User track privacy protection method and system fusing generative model and three-stage personalized differential privacy

Abstract

The invention discloses a user track privacy protection method and system fusing a generative model and three-stage personalized differential privacy, and relates to the technical field of data security, and the method comprises the steps: obtaining real user track data, and carrying out the preprocessing of the real user track data to form standardized track representation; training a depth generation model to generate synthetic trajectory data; performing density clustering on the synthetic trajectory to obtain a feature region and calculating a region centroid; point-level privacy budgets are distributed according to three stages of the area proportion, the centroid distance and the global density; constructing an adaptive elliptical disturbance domain based on the regional covariance, sampling candidate confusion points in the disturbance domain and applying road feasibility screening; and calculating a score based on the privacy budget and the utility function, and sampling replacement points by adopting an index mechanism to output a privacy protection trajectory. According to the technical scheme, the track structure continuity and the road semantic consistency are remarkably improved under the condition that the differential privacy constraint is met, and the privacy intensity and the data availability are both considered.

Description

A method and system for protecting user trajectory privacy by integrating generative models and three-stage personalized differential privacy. Technical Field

[0001] This invention relates to the field of data security technology, and in particular to a user trajectory privacy protection method that integrates generative models and three-stage personalized differential privacy, as well as a user trajectory privacy protection system that integrates generative models and three-stage personalized differential privacy. Background Technology

[0002] With the development of mobile internet, the Internet of Things, and location-based services (LBS), smartphones, in-vehicle terminals, and wearable devices continuously collect and upload location information, forming massive and multi-dimensional user trajectory data. Trajectory data has significant application value in scenarios such as urban traffic optimization, business district selection, personalized travel recommendations, and emergency command and dispatch. However, trajectory data inherently carries sensitive information such as residence, workplace, and travel time patterns, making it vulnerable to identity and behavioral inferences through trajectory re-identification and external information correlation, posing a significant risk of privacy breaches.

[0003] To achieve secure publication and sharing of trajectory data, existing solutions mainly fall into three categories: anonymization / generalization, differential privacy perturbation, and deep generative model synthesis. Anonymization / generalization methods typically reduce linkability through spatiotemporal gridding, time segmentation, and fragment suppression. However, in high-dimensional, long-sequence trajectory scenarios, strong generalization is often required, leading to disruption of the trajectory's spatiotemporal continuity and decreased accuracy in upper-layer applications. Differential privacy trajectory perturbation methods provide formal privacy guarantees by injecting noise into trajectory point coordinates or statistics. However, traditional methods often employ uniform or coarse-grained privacy budget allocation strategies, making it difficult to finely distinguish between high-density hotspots and low-density sensitive areas. This can easily result in excessive local perturbation or insufficient local protection, leading to an imbalance between privacy and utility.

[0004] Deep generative modeling methods generate synthetic trajectories by learning the spatiotemporal distribution, behavioral patterns, and semantic attributes of real trajectories, thereby reducing the risk of directly publishing the original trajectories. However, relying solely on generative models cannot provide rigorously provable privacy guarantees and is still susceptible to attacks such as member inference and model inversion, which could leak user behavior patterns from the training set. Existing techniques that combine generative models with differential privacy can improve privacy strength, but there is still room for optimization in terms of privacy budget allocation granularity, global-local collaborative protection in complex scenarios, and preservation of trajectory geometry and behavioral patterns.

[0005] Therefore, there is an urgent need for a privacy protection scheme for trajectory data release: while providing formal privacy protection, it should achieve more granular budget control for different regional densities and individual sensitivities, reduce the damage of disturbances to the spatiotemporal structure and behavioral characteristics of trajectories, and improve the usability and stability of privacy-protected trajectories in applications such as hotspot identification, trajectory clustering, and behavioral analysis. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a user trajectory privacy protection method and system that integrates generative models and three-stage personalized differential privacy. By training a deep generative model based on real user trajectories and generating a synthetic trajectory dataset, it avoids directly publishing and perturbing real trajectories. A three-stage personalized privacy budget allocation mechanism is employed, progressively refining the privacy budget from the region level to the point level and updating it with density adjustment. This allows different feature regions and trajectory points to receive differentiated protection strengths, reducing the problems of excessive perturbation or insufficient protection caused by traditional unified budgets. An adaptive elliptical perturbation domain consistent with the main trajectory direction is constructed based on the feature region covariance. After generating candidate confusion points within the perturbation domain, spatial constraints such as road feasibility are applied to ensure that replacement points meet spatial feasibility requirements. Replacement points are determined by sampling from the candidate confusion points through a preset utility function and exponential mechanism. While satisfying differential privacy constraints, this approach maintains the trajectory geometry and local continuity as much as possible, improving the usability and stability of privacy-protected trajectory data.

[0007] To achieve the above objectives, this invention provides a user trajectory privacy protection method that integrates a generative model and three-stage personalized differential privacy, comprising: acquiring a real user trajectory dataset, and performing trajectory feature extraction and preprocessing on the real user trajectory dataset to obtain a standardized trajectory representation, wherein the real user trajectory dataset includes a trajectory sequence composed of multiple trajectory points; training a deep generative model based on the standardized trajectory representation, and generating a synthetic trajectory dataset by the deep generative model; extracting feature regions from the synthetic trajectory dataset to obtain multiple feature regions, and determining the centroid of each feature region; and performing a three-stage personalized privacy budget allocation on the trajectory points in the synthetic trajectory dataset to obtain... A point-level privacy budget is generated for each trajectory point. A covariance matrix is ​​calculated for the trajectory point set within each feature region, and the principal axis direction and axis length scale parameters of the ellipse are determined based on the covariance matrix to construct an adaptive elliptical perturbation domain. A candidate confusion point set is generated within the elliptical perturbation domain, comprising multiple candidate confusion points, and spatial feasibility constraints are applied to the candidate confusion point set for screening. A utility score for the candidate confusion points is calculated based on the point-level privacy budget and a preset utility function, and an exponential mechanism is used to sample and determine replacement points from the candidate confusion point set according to the utility score. The corresponding trajectory points in the synthetic trajectory dataset are then perturbed to replace these replacement points, resulting in a privacy-preserving trajectory dataset.

[0008] In the above technical solution, preferably, the specific process of extracting and preprocessing trajectory features from the real user trajectory dataset includes: cleaning and sorting the trajectory sequence by time, and padding the trajectory sequence to unify the trajectory length, while setting mask identifiers for the padding positions; normalizing the latitude and longitude information in the trajectory sequence and converting the latitude and longitude information into coordinate offsets relative to the centroid of the trajectory point set; and performing one-hot encoding on the time attribute and location category attribute in the trajectory sequence to obtain the standardized trajectory representation.

[0009] In the above technical solution, preferably, the deep generative model is a CNN-LSTM-GAN model, including a trajectory attribute embedding layer, a convolutional feature extraction layer, a long short-term memory network layer, and a generator and discriminator of a generative adversarial network; the trajectory attribute embedding layer is used to map the standardized trajectory representation to a high-dimensional semantic space; the convolutional feature extraction layer is used to extract local movement trend features; the long short-term memory network layer is used to model the time-dependent features of the trajectory; the generator is used to generate synthetic trajectory data, and the discriminator is used to distinguish between true and false input trajectories; the generator and the discriminator are trained adversarially to obtain the deep generative model.

[0010] In the above technical solution, preferably, the feature region extraction uses the density clustering algorithm DBSCAN to cluster the trajectory points in the synthetic trajectory dataset, and determines each density cluster obtained by clustering as the feature region, and calculates the corresponding region centroid based on the spatial coordinates of the trajectory points in each feature region.

[0011] In the above technical solution, preferably, the specific process of the three-stage personalized privacy budget allocation includes: First, for each feature region of each trajectory, a region-level privacy budget is determined based on the number of trajectory points within the feature region and the total number of trajectory points; Second, based on the distance between each trajectory point within the feature region and the centroid of the corresponding region, and combined with user sensitivity parameters and a preset privacy level list, the region-level privacy budget is allocated as a point-level privacy budget corresponding to each trajectory point; Third, density clustering is performed on the global trajectory point set in the synthetic trajectory dataset to obtain multiple density clusters, a density adjustment coefficient is calculated based on the point density of each density cluster and preset adjustment parameters, and the point-level privacy budget of the trajectory points belonging to the corresponding density cluster is scaled and updated based on the density adjustment coefficient.

[0012] In the above technical solution, preferably, the specific process of constructing the adaptive elliptical perturbation domain includes: calculating the covariance matrix of the spatial coordinates of the trajectory points within the feature region, and determining the principal axis direction and semi-axis scale parameters of the ellipse based on the covariance matrix; using the centroid of the region as the center of the ellipse, and constructing the adaptive elliptical perturbation domain based on the principal axis direction and the semi-axis scale parameters.

[0013] In the above technical solution, preferably, the specific process of generating a set of candidate confusion points in the elliptical perturbation domain includes: modeling the spatial distribution of trajectory points in the feature region based on kernel density estimation or Gaussian mixture model, and sampling to generate the candidate confusion points in the elliptical perturbation domain; the spatial feasibility constraint includes road feasibility constraint, which is used to calculate the minimum distance from the candidate confusion point to the nearest road segment in the road network based on the road network, and retaining candidate confusion points whose minimum distance is not greater than a preset road buffer threshold.

[0014] In the above technical solution, preferably, the preset utility function includes a regional consistency term and a local consistency term; the regional consistency term is used to calculate a regional consistency score based on the distance between the candidate confusion point and the centroid of the region and the minimum distance from the candidate confusion point to the nearest road segment in the road network; the local consistency term is used to calculate a local consistency score based on the direction difference and curvature difference between the candidate confusion point and the previous and next trajectory points of the trajectory point to be replaced; the utility score is a weighted combination of the regional consistency score and the local consistency score; the index mechanism is used to index the utility score based on the point-level privacy budget to determine the sampling probability of each candidate confusion point, and to sample and determine the replacement point according to the sampling probability.

[0015] In the above technical solution, preferably, the regional consistency item includes a centroid consistency sub-item and a road consistency sub-item; the centroid consistency sub-item is used to calculate a centroid consistency score based on the distance between the candidate confusion point and the regional centroid; the road consistency sub-item is used to calculate a road consistency score based on the minimum distance from the candidate confusion point to the nearest road segment in the road network; the regional consistency score is a weighted combination of the centroid consistency score and the road consistency score, wherein the scores participating in the combination are normalized; the local consistency item includes a distance consistency sub-item and a shape consistency sub-item; the distance ... local consistency score is calculated based on the minimum distance from the candidate confusion point to the nearest road segment in the road network; the regional consistency score is a weighted combination of the centroid consistency The spatial distance between the selected confusion point and the trajectory point to be replaced is used to calculate the distance consistency score; the shape consistency sub-item is used to calculate the direction consistency score based on the direction vector difference formed by the candidate confusion point and the previous trajectory point and the next trajectory point, respectively, and to calculate the curvature continuity score based on the second difference of the previous trajectory point, the candidate confusion point, and the next trajectory point; the shape consistency score is a weighted combination of the direction consistency score and the curvature continuity score; the local consistency score is a weighted combination of the distance consistency score and the shape consistency score, wherein each weighted combination adopts a preset weight and the scores participating in the combination are normalized.

[0016] This invention also proposes a user trajectory privacy protection system that integrates a generative model and three-stage personalized differential privacy, used to implement the user trajectory privacy protection method that integrates a generative model and three-stage personalized differential privacy as disclosed in any of the above technical solutions. The system includes: a trajectory acquisition module for acquiring a real user trajectory dataset, wherein the real user trajectory dataset includes a trajectory sequence composed of multiple trajectory points; a trajectory preprocessing module connected to the trajectory acquisition module for extracting and preprocessing trajectory features from the real user trajectory dataset and outputting a standardized trajectory representation; a synthetic trajectory generation module connected to the trajectory preprocessing module for training a deep generative model based on the standardized trajectory representation and generating a synthetic trajectory dataset from the deep generative model; a feature region extraction module connected to the synthetic trajectory generation module for extracting feature regions from the synthetic trajectory dataset to obtain multiple feature regions and determining the centroid of each feature region; and a privacy budget allocation module connected to the feature region extraction module for allocating privacy budgets to the synthetic trajectory dataset. The dataset performs a three-stage personalized privacy budget allocation on trajectory points to obtain a point-level privacy budget corresponding to each trajectory point. A perturbation domain construction module, connected to the feature region extraction module, calculates the covariance matrix for the trajectory point set within each feature region and determines the elliptical principal axis direction and axis length scale parameters based on the covariance matrix, thereby constructing an adaptive elliptical perturbation domain. A confusion point generation and filtering module, connected to the perturbation domain construction module, generates a candidate confusion point set within the adaptive elliptical perturbation domain and applies spatial feasibility constraints to the candidate confusion point set for filtering. A utility scoring and sampling module, connected to the confusion point generation and filtering module, calculates the utility score of the candidate confusion points based on the point-level privacy budget and a preset utility function, and uses an exponential mechanism to sample replacement points from the candidate confusion point set according to the utility score. A perturbation replacement output module, connected to the utility scoring and sampling module, performs replacement perturbations on the corresponding trajectory points in the synthetic trajectory dataset to generate and output a privacy-preserving trajectory dataset.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Balance between privacy protection and data availability. The present invention generates synthetic trajectory data through a deep generative model, avoiding direct exposure and disturbance of the real trajectory from the source, thus effectively reducing the risk of privacy leakage. At the same time, the generated synthetic trajectory is highly similar to the real data in terms of spatiotemporal distribution and behavioral patterns, ensuring that the data still has high availability under the premise of privacy protection.

[0018] (2) Personalized differential privacy budget allocation: Through three-stage personalized privacy budget allocation, the allocation of privacy budget is precisely controlled according to the sensitivity of different trajectory points and the density of the area they are in. High-sensitivity areas receive more privacy protection, while low-sensitivity areas are not excessively disturbed, thereby optimizing the balance between privacy protection and data utility and improving the efficiency of privacy protection.

[0019] (3) Adaptive Elliptical Perturbation Domain Generation and Spatial Constraints This invention constructs an adaptive elliptical perturbation domain by calculating the covariance matrix of trajectory points within the feature region, ensuring that the perturbation points remain reasonable in space. At the same time, spatial constraints such as road feasibility are applied to ensure that the candidate confusion points are feasible on the road network, thereby improving the rationality and availability of data replacement.

[0020] (4) Combination of utility score and index mechanism By calculating the utility score of candidate confusion points and using the index mechanism to sample and determine replacement points according to the utility score, this invention can maximize the spatial consistency and local continuity of trajectory data while ensuring privacy protection, and further improve the reliability and accuracy of privacy-protected trajectory data. Attached Figure Description

[0021] Figure 1 is a flowchart illustrating the fusion generation model and the three-stage personalized differential privacy user trajectory privacy protection method disclosed in an embodiment of the present invention; Figure 2 is an architectural diagram illustrating the deep generation model disclosed in an embodiment of the present invention; Figure 3 is a schematic diagram illustrating the principle of feature region extraction disclosed in an embodiment of the present invention; Figure 4 is a schematic diagram illustrating the adaptive elliptical perturbation domain disclosed in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0023] The present invention will now be described in further detail with reference to the accompanying drawings: As shown in Figure 1, the user trajectory privacy protection method according to the present invention, which integrates a generative model and three-stage personalized differential privacy, includes: acquiring a real user trajectory dataset, wherein the real user trajectory dataset includes a trajectory sequence composed of multiple trajectory points, and performing trajectory feature extraction and preprocessing on the real user trajectory dataset to obtain a standardized trajectory representation, so as to unify the expression granularity and data form of the trajectory, and provide consistent input for subsequent generative model training and privacy mechanism calculation.

[0024] A deep generative model is trained based on standardized trajectory representation, and a synthetic trajectory dataset is generated by the deep generative model. This allows subsequent privacy processing to be carried out on synthetic data, reducing the direct exposure of real trajectories.

[0025] Feature regions are extracted from the synthetic trajectory dataset to obtain multiple feature regions, and the centroid of each feature region is determined so that subsequent privacy budget allocation and perturbation domain construction have a regional reference benchmark.

[0026] A three-stage personalized privacy budget allocation is performed on the trajectory points in the synthetic trajectory dataset to obtain the point-level privacy budget corresponding to each trajectory point, so that different trajectory points achieve a differentiated balance between privacy strength and data availability.

[0027] The covariance matrix is ​​calculated for the set of trajectory points in each feature region, and the direction of the principal axis and the axis length scale parameters of the ellipse are determined based on the covariance matrix, thereby constructing an adaptive elliptical perturbation domain to reduce the structural distortion caused by the traditional fixed perturbation domain.

[0028] A candidate confusion point set is generated within the elliptical perturbation domain. The candidate confusion point set includes multiple candidate confusion points, and spatial feasibility constraints are applied to the candidate confusion point set for screening to ensure that the candidate points meet the spatial semantic rationality.

[0029] The utility score of candidate confusion points is calculated based on point-level privacy budget and preset utility function. Then, an index mechanism is used to sample and determine replacement points from the candidate confusion point set according to the utility score. The corresponding trajectory points in the synthetic trajectory dataset are replaced and perturbed, and the privacy-preserving trajectory dataset is output.

[0030] Specifically, this method provides a complete trajectory privacy publishing process, from trajectory generation, feature region extraction, personalized privacy budget allocation to differential privacy perturbation. It effectively resists privacy attacks such as member inference and model inversion, while preserving the spatiotemporal structure and behavioral characteristics of the trajectory to a high degree. This ensures that the published trajectory data remains suitable for higher-level data analysis tasks such as hotspot region identification, trajectory clustering, behavior analysis, and path prediction. The entire process boasts advantages such as high automation, strong privacy protection, and good data availability, making it suitable for practical applications requiring trajectory data sharing, public publishing, and distributed computing.

[0031] In this implementation, a closed-loop process is formed, consisting of generative model synthesis, three-stage personalized budgeting, adaptive perturbation domain, feasibility screening, and index mechanism replacement. Under the premise of satisfying differential privacy constraints, the continuity of trajectory structure and consistency of behavioral semantics are improved.

[0032] Trajectory data typically contains multi-dimensional information such as latitude and longitude, timestamps, and location categories, and is characterized by variable length, uneven time distribution, and sparse categories. Without preprocessing, this can affect model convergence and privacy budget calculations. Therefore, in the above implementation, preferably, trajectory feature extraction and preprocessing are performed on the real user trajectory dataset to convert the raw trajectory data into a standardized format suitable for generative model training and differential privacy processing. Specifically, this process includes: cleaning and sorting the trajectory sequences by time, and padding the sequences to unify trajectory length, while setting mask identifiers for the padding positions to ensure consistency in length dimension and identifiability of the padding points during subsequent model training and privacy calculations; centroid normalization of the latitude and longitude information in the trajectory sequences, converting the latitude and longitude information into coordinate offsets relative to the centroid of the trajectory point set to reduce the impact of absolute coordinate scale differences on model convergence and perturbation domain scale estimation; and one-hot encoding of the time attributes (day, hour) and location category attributes in the trajectory sequences to form a standardized trajectory representation that can be directly input into the generative model, improving the model's ability to express semantic attributes.

[0033] In this implementation, time series normalization, spatial normalization, and attribute encoding reduce the risks of training instability and budget calculation bias caused by variable trajectory length, uneven time distribution, and sparse categories.

[0034] As shown in Figure 2, in the above implementation, preferably, CNN-LSTM-GAN is used as a deep generative model, including a trajectory attribute embedding layer, a convolutional feature extraction layer, a long short-term memory network layer, and a generator and discriminator of a generative adversarial network, in order to learn the spatial continuity, temporal dependence and behavioral patterns of real trajectories.

[0035] The trajectory attribute embedding layer is used to map the standardized trajectory representation to a high-dimensional semantic space to unify the fusion expression of latitude and longitude offset, time attributes and location categories in the same representation domain; the convolutional feature extraction layer is used to extract local movement trend features, which can capture local movement trends such as short-range straight-line movement, small-range turning back and forth and continuous acceleration changes in a local range. At the same time, it can perceive micro-structural features such as fine-grained dwell time and repetitive commuting patterns near POI (Point of Interest), and has a stronger response to high-frequency abrupt changes or small fluctuations in the trajectory, thereby extracting more refined local spatiotemporal features.

[0036] Long Short-Term Memory (LSTM) layers are used to model the temporal dependencies of trajectories, representing long-range movement patterns. Building upon the local feature extraction performed by CNNs, LSTMs can focus on modeling long-term dependencies and global behavioral logic, significantly reducing the learning burden of sequence models and improving overall expressive power.

[0037] The generator produces synthetic trajectory data, while the discriminator distinguishes between real and fake input trajectories. The generator and discriminator iteratively update within an adversarial training framework to obtain converged generative model parameters. This ensures that the generated trajectories realistically simulate the joint spatial, temporal, and class distribution of real trajectories, maintaining overall statistical realism without directly corresponding to any real user. Furthermore, the detailed information provided by the CNN makes the generated trajector more realistic in its local structure, prompting faster convergence and increasing the difficulty for the discriminator to distinguish between real and fake trajectories, thereby improving the stability and generation quality of the GAN training.

[0038] In this implementation, the distribution fit of the synthesized trajectory is improved while simultaneously covering both spatial local patterns and temporal dependent structures, providing a more stable synthesis basis for subsequent differential privacy perturbations and reducing the usability loss of the perturbated trajectory.

[0039] As shown in Figure 3, in the above embodiment, preferably, the feature region extraction uses the density clustering algorithm DBSCAN to cluster the trajectory points in the synthetic trajectory dataset. Each density cluster obtained from the clustering is determined as a feature region, and the corresponding region centroid is calculated based on the spatial coordinates of the trajectory points within each feature region. The feature regions are used to characterize trajectory hotspots and behavioral pattern clusters, and the region centroid serves as a region-level reference center, participating in subsequent point-level privacy budget allocation, perturbation domain center localization, and region consistency score calculation.

[0040] Specifically, DBSCAN can effectively identify density clusters of arbitrary shapes, making it suitable for dividing irregular routes, densely populated areas of stops, and sparsely populated areas of movement paths. Through density analysis, it can automatically distinguish: 1. Hotspot areas (high density): such as residential areas, workplaces, campuses, and other frequently visited locations.

[0041] 2. Sparse areas (low density): These may be related to the user's personalized behavior and are more sensitive to privacy.

[0042] During implementation, the following steps are performed: (1) Calculation of spatial neighborhood of trajectory points: Search for local dense areas based on the preset neighborhood radius ε and the minimum number of points MinPts.

[0043] (2) Feature region identification: Divide the trajectory points with reachable density into the same feature region.

[0044] (3) Calculate the spatial centroid for each region, which will be used for subsequent personalized privacy budget allocation.

[0045] In this implementation, by using density-driven region partitioning and centroid extraction, the privacy budget and perturbation domain are organized around the real distribution pattern, avoiding the homogenization of hotspot and sparse regions caused by processing at a globally uniform scale.

[0046] In the above implementation, in order to maximize the availability of the trajectory under differential privacy protection, this invention proposes a three-stage, multi-granularity, density-adaptive personalized privacy budget allocation mechanism. Through the synergistic effect of regional features, point features and global density, a different ε privacy budget is allocated to each trajectory point. The first two stages only apply to a single trajectory, while the third stage makes a global adjustment based on all trajectory points, which is more in line with the actual situation.

[0047] Preferably, the specific process of the three-stage personalized privacy budget allocation includes: In the first stage, for each feature region of each trajectory, a region-level privacy budget (ε) is determined based on the number of trajectory points within the feature region and the total number of trajectory points. r This establishes a regional budget baseline by linking the regional budget to the region's proportion within a single trajectory; specifically, based on the differences in each feature region, the privacy budget allocated to each feature region is determined using ε. r Adjustments are made, setting the privacy budget increment to Δε. Among them, t r This is the count of trajectory points within the current feature region, where t is the total number of user trajectories, and K represents the total number of centroids in the region.

[0048] In the second stage, based on the distance between each trajectory point within the feature region and the centroid of the corresponding region, and combined with user sensitivity parameters and a preset privacy level list, the region-level privacy budget is allocated to the point-level privacy budget (ε) corresponding to each trajectory point. p This allows for differentiated budget intensity for trajectory points at different locations within the same region; specifically, the trajectory points in the feature region are closely clustered and densely packed, and different trajectory points have different privacy levels L = {w1, w2, ..., w...}. n Each trajectory point has a different privacy budget allocated to different privacy levels. The privacy of a trajectory point is determined by its distance from the centroid of the region. Where d(pi, ck) represents the distance between trajectory point pi and the centroid of the region ck. The closer the position pi is to the centroid of the region, the higher the user-specified sensitivity s and the higher the privacy w. All positions in the user trajectory are assigned different privacy budgets based on privacy. Trajectory points with higher privacy requirements need higher privacy protection strength and therefore need to be assigned smaller ε. Privacy budget and privacy are inversely proportional. Combining privacy w and regional privacy budget ε r To calculate the privacy budget allocated to each trajectory point: Where, ε r For the regional privacy budget, w piAs can be seen from the formula, the greater the privacy of user trajectory points, the smaller the allocated privacy budget, indicating a greater degree of privacy protection, and vice versa.

[0049] In the third stage, density clustering is performed on the global trajectory point set in the synthetic trajectory dataset to obtain multiple density clusters. Based on the point density of each density cluster and preset adjustment parameters, the density adjustment coefficient (α) is calculated. k The coefficients are used to scale and update the point-level privacy budget of trajectory points belonging to the corresponding density clusters based on the density adjustment coefficients, so that the budget can be consistent at the global density structure level.

[0050] Specifically, to further optimize the strength of privacy protection and make it more precise and reasonable, a third phase of global privacy budget adjustment is introduced. On the one hand, this avoids overprotecting common areas. In many application scenarios, there are frequently visited areas (such as tourist attractions, commercial districts, and public transportation stations). Data in these areas typically does not contain sensitive personal information, therefore, the privacy protection requirements for these areas are relatively low. If privacy budget allocation is based solely on personalized privacy protection mechanisms, it may lead to overprotection of these common areas (e.g., allocating excessively large privacy budgets), resulting in reduced data usability. On the other hand, it ensures adequate protection for sensitive areas. Some areas may contain higher levels of personal privacy information due to less access or less public attention. For example, a user's home address and workplace are often not frequently found in public areas but are closely related to their personal behavior. While personalized privacy protection can adjust the privacy budget based on the sensitivity of trajectory points, without global privacy protection considerations, data in sensitive areas may not receive sufficient privacy protection. Global privacy budget adjustment can identify these low-density, sensitive areas and appropriately increase their privacy budgets according to their privacy protection needs, ensuring that data in these areas receives stricter privacy protection. In this way, data from sensitive areas will not be leaked due to an excessively low privacy budget.

[0051] During implementation, based on global DBSCAN clustering of all trajectory points, the density structure of the entire dataset is analyzed, and the density adjustment coefficient α is calculated. k : Here, γ is an adjustment parameter used to control the degree of influence of cluster density on privacy budget adjustment. A larger γ will make the privacy budget adjustment more significant in areas with higher density. k ) is cluster C k The density represents the density of trajectory points within a cluster; the higher the density, the higher the privacy budget adjustment factor α. k The smaller.

[0052] Ultimately, the privacy budget for each trajectory point is determined by the following formula: in, The second phase of computation calculates a personalized privacy budget, taking into account the sensitivity of trajectory points and the distance to the region's centroid, α. k It is the cluster C to which the trajectory point belongs. k The adjustment factor is calculated based on the density of the cluster.

[0053] In the third phase, DBSCAN clustering is used to assign all trajectory points to different clusters (regions), and the privacy budget for each cluster is adjusted according to the cluster density. For clusters with higher density (such as popular areas or frequently accessed areas), the privacy budget will be appropriately increased; while for clusters with lower density (such as sensitive areas or less frequently accessed areas), the privacy budget will be correspondingly decreased. In this way, personalized privacy protection and global privacy requirements are effectively balanced, significantly improving the security and usability of trajectory publishing, and making privacy budget allocation shift from "homogeneous" to "personalized".

[0054] This implementation achieves multi-granularity budget linkage from region level to point level to global density adjustment, and a collaborative ε-allocation method at the region, point, and global levels. This overcomes the shortcomings of existing DP methods in providing personalized protection based on behavioral sensitivity differences, resulting in stronger protection for sensitive and sparse areas, reduced disturbances in non-sensitive areas, and significantly improved data availability. Simultaneously, it reduces unnecessary disturbances in high-density hotspot areas, enhancing overall availability. Furthermore, the three-stage DP mechanism is universal and can be extended to the protection of other types of sequence data. It is not only applicable to trajectories but also to user behavior sequences, sensor data, and point-of-interest access sequences, offering superior scalability compared to existing trajectory-specific methods.

[0055] As shown in Figure 4, in the above embodiment, preferably, the specific process of constructing the adaptive elliptical perturbation domain includes: calculating the covariance matrix of the spatial coordinates of the trajectory points within the feature region, and determining the principal axis direction and semi-axis scale parameters of the ellipse based on the covariance matrix, so that the perturbation domain conforms to the principal direction distribution of the trajectory points in the region in terms of direction and reflects the expansion shape of the region in terms of scale; using the centroid of the region as the center of the ellipse, constructing the adaptive elliptical perturbation domain based on the principal axis direction and semi-axis scale parameters.

[0056] In the implementation process, in order to avoid the distortion of trajectory shape caused by the traditional circular perturbation domain, an elliptical perturbation domain is constructed based on the covariance matrix of the feature region. This allows the perturbation region to adapt to the trajectory directionality and local geometry, greatly improving the trajectory smoothness and analysis effect.

[0057] Specifically, given trajectory point p j First, based on its cluster region C kConstruct a local perturbation domain, and then use the covariance moments Σ of the point set in this region. k Estimate the main direction and expansion pattern of the region, and define the perturbation region as surrounding the cluster center C. k Ellipse E k : in, , Represents Euclidean distance. This represents the distance from the j-th trajectory point to the cluster center.

[0058] This elliptical domain can adaptively adjust the perturbation range according to the main direction and expansion degree of the cluster, thereby avoiding the generation of a large number of invalid candidate points in the sparse direction of the circular domain.

[0059] In this embodiment, the perturbation domain changes from a fixed shape to an elliptical structure consistent with the statistical morphology of the region, which reduces the invalid candidate points and trajectory distortion generated by the circular perturbation domain in the sparse direction, and improves the smoothness and structural consistency of the perturbation trajectory.

[0060] In the above embodiments, preferably, the specific process of generating a candidate confusion point set within the elliptical perturbation domain includes: modeling the spatial distribution of trajectory points within the feature region based on kernel density estimation or Gaussian mixture model, using the modeling result as the prior spatial distribution of the region, and sampling within the elliptical perturbation domain according to the prior distribution to generate a candidate confusion point set, so that the candidate points are more concentrated in regional hotspots and high-probability spatial locations.

[0061] Then, spatial feasibility constraints are applied to the candidate confusion point set. The road feasibility constraint calculates the minimum distance from the candidate confusion point to the nearest road segment through the road network, and retains candidate confusion points whose minimum distance is not greater than a preset road buffer threshold, so as to filter candidate points that fall into buildings, water bodies or impassable areas.

[0062] Specifically, in the perturbation domain E k To make the candidate confusion points more consistent with the regional hotspot distribution, kernel density estimation (KDE) or Gaussian mixture modeling (GMM) is performed on the point set within the cluster to obtain the spatial density function of the cluster. Candidate points are then randomly sampled from the following truncated distribution: in, This represents the probability of a trajectory point appearing at different locations within the cluster; the closer to the high-density region, the higher the probability. The larger it is, the closer it is to the low-frequency region. The smaller the density, the more concentrated the candidate confusion points will be within a reasonable region.

[0063] in, As an indicator function, it ensures that candidate points are generated only within the perturbation domain. This strategy makes it more likely that candidate points will appear in high-density (hotspot) areas, thereby increasing the possibility of subsequent exponential mechanisms selecting efficient points.

[0064] After completing the construction of the elliptical perturbation domain based on the feature region covariance matrix, in order to further ensure the spatial feasibility and behavioral semantic consistency of the perturbed trajectory points in the real traffic environment, road feasibility constraints are introduced in the differential privacy post-processing stage.

[0065] Specifically, road network data within the study area is acquired in advance, and the road network is represented as a set of road segments or line elements: For any candidate perturbation point p within the elliptical perturbation domain, the minimum distance to the nearest road is defined as: Considering positioning errors and map matching deviations, a road buffer threshold is set. Only candidate perturbation points that meet the following conditions are retained: Therefore, the generation process of candidate confusion points must simultaneously satisfy the prior distribution of the feature region, the constraint of the elliptical perturbation domain, and the constraint of road feasibility. Its sampling distribution can be expressed as: in, E represents the prior spatial distribution of the region obtained based on kernel density estimation or Gaussian mixture models. k The elliptical perturbation domain represents the feature region to which the trajectory point belongs. Through the aforementioned road feasibility constraints, candidate perturbation points can be effectively prevented from falling into buildings, water bodies, or other impassable areas, thereby improving the spatial rationality of the trajectory after privacy perturbation. This road feasibility constraint only applies to the candidate point selection process in the differential privacy post-processing stage and does not depend on the original trajectory data or generation model parameters; therefore, it does not weaken the existing differential privacy protection guarantees.

[0066] Candidate set size: for each p in the perturbation domain j Generate n candidate confusion points, denoted as Because differential privacy pairs with post-processing closures, this step does not change the privacy guarantees.

[0067] In the implementation process, the construction method includes: (1) calculating the covariance matrix Σ of the feature region point set; (2) solving for the principal direction vector and scale; (3) constructing an elliptical perturbation domain E consistent with the principal direction of the trajectory. k (4) Based on KDE or GMM as priors and road constraints, candidate confusion points are generated from within the ellipse.

[0068] The candidate points generated by this module are more consistent with the actual behavior distribution and can effectively maintain the directional continuity and local shape of the trajectory.

[0069] In this implementation, the candidate point generation simultaneously satisfies the constraints of the elliptical perturbation domain, the regional prior distribution, and the road feasibility, thereby improving the traffic semantic rationality of the perturbation points. Meanwhile, the candidate selection falls under the category of differential privacy post-processing and does not change the existing privacy guarantees.

[0070] In the above embodiments, preferably, the preset utility function includes a regional consistency term and a local consistency term; the regional consistency term E i The regional consistency score is calculated based on the distance between candidate confusion points and the region centroid, as well as the minimum distance from the candidate confusion point to the nearest road segment in the road network, to constrain the candidate points to be consistent with the original distribution in terms of regional structure and road semantics; the local consistency term C o The local consistency score is calculated based on the direction difference and curvature difference between the candidate confusion point and the previous and next trajectory points of the trajectory point to be replaced, so as to constrain the replacement trajectory to maintain coherence in terms of local direction continuity and curvature smoothness; the utility score is a weighted combination of the regional consistency score and the local consistency score.

[0071] The indexing mechanism is used to index the utility score based on the point-level privacy budget, obtain the sampling probability of candidate confusion points, and determine the replacement points by sampling according to the sampling probability, and perform replacement perturbation on the corresponding trajectory points in the synthetic trajectory dataset.

[0072] In this implementation, under point-level budget control, sampling is performed through a utility-driven exponential mechanism, which balances privacy strength and structural fidelity, so that the perturbation results maintain higher analyzability in terms of regional structure, road semantics, directional continuity and curvature smoothness.

[0073] In the above embodiments, preferably, the regional consistency term E i This includes the centroid consistency sub-item and the road consistency sub-item. The centroid consistency sub-item is used to calculate the centroid consistency score based on the distance between the candidate confusion point and the region centroid. It represents the distance that the disturbance point should maintain from the region centroid, which is the basis for maintaining the region structure. Its definition is: The road consistency sub-item is used to calculate the road consistency score based on the minimum distance from the candidate disturbance point to the nearest road segment in the road network. It measures the spatial fit between the candidate disturbance point and the road network, and is defined as follows: The regional consistency score is a weighted combination of the centroid consistency score and the road consistency score, with the scores participating in the combination being normalized, as shown below: .

[0074] Local consistency term C oThis includes a distance consistency sub-item and a shape consistency sub-item; the distance consistency sub-item is used to calculate the distance consistency score based on the spatial distance between the candidate confusion point and the trajectory point to be replaced, determining the closeness of the candidate point to the origin, and is defined as follows: To avoid abrupt changes in direction or abnormal curvature in the trajectory after perturbation, a shape consistency sub-item is added. This sub-item is used to calculate the direction consistency score based on the difference in direction vectors formed by the candidate confusion point and the previous and subsequent trajectory points, respectively. The curvature coherence score is calculated based on the second-order difference between the previous trajectory point, the candidate confusion point, and the subsequent trajectory point. Its definition is as follows: ;in, Describe the consistency between the direction vector of the previous trajectory and the direction after the disturbance. Describes the coherence of the second-order difference (curvature) of the trajectory; The internal weights of the two shape features.

[0075] The shape consistency score is a weighted combination of the orientation consistency score and the curvature coherence score, and is expressed as follows: Local consistency score is a weighted combination of distance consistency score and shape consistency score, and is expressed as follows: .

[0076] Each weighted combination uses a preset weight. and The scores of the combined data are normalized. Specifically, the multi-factor scoring function considers behavioral characteristics such as density, direction, and curvature, making the final confusion point selection more consistent with behavioral logic and spatiotemporal structure. This ensures that the perturbed trajectory still has a real movement pattern, satisfying privacy constraints while maximizing usability.

[0077] Furthermore, regarding the calculation of sampling probability, based on the definition of the exponential mechanism, the probability of the sensitivity of the scoring mechanism can be derived using the following equation: Finally, the sampling probability of generating confused points in the perturbation domain of each trajectory point is calculated. According to the definition of the exponential mechanism, the point to be confused, p... cv The probability of being selected as a trajectory point confusion point can be calculated as follows: Expand into the normalized complete form: The original trajectory point p is calculated using the trajectory confusion function in the above formula. i j With the confusion point p generated in each elliptical perturbation domain cv The probability of confusion P(p cv |p i j Then randomly select a p. cvAs replacement trajectory points, all trajectory points can be replaced and then used for publishing.

[0078] In this implementation, by decomposing and normalizing weighted components, regional structure preservation, road semantic fit, replacement amplitude control, and local shape smoothing layering are incorporated into the scoring system, improving the stability of the index mechanism sampling and the structural continuity of the perturbed trajectory. By introducing road feasibility constraints and a road consistency scoring mechanism in the differential privacy post-processing stage, perturbed trajectory points are effectively prevented from falling into non-road or impassable areas, significantly improving the traffic semantic rationality and engineering usability of the published trajectory while ensuring strict differential privacy protection.

[0079] This invention also proposes a user trajectory privacy protection system that integrates a generative model and three-stage personalized differential privacy, used to implement the user trajectory privacy protection method that integrates a generative model and three-stage personalized differential privacy as disclosed in any of the above embodiments. The system includes: a trajectory acquisition module for acquiring a real user trajectory dataset, which includes a trajectory sequence composed of multiple trajectory points; a trajectory preprocessing module connected to the trajectory acquisition module for extracting and preprocessing trajectory features from the real user trajectory dataset and outputting a standardized trajectory representation; a synthetic trajectory generation module connected to the trajectory preprocessing module for training a deep generative model based on the standardized trajectory representation and generating a synthetic trajectory dataset from the deep generative model; a feature region extraction module connected to the synthetic trajectory generation module for extracting feature regions from the synthetic trajectory dataset to obtain multiple feature regions and determining the centroid of each feature region; and a privacy budget allocation module connected to the feature region extraction module for allocating privacy budgets to the synthetic trajectory dataset. The trajectory points in the trajectory dataset undergo a three-stage personalized privacy budget allocation to obtain a point-level privacy budget corresponding to each trajectory point. A perturbation domain construction module, connected to the feature region extraction module, calculates the covariance matrix for the trajectory point set within each feature region and determines the elliptical principal axis direction and axis length scale parameters based on the covariance matrix, thereby constructing an adaptive elliptical perturbation domain. A confusion point generation and filtering module, connected to the perturbation domain construction module, generates a candidate confusion point set within the adaptive elliptical perturbation domain and applies spatial feasibility constraints to the candidate confusion point set for filtering. A utility scoring and sampling module, connected to the confusion point generation and filtering module, calculates the utility score of the candidate confusion points based on the point-level privacy budget and a preset utility function, and uses an exponential mechanism to sample replacement points from the candidate confusion point set according to the utility score. A perturbation replacement output module, connected to the utility scoring and sampling module, performs replacement perturbations on the corresponding trajectory points in the synthetic trajectory dataset to generate and output a privacy-preserving trajectory dataset.

[0080] During implementation, highly practical privacy-preserving trajectory data is generated without exposing real trajectory information. Using real user trajectories as input, deep neural networks are used to learn the behavioral patterns of trajectories in multiple dimensions such as space, time, and category. Convolutional neural networks are used to extract local temporal features, and LSTM networks are combined to model the long-range dependencies of trajectories. Finally, a generative adversarial network (GAN) is used to build an adversarial learning mechanism between real and synthetic trajectories, thereby generating high-quality synthetic trajectories that closely approximate real data in overall distribution.

[0081] After generating the trajectory, to further enhance privacy protection and ensure strict mathematical privacy guarantees during trajectory publishing, the local behavioral patterns of the trajectory are analyzed based on the DBSCAN density clustering algorithm to identify various density feature regions, thereby constructing the local feature domain of the trajectory structure. Subsequently, a three-stage personalized differential privacy budget allocation strategy is adopted: the first stage allocates a region-level privacy budget based on the region density to achieve differentiated protection for hotspot and sensitive regions; the second stage allocates a point-level privacy budget based on the distance between the trajectory point and the region centroid and the user-specified sensitivity; the third stage uniformly adjusts the privacy budget based on the global trajectory density, so that the protection strength takes into account both local and global characteristics.

[0082] After the privacy budget is allocated, an adaptive perturbation domain construction method is adopted. By calculating the covariance matrix of the feature regions, an elliptical perturbation region is formed, ensuring that the perturbation direction is consistent with the main trajectory direction, thereby avoiding the trajectory structure distortion problem caused by the traditional circular perturbation domain. Subsequently, candidate confusion points are generated within the perturbation domain by combining kernel density estimation (KDE) or Gaussian mixture model (GMM) to improve the spatial rationality and structural consistency of the confusion points.

[0083] To maximize trajectory usability while satisfying differential privacy, a comprehensive utility scoring function is employed, incorporating four factors into the scoring model: regional density consistency of candidate points, distance from the original point, continuity of motion direction, and curvature smoothness. An exponential mechanism is used to sample from the candidate point set, ultimately outputting trajectory points that meet differential privacy requirements, thus generating a privacy-preserving trajectory with well-maintained overall structure and behavior patterns.

[0084] In this implementation, the end-to-end closed loop of generation, clustering, budgeting, perturbation domain, candidate, scoring, sampling, replacement and release is solidified in a modular pipeline manner, which reduces the complexity of engineering integration and improves process automation and reusability.

[0085] According to the user trajectory privacy protection system with fusion generation model and three-stage personalized differential privacy disclosed in the above embodiments, the functions to be implemented by each module are consistent with the steps of the user trajectory privacy protection method with fusion generation model and three-stage personalized differential privacy disclosed in the above embodiments. In the implementation process, refer to the above embodiments for operation, and will not be repeated here.

[0086] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A user trajectory privacy protection method that integrates generative models and three-stage personalized differential privacy, characterized in that, include: A real user trajectory dataset is acquired, and trajectory features are extracted and preprocessed to obtain a standardized trajectory representation. The real user trajectory dataset includes trajectory sequences composed of multiple trajectory points. A deep generation model is trained based on the standardized trajectory representation, and the deep generation model generates a synthetic trajectory dataset. Feature regions are extracted from the synthetic trajectory dataset to obtain multiple feature regions, and the centroid of each feature region is determined. A three-stage personalized privacy budget allocation is performed on the trajectory points in the synthetic trajectory dataset to obtain a point-level privacy budget corresponding to each trajectory point. For each feature region... The covariance matrix of the trajectory point set is calculated, and the direction of the ellipse principal axis and the axis length scale parameters are determined based on the covariance matrix to construct an adaptive elliptical perturbation domain. A candidate confusion point set is generated within the elliptical perturbation domain, which includes multiple candidate confusion points. Spatial feasibility constraints are applied to the candidate confusion point set for screening. The utility score of the candidate confusion points is calculated based on the point-level privacy budget and a preset utility function. An exponential mechanism is used to sample and determine replacement points from the candidate confusion point set according to the utility score. The corresponding trajectory points in the synthetic trajectory dataset are then perturbed to replace the perturbation points, and a privacy-preserving trajectory dataset is output.

2. The user trajectory privacy protection method based on the fusion generation model and three-stage personalized differential privacy as described in claim 1, characterized in that, The specific process of trajectory feature extraction and preprocessing of the real user trajectory dataset includes: cleaning and sorting the trajectory sequence by time, and padding the trajectory sequence to unify the trajectory length, while setting mask identifiers for the padding positions; normalizing the latitude and longitude information in the trajectory sequence and converting the latitude and longitude information into coordinate offsets relative to the centroid of the trajectory point set; and performing one-hot encoding on the time attribute and location category attribute in the trajectory sequence to obtain the standardized trajectory representation.

3. The user trajectory privacy protection method based on the fusion generation model and three-stage personalized differential privacy as described in claim 2, characterized in that, The deep generative model is a CNN-LSTM-GAN model, which includes a trajectory attribute embedding layer, a convolutional feature extraction layer, a long short-term memory network layer, and a generator and discriminator of a generative adversarial network. The trajectory attribute embedding layer is used to map the standardized trajectory representation to a high-dimensional semantic space. The convolutional feature extraction layer is used to extract local movement trend features. The long short-term memory network layer is used to model the time-dependent features of the trajectory. The generator is used to generate synthetic trajectory data, and the discriminator is used to distinguish between real and fake input trajectories. The generator and the discriminator are trained adversarially to obtain the deep generative model.

4. The method according to claim 1, characterized in that, The feature region extraction uses the density clustering algorithm DBSCAN to cluster the trajectory points in the synthetic trajectory dataset, and determines each density cluster as the feature region, and calculates the corresponding region centroid based on the spatial coordinates of the trajectory points in each feature region.

5. The user trajectory privacy protection method based on the fusion generation model and three-stage personalized differential privacy as described in claim 4, characterized in that, The specific process of the three-stage personalized privacy budget allocation includes: First, for each feature region of each trajectory, a region-level privacy budget is determined based on the number of trajectory points within the feature region and the total number of trajectory points. Second, based on the distance between each trajectory point within the feature region and the centroid of the corresponding region, and combined with user sensitivity parameters and a preset privacy level list, the region-level privacy budget is allocated to a point-level privacy budget corresponding to each trajectory point. Third, density clustering is performed on the global trajectory point set in the synthetic trajectory dataset to obtain multiple density clusters. A density adjustment coefficient is calculated based on the point density of each density cluster and preset adjustment parameters. The point-level privacy budget of the trajectory points belonging to the corresponding density cluster is then scaled and updated based on the density adjustment coefficient.

6. The user trajectory privacy protection method based on the fusion generation model and three-stage personalized differential privacy as described in claim 5, characterized in that, The specific process of constructing the adaptive elliptical perturbation domain includes: calculating the covariance matrix of the spatial coordinates of the trajectory points within the feature region, and determining the principal axis direction and semi-axis scale parameters of the ellipse based on the covariance matrix; using the centroid of the region as the center of the ellipse, and constructing the adaptive elliptical perturbation domain based on the principal axis direction and the semi-axis scale parameters.

7. The user trajectory privacy protection method based on the fusion generation model and three-stage personalized differential privacy as described in claim 6, characterized in that, The specific process of generating a set of candidate confusion points within the elliptical perturbation domain includes: modeling the spatial distribution of trajectory points within the feature region based on kernel density estimation or Gaussian mixture model, and sampling to generate the candidate confusion points within the elliptical perturbation domain; the spatial feasibility constraint includes road feasibility constraint, which is used to calculate the minimum distance from the candidate confusion point to the nearest road segment in the road network based on the road network, and retaining candidate confusion points whose minimum distance is not greater than a preset road buffer threshold.

8. The user trajectory privacy protection method based on the fusion generation model and three-stage personalized differential privacy as described in claim 7, characterized in that, The preset utility function includes a regional consistency term and a local consistency term; the regional consistency term is used to calculate a regional consistency score based on the distance between the candidate confusion point and the centroid of the region and the minimum distance from the candidate confusion point to the nearest road segment in the road network; the local consistency term is used to calculate a local consistency score based on the direction difference and curvature difference between the candidate confusion point and the previous and next trajectory points of the trajectory point to be replaced; the utility score is a weighted combination of the regional consistency score and the local consistency score; the indexing mechanism is used to index the utility score based on the point-level privacy budget to determine the sampling probability of each candidate confusion point, and to sample and determine the replacement point according to the sampling probability.

9. The user trajectory privacy protection method based on the fusion generation model and three-stage personalized differential privacy as described in claim 8, characterized in that, The region consistency item includes a centroid consistency sub-item and a road consistency sub-item; the centroid consistency sub-item is used to calculate the centroid consistency score based on the distance between the candidate confusion point and the region centroid; The road consistency sub-item is used to calculate the road consistency score based on the minimum distance from the candidate confusion point to the nearest road segment in the road network; the regional consistency score is a weighted combination of the centroid consistency score and the road consistency score, wherein the scores participating in the combination are normalized; the local consistency item includes a distance consistency sub-item and a shape consistency sub-item; the distance consistency sub-item is used to calculate the distance consistency score based on the spatial distance between the candidate confusion point and the trajectory point to be replaced; the shape consistency sub-item is used to calculate the direction consistency score based on the direction vector difference formed by the candidate confusion point and the previous trajectory point and the next trajectory point, respectively, and to calculate the curvature coherence score based on the second difference of the previous trajectory point, the candidate confusion point, and the next trajectory point; the shape consistency score is a weighted combination of the direction consistency score and the curvature coherence score; the local consistency score is a weighted combination of the distance consistency score and the shape consistency score, wherein each weighted combination uses a preset weight and the scores participating in the combination are normalized.

10. A user trajectory privacy protection system that integrates a generative model with three-stage personalized differential privacy, characterized in that, A user trajectory privacy protection method for implementing the fusion generative model and three-stage personalized differential privacy as described in any one of claims 1 to 9, comprising: a trajectory acquisition module for acquiring a real user trajectory dataset, the real user trajectory dataset including a trajectory sequence composed of multiple trajectory points; a trajectory preprocessing module connected to the trajectory acquisition module for extracting and preprocessing trajectory features from the real user trajectory dataset and outputting a standardized trajectory representation; a synthetic trajectory generation module connected to the trajectory preprocessing module for training a deep generative model based on the standardized trajectory representation and generating a synthetic trajectory dataset by the deep generative model; a feature region extraction module connected to the synthetic trajectory generation module for extracting feature regions from the synthetic trajectory dataset to obtain multiple feature regions and determining the centroid of each feature region; and a privacy budget allocation module connected to the feature region extraction module for performing three-stage personalized privacy on the trajectory points in the synthetic trajectory dataset. The system comprises the following modules: a budget allocation module to obtain a point-level privacy budget corresponding to each trajectory point; a perturbation domain construction module connected to the feature region extraction module to calculate the covariance matrix for the trajectory point set within each feature region, and to determine the elliptical principal axis direction and axis length scale parameters based on the covariance matrix, thereby constructing an adaptive elliptical perturbation domain; a confusion point generation and filtering module connected to the perturbation domain construction module to generate a candidate confusion point set within the adaptive elliptical perturbation domain, and to filter the candidate confusion point set by applying spatial feasibility constraints; a utility scoring and sampling module connected to the confusion point generation and filtering module to calculate the utility score of the candidate confusion points based on the point-level privacy budget and a preset utility function, and to sample replacement points from the candidate confusion point set according to the utility score using an exponential mechanism; and a perturbation replacement output module connected to the utility scoring and sampling module to perform replacement perturbation on the corresponding trajectory points in the synthetic trajectory dataset to generate and output a privacy-preserving trajectory dataset.

Images