information, semantic information and trajectory behavior features.

[0040] S5. Sensitive Location Identification: Execute a custom sensitive location identification algorithm that combines stop point detection and semantic weighting on the trajectory data, calculate the sensitivity and set a sensitivity threshold to determine sensitive location points in the trajectory.

[0041] S51. Extracting the dwell point: When in a continuous sampling sequence The sum of the time intervals between consecutive sampling points is greater than a preset time threshold. And the movement distance is less than the preset spatial threshold. If the area is identified as a potential dwelling area, the coordinates of the dwelling point are determined by averaging the coordinates of all sampling points within that area, and the total time span is recorded as the actual dwelling time. .

[0042] S52. Sensitive Location Determination: Traverse the set of dwell points and calculate the comprehensive determination index. :

[0043] The semantic weight corresponding to the point of interest type of dwell point i. This represents the frequency with which users access this area.

[0044] S6. Privacy Budget Allocation: Targeting identified sensitive locations The position sensitivity calculated based on step S5 Dynamically allocate differentiated total privacy budgets to different levels of sensitivity for different levels of stay. The specific steps are as follows:

[0045] S61. Dynamic Calculation of Total Budget: The system utilizes an exponential decay function, based on sensitivity... Calculate the total privacy budget specific to this location. Highly sensitive locations will receive a smaller overall budget to ensure stronger privacy protection.

[0046] S62. Budget Allocation: Based on the differential privacy sequential combination theorem, the total budget is allocated... According to the preset ratio It is divided into two sub-budgets, one for macro and one for micro confounding:

[0047] Roadside privacy budget The mechanism assigned to step S7 is used for filtering target confusing road segments, mainly targeting the concealment of macroscopic spatial locations.

[0048] Location privacy budget The Laplace sampling of the dwell point is assigned to step S8, which mainly targets the micro-obfuscation of local dwell behavior.

[0049] Note: Usually set This ensures that the vast majority of the privacy budget is used to protect location offsets at the road network level, while the remaining smaller budget is sufficient to generate moderately discrete point clusters on local road segments.

[0050] S7. Confused Road Segment Filtering: An initial road segment set is constructed based on the road network topology with sensitivity as the benchmark. Filtering is performed by combining semantic similarity and topological reachability, and adaptive expansion of the search range is supported to ensure the number of candidate road segments. Finally, a utility function combining semantics and spatial distance is constructed, incorporating the road segment privacy budget allocated in step S6. The probability is calculated using a differential privacy index mechanism to select target confusing road segments. The specific steps are as follows:

[0051] S71. Based on sensitive locations sensitivity Dynamically calculate the required number of candidate road segments. The calculation formula is:

[0052] in, The number of basic candidate road segments (minimum safety threshold). For sensitivity weighting coefficients, rounding up ensures... It must be a valid positive integer. This mechanism allows road segments with higher sensitivity to adaptively trigger a larger-scale candidate set search;

[0053] S72. Starting with the original sensitive road segments, and relying on the physical topology of the road network, breadth-first search (BFS) is used to expand outwards to adjacent road networks layer by layer. During the traversal, each road segment undergoes both semantic similarity evaluation and topological reachability verification; if both are satisfied, it is added to the candidate road segment set. The process continues until the number of road segments in the set reaches the target number. The search will terminate immediately upon that time.

[0054] S73. Constructing the comprehensive utility function: Constructing the utility function To comprehensively consider the candidate road sections With the original sensitive road section The degree of semantic and spatial structural fit, with its range normalized to [value range missing]. :

[0055] in, These are the weighting coefficients.

[0056] Semantic utility The road segment representation vector output during the feature representation learning stage is utilized. (Original road section) and (Traverse the road segments) and calculate the normalized cosine similarity between the two:

[0057] Trajectory structure utility Sensitive road sections With candidate road sections This is mapped to directed line segments in two-dimensional space, taking into account the spatial deviations between the two in three dimensions: perpendicularity, parallelism, and angle. First, the combined line segment distance is calculated. :

[0058] in, The vertical distance represents the lateral spatial offset between the two road segments; The parallel distance represents the longitudinal spatial offset between two road segments; The angular distance represents the angle of deflection between the two road segments. The corresponding distance weight parameters satisfy the condition that the sum of the weights is 1.

[0059] To satisfy the requirement of bounded utility function in the differential privacy index mechanism, an exponential decay function is used to map the comprehensive distance to trajectory structure utility (the smaller the distance, the higher the utility). The maximum possible composite line segment distance (normalization constant) in the candidate set:

[0060] S74 is based on a differential privacy index mechanism, combined with the allocated road segment privacy budget. and the global sensitivity of the utility function. (Since the utility function has been normalized to) The candidate road segment set is calculated according to the following probability formula. The probability of selection for each candidate road segment And randomly select one of the following road segments based on probability as the final target for confusion:

[0061] S8. Sensitive Location Replacement: The sensitive location obfuscation process in step S8 is as follows, with the specific method for each step being:

[0062] S81. New Dwelling Center Location: First, on the selected target confusion section, find the location node with the highest semantic similarity to the original sensitive location point in the local micro-environment. Establish this node as the new dwelling center. This serves as the reference point for generating subsequent confusion trajectory points.

[0063] S82. Sampling Point Count Statistics: To maintain the continuity of trajectory data in terms of time series and data scale, the total number of actual GPS trajectory points corresponding to the original sensitive location points is counted, denoted as . The generalization phase will generate an equal number of obfuscated trajectory points on the target road segment.

[0064] S83. Reconstructing the Probability Density Function: To add Laplace noise that satisfies differential privacy on a two-dimensional geographic plane, this study... Reconstruct the probability density function in polar coordinates with the origin. Combine this with the assigned location privacy budget. Set random variables The variable represents the drift distance of the sampling point from the center. This represents the drift direction angle. The two-dimensional Laplace mechanism requires its joint probability density to be spatially exponentially decaying and isotropic. and The marginal probability density functions are derived as follows:

[0065] The above formula shows that the direction angle exist The interval follows a uniform distribution; while the drift distance The probability density decreases exponentially with increasing distance and is affected by privacy budgets. Strict control.

[0066] S84. Laplace sampling generation: for drift distance The sampling, its cumulative distribution function (CDF) contains forms such as The term. If the traditional inverse transform sampling method is used to directly invert the CDF, it will inevitably introduce a special Lambert W function to solve the complex transcendental equations. To avoid this computational bottleneck, this scheme seeks an equivalent alternative based on statistical principles. Based on statistical principles, the probability density function Essentially equivalent to a shape parameter of 2 and a scale parameter of... gamma distribution According to the properties of probability theory, this gamma distribution can be considered as having two independent means. The sum of exponentially distributed variables.

[0067] Based on this isomorphic property, this algorithm adopts an efficient equivalent sampling strategy: firstly, in Two uniformly distributed random variables are generated independently within the interval. and Two independent exponentially distributed samples are generated by logarithmic transformation, and then summed to obtain the final drift distance. :

[0068] At the same time, a uniformly distributed random angle is generated. Finally, by performing a trigonometric mapping from standard polar coordinates to rectangular coordinates, the first... a generalized trajectory point Specific two-dimensional coordinates:

[0069] go through After one independent sampling cycle, a set of confused trajectory sampling points that meets both strict differential privacy protection requirements and conforms to the characteristics of real spatial scattering distribution can be generated around the new stop center of the target confused road segment.

[0070] S9. Trajectory Spatiotemporal Reconstruction: To address the local trajectory spatiotemporal discontinuity problem caused by obfuscation, a path planning-based local reconstruction mechanism is adopted to ensure the continuity and authenticity of the obfuscated trajectory in both spatial road network topology and time. Combined with... Figure 4 As shown, the specific process is as follows:

[0071] S91. Splicing Point Location: Obtain sensitive location points in the original trajectory. The new stay center generated in step S81 and adjacent to the original trajectory Preceding stop and subsequent stops Sensitive location points and previous stop Mark the midpoint between them as A, and then mark the sensitive location points. and subsequent stops The midpoint between them is marked as B. Inter-addressing to extract the preceding splice point and in the trajectory segment Inter-addressing to extract post-order splice points Among them, ensure points To the original sensitive point The distance to its new stay center The distance is close ( ),point Time The distance to its new stay center The distance is close ( ).

[0072] S92. Path Reconstruction: Remove nodes that are adjacent to previous join points. and subsequent splicing points Trajectory segments between them. Connecting the preceding splicing points. Using the coordinates as the starting point, the new stay center The coordinates are used as the waypoints and subsequent splicing points. Using the coordinates as the endpoint, the system calls the driving route planning API of online map open platforms (such as Gaode, Baidu, or OpenRouteService) to obtain the sequence of entry road segment nodes that fit the real road topology.

[0073] S93. Global Time Update: A time-based extrapolation mechanism based on trajectory departure time is used to re-assign timestamps to the entire spatially reconstructed trajectory. The specific steps are as follows:

[0074] S931 extracts the initial departure time of the entire original trajectory as the time reference, and performs sequential calculations along the node sequence of the spliced ​​new trajectory.

[0075] For newly added road sections, S932 calculates the travel time difference between nodes by combining the spatial distance between adjacent points and the physical speed limit of the road network, and accumulates them to the timestamp one by one; when extrapolating to the new stop center... At that time, the sensitive point described in S5 is directly extracted to its original position. Actual stay time This data is accumulated to the current timestamp to ensure that the user's actual dwell time is preserved.

[0076] S933. When multiple sensitive locations are identified, steps S91 to S93 are repeated sequentially for each sensitive point according to the travel time sequence, continuously performing local spatial replacement and deduction until all sensitive points in the trajectory have been processed. Finally, based on the continuously accumulated time difference, the final arrival time of the entire trajectory is derived, completing the global closed loop of spatiotemporal consistency.

Claims

1. A vehicle trajectory privacy protection method based on semantic awareness and differential privacy, characterized in that, Includes the following steps: S1. Data Input: Input the vehicle trajectory dataset and the corresponding city road network data; S2. Road segment extraction: Based on the road network topology, the input urban road network is divided into regular road segments; S3. Semantic Information Mining: Obtain the translation distances on the left and right sides of each road segment. Within the region, the data on points of interest are used to extract their type and density information, and these are combined with the inherent semantic attributes of the road segment (including road type, road segment length, number of lanes, maximum driving speed, and in-degree and out-degree in the road network map) to form a set of semantic features for the road segment. S4. Road segment representation generation: Based on the structural characteristics and semantic features of road segments, a transition probability matrix is ​​constructed by statistically analyzing the number of transitions of sub-road segments in trajectory data. This matrix is ​​then used as prior information to embed into the attention calculation mechanism of the graph attention network. Through self-supervised training, a road segment representation vector that integrates physical structure, semantic information, and trajectory behavior patterns is generated. S5. Sensitive Location Identification: Extract potential dwell points from continuous sampling sequences based on time and space thresholds, and calculate the sensitivity of each dwell point by combining dwell time, semantic weight of interest points and access frequency, and set sensitivity thresholds to determine sensitive location points in the trajectory; S6. Privacy Budget Allocation: For the identified sensitive locations, based on the calculated location sensitivity, a differentiated total privacy budget is dynamically allocated using an exponential decay function. The total privacy budget is then divided into a road segment privacy budget for filtering obfuscated road segments and a location privacy budget for sampling sensitive points, according to the serial combination theorem of differential privacy. S7. Filtering of Confused Road Segments: The number of target candidate road segments is dynamically determined based on the sensitivity of sensitive locations. Starting from the road segment where the sensitive location is located, the breadth-first search is used to expand outward based on the road network topology. A set of candidate road segments is constructed by combining semantic similarity and topological reachability. A utility function combining semantic and spatial structural distance is constructed. The road segment privacy budget described in S6 is introduced and the differential privacy index mechanism is applied to calculate the probability and select the target confused road segments. S8. Sensitive Location Replacement: Locate the location with the highest semantic similarity on the target confusing road segment as the new stopping center, introduce the location privacy budget described in S6, and generate generalized sampling point coordinates containing drift distance and direction angle based on the two-dimensional plane Laplace mechanism in polar coordinate system and the equivalent sampling strategy of gamma distribution, so as to replace the trajectory points of the original sensitive location; S9. Trajectory Spatiotemporal Reconstruction: Perform local spatial reconstruction based on online path planning on the trajectory after replacing sensitive points, and use a deduction mechanism based on departure time to complete the global timestamp update.

2. The method according to claim 1, characterized in that, The specific process of regularizing sub-road segment division in step S2 is as follows: Physical nodes such as intersections and road junctions are used as dividing boundaries; Based on the preset sub-segment length threshold and floating range Perform length-constrained slicing to generate length constraints in A collection of road network sub-segments.

3. The method according to claim 1, characterized in that, The specific process of generating road segment representation vectors through self-supervised task training in step S4 is as follows: Extract the attributes of each road segment, and use the set of semantic features of the road segments constructed in step S3 of right 1 as the semantic attributes of each road segment. At the same time, count the number of transitions between road segments in the trajectory as additional feature input. Based on the road network topology and road segment transfer relationships, a road segment-level graph structure is constructed, forming nodes and their adjacency relationships. Nodes represent road segments, and edges represent the physical connections and trajectory transfer probabilities between road segments. Input the segment-level graph structure into the graph attention network, and calculate the attention weights between nodes by combining the transfer prior relationship; The features of some road segments are randomly masked to perform a mask recovery task, and positive and negative sample pairs are constructed based on road segment similarity to perform a contrast learning task. The recovery loss generated by the mask recovery task and the contrastive loss generated by the contrastive learning task are weighted and fused to update the model parameters to achieve joint optimization and output the road segment representation vector.

4. The method according to claim 1, characterized in that, The specific process of sensitive location identification in step S5 is as follows: Determine whether the sum of the time intervals of consecutive sampling points in the trajectory is greater than a preset time threshold. And whether the moving distance is less than the preset spatial threshold. If so, it is identified as a potential dwelling area, and the average of the coordinates of all sampling points within that area is calculated as the dwelling point coordinates. The total time span is recorded as the actual dwelling time. ; Retrieve semantic weights based on the type of interest matched with the stop point and assign them custom values. And calculate the frequency of user visits to this area. The The ratio of the number of visits to this stop area to the total number of visits to all locations on the trajectory; The following formula is used to calculate the comprehensive judgment index as the sensitivity of this stop point. : Determine the calculated sensitivity Is it greater than the preset threshold? If so, then that point is officially identified as a sensitive location point in the trajectory.

5. The method according to claim 1, characterized in that, The specific process of privacy budget allocation in step S6 is as follows: Using the exponential decay function, based on the sensitivity obtained in step S4 Calculate the total privacy budget specific to this location. The formula is: Based on the serial combination theorem of differential privacy, the total budget is... According to the preset ratio Divide into road segment privacy budget Location privacy budget ;in A screening mechanism is assigned to target confusing road segments; , and assign Laplace sampling for the dwell points.

6. The method according to claim 1, characterized in that, The specific process for screening confusing road sections in step S7 is as follows: First, based on sensitive location points sensitivity Dynamically calculate the number of target candidate road segments The formula is: in, Based on the number of candidates, For sensitivity weighting coefficients, rounding up ensures... It is a positive integer; Then, starting from the original sensitive road segment, and relying on the physical topology of the road network, breadth-first search is used to expand to the surrounding adjacent road network layer by layer; During the traversal process, the semantic similarity between the traversed road segments and the original road segment representation vectors is calculated. : The road segment representation vector output during the feature representation learning phase (Original road section) and (Traverse the road segments) Calculate the normalized cosine similarity between the two segments; and verify topological reachability. Add road segments that meet the preset semantic similarity threshold and have physical reachability to the candidate road segment set until the number of road segments in the candidate road segment set reaches the target number. Stop searching at this time; Constructing a comprehensive utility function Taking into account the candidate road sections With the original sensitive road section The degree of semantic and spatial structural fit, with its range normalized to [value range missing]. : in, These are the weighting coefficients. Let the semantic similarity between the above road segment and the original road segment representation vector be denoted as . To normalize the trajectory structure effectiveness, it considers the spatial deviation of two road segments in three dimensions: perpendicular, parallel, and angular. The formula is as follows: First, calculate its comprehensive line segment distance. : in, The vertical distance represents the lateral spatial offset between the two road segments; The parallel distance represents the longitudinal spatial offset between two road segments; The angular distance represents the angle of deflection between the two road segments. The corresponding distance weight parameters satisfy the condition that the sum of the weights is 1; An exponential decay function is used to map the overall distance to trajectory structure utility (the smaller the distance, the higher the utility): in, The maximum possible composite line segment distance (normalization constant) in the candidate set; Finally, based on the differential privacy index mechanism, the road segment privacy budget allocated in step S6 is introduced. Calculate the candidate road segment set Selection probability of each road segment : One road segment is randomly selected as the target confusion segment based on probability.

7. The method according to claim 1, characterized in that, In step S8, location points are randomly sampled using the Laplace mechanism and used to replace the original sensitive location points in the trajectory. Locate the node with the highest local semantic similarity to the original sensitive location point on the target confusing road segment and establish it as the new dwell center. ; Count the total number of actual GPS track points corresponding to the original sensitive location points. ; With the new stay center Reconstruct the two-dimensional probability density function in polar coordinates at the origin, and combine it with the location privacy budget allocated in step S6. An equivalent sampling strategy is implemented by utilizing the isomorphism between the gamma distribution and the exponential distribution, as follows; Combined with the allocated location privacy budget ,set up This indicates the drift distance of the sampling point from the center. Indicates the drift direction angle. and The marginal probability density function is: Due to drift distance Since its cumulative distribution function cannot be inverted, an equivalent gamma distribution is used as a substitute. Two uniformly distributed random variables are generated independently within the interval. and Two independent exponentially distributed samples are generated by logarithmic transformation and added together to calculate the drift distance of the generalized trajectory points from the center. : Regenerate a drift direction angle that follows a uniform distribution. By mapping polar coordinates to rectangular coordinates, the first... Two-dimensional coordinates of a generalized trajectory point : Perform independent sampling repeatedly until a result is generated. A generalized trajectory sampling point is used to replace the original sensitive location point.

8. The method according to claim 1, characterized in that, The specific process of trajectory spatiotemporal reconstruction in step S9 is as follows: Obtain sensitive location points in the original trajectory New Stay Center and adjacent to the original trajectory Preceding stop and subsequent stops ; sensitive location points and previous stop The midpoint between them is marked as Sensitive location points and subsequent stops The midpoint between them is marked as ;exist Inter-addressing to extract the preceding splice point and in the trajectory segment Inter-addressing to extract post-order splice points ; Remove points between preceding splice points and subsequent splicing points Trajectory segments between; connecting preceding points Using the coordinates as the starting point, the new stay center The coordinates are used as waypoints and subsequent splicing points. Using the coordinates of the destination as the endpoint, the driving route planning API of the online map open platform is called to obtain the sequence of entry road segment nodes that fit the real road topology; The initial departure time of the entire original trajectory is extracted as the time reference, and the calculation is performed sequentially along the node sequence of the spliced ​​new trajectory. For newly added road segments, the travel time difference between nodes is calculated by combining the spatial distance between two adjacent points and the physical speed limit of the road network, and then accumulated to the timestamp one by one. When the deduction is extended to the new residence center At that time, the sensitive point described in step S5 is extracted at its original position. Actual stay time As a constant, it is accumulated to the current timestamp; if there are multiple sensitive locations in the trajectory, the above local spatial replacement and deduction steps are repeated in sequence according to the travel time, and the final arrival time of the entire trajectory is naturally derived based on the continuously accumulated time difference.

Images