An interest region-oriented local differential privacy trajectory publishing algorithm

Abstract

This invention discloses a local differential privacy trajectory publishing algorithm oriented towards regions of interest, comprising the following steps: Step 1, extracting a set of safe regions from the original user trajectory; Step 2, traversing each trajectory point in the original user trajectory in chronological order, and further dividing the trajectory points falling into the safe regions to obtain discrete candidate domains and nearest candidate points; Step 3, performing distance-layered restricted random response within the discrete candidate domains to generate perturbation points, wherein the perturbation points and trajectory points not falling into any safe regions constitute the published trajectory. This invention features a finer-grained trade-off between privacy strength and trajectory availability.

Description

Technical Field

[0001] This invention relates to the field of rescue equipment technology, and more specifically, to a local differential privacy trajectory publishing algorithm oriented towards regions of interest. Background Technology

[0002] With the rapid development of location-based network services, user trajectory data is widely used in navigation, location-based social networking and other scenarios. However, trajectory data has significant spatiotemporal correlation and re-identifiability, which can easily leak users' sensitive location information.

[0003] Existing methods for protecting trajectory privacy mainly include trajectory generalization, trajectory suppression, trajectory encryption, dynamic pseudonymization, and differential privacy. While these methods can reduce the risk of privacy leakage to some extent, they typically rely on specific assumptions and may affect service accuracy and data availability. Especially in trajectory publishing related to Areas of Interest (AoI), existing methods lack effective constraints on dwell structure and trajectory continuity, and random perturbations can easily lead to trajectory deformation and long jumps. Therefore, it is difficult to protect AoI privacy while ensuring trajectory availability. Summary of the Invention

[0004] The purpose of this invention is to design and develop a local differential privacy trajectory publishing algorithm for regions of interest. By using an algorithm structure that first structures and then perturbs, the distortion propagation caused by unconstrained randomization is reduced, the usability of the trajectory is improved, and the ability to maintain the dwell structure is enhanced.

[0005] The technical solution provided by this invention is as follows:

[0006] A local differential privacy trajectory publishing algorithm oriented towards regions of interest includes the following steps:

[0007] Step 1: Collect raw user trajectories including location coordinates and corresponding timestamps;

[0008] Step 2: Extract a set of safe zones from the original user trajectory;

[0009] Step 3: Traverse each trajectory point in the original user trajectory in chronological order, and further divide the trajectory points that fall into the safe area into regions to obtain discrete candidate domains and nearest candidate points;

[0010] Step 4: Perform distance-stratified restricted random response within the discrete candidate domain to generate perturbation points. These perturbation points and trajectory points that do not fall into any safe zone constitute the published trajectory.

[0011] Preferably, step two specifically includes:

[0012] Extract the set of dwell intervals, the set of interest areas, and the set of safe areas sequentially from the original user trajectory.

[0013] Preferably, the extraction of the set of safe zones specifically includes the following steps:

[0014] Step 1: Preprocess the original user trajectory;

[0015] Step 2: Process the preprocessed user trajectory using a hidden semi-Markov model to obtain a state sequence;

[0016] Step 3: Scan the state sequence in order, extract the continuous dwell state intervals as dwell intervals, and use the dwell intervals that satisfy the dwell constraints as elements to form a set of dwell intervals;

[0017] Step 4: Compress the dwell intervals in the dwell interval set into pairs, and use the pairs as elements to form a compressed set;

[0018] Step 5: Perform density clustering on the compressed set using DBSCAN, grouping spatially adjacent dwelling positions into regions of interest, and constructing a set of regions of interest using these regions as elements.

[0019] Step 6: Buffer each region of interest outward to obtain a set of safe regions.

[0020] Preferably, the preprocessing includes:

[0021] The original user trajectories are sorted by time and then cleaned.

[0022] Preferably, the hidden semi-Markov model specifically includes:

[0023] Hidden state set: The set of user trajectory states;

[0024] State transition probability: ;

[0025] Launch probability: ;

[0026] Display duration distribution: ;

[0027] The observation sequence is obtained by extracting trajectory point information in chronological order from the preprocessed user trajectory. The hidden semi-Markov model models the trajectory state through state transition probabilities, emission probabilities, and explicit duration distributions, and decodes the state of the observation sequence to obtain the first... The state corresponding to each trajectory point ;

[0028] In the formula, Represents the first intermediate state variable. Represents the second intermediate state variable, and , Indicates the first The trajectory state of each trajectory point is determined by... Transfer to The probability of; Indicates that the i-th trajectory state is When its corresponding observation item The probability of occurrence Indicates trajectory state continued The probability of a series of consecutive trajectory points Represents the observation sequence The Middle One observation item, For the first The instantaneous velocity of each trajectory point For the first The spatial distance between each trajectory point and the previous trajectory point Indicates a stationary state. Indicates the movement status. Indicates the duration of the state. For the first The state of each trajectory point For the first The instantaneous velocity of each trajectory point.

[0029] Preferably, the dwell constraint is:

[0030] and ;

[0031] in, The duration of the stay interval. As a minimum stay duration constraint, The spatial radius of the trajectory points within the dwell interval. This is a constraint on the maximum dwell radius.

[0032] Preferably, step three specifically includes:

[0033] For trajectory points that fall into the safe zone, construct discrete candidate domains within the corresponding safe zone. Discrete candidate domain The candidate point with the smallest spatial distance from the current trajectory point is the nearest candidate point.

[0034] Wherein, the discrete candidate domain Within its corresponding security area and the first Candidate points are formed by the spatial distance between trajectory points not exceeding a preset threshold.

[0035] Preferably, the generation of the disturbance point in step four includes the following steps:

[0036] Step a: Divide the discrete candidate domain into A ring-shaped distance layer ;

[0037] Step b: Obtain perturbation points by sampling in different distance layers according to the sampling rules.

[0038] Preferably, in step a... The radial spacing between the annular distance layers is equal.

[0039] Preferably, the sampling rule is as follows:

[0040] If the initial random number Then in the discrete candidate domain The disturbance point is obtained by uniform random sampling in the middle; otherwise, random numbers are generated again.

[0041] If the random number is generated again If so, the nearest candidate point is selected as the perturbation point;

[0042] If the random number is generated again and Then, any layer in the other annular distance layers can be selected and the perturbation points are obtained by uniformly sampling within the layer;

[0043] in, As the probability of being at the bottom, Select the probability for the nearest layer.

[0044] The beneficial effects of this invention are as follows:

[0045] (1) The local differential privacy trajectory publishing algorithm designed and developed by this invention for regions of interest follows the design paradigm of first structuring and then perturbation, and combines privacy mechanism with travel behavior structure to reduce the distortion propagation caused by unconstrained randomization.

[0046] (2) The local differential privacy trajectory publishing algorithm designed and developed by the present invention for the region of interest is used by the terminal to model and detect the dwell structure of a single trajectory, and to characterize the continuity and stability of the dwell state by explicit duration modeling, thereby reducing false detection and fragmented segmentation caused by short pauses and positioning noise, and thus obtaining a more reliable set of dwell points and dwell segment boundaries.

[0047] (3) The local differential privacy trajectory publishing algorithm designed and developed by the present invention for interest regions automatically constructs interest regions based on the dwell detection results, regionalizes the dwell semantics and individual activity range, and further determines the local security areas that need to be protected by interest regions, so that the protected objects are transformed from the entire trajectory points to the sensitive areas related to dwelling, reducing the invalid disturbance to non-sensitive road sections.

[0048] (4) The local differential privacy trajectory publishing algorithm designed and developed by this invention for regions of interest performs region determination on trajectory points and performs local randomization publishing only on points that fall into the safe region: a discrete candidate range is constructed within the safe region, and a distance-layered random response is adopted to make the output more biased towards local neighbors, thereby constraining the spatial span of the disturbance from the mechanism level, so as to suppress the deformation accumulation caused by cross-regional jumps and far jumps, and avoid the deformation from spreading and expanding along the time series; when the candidate range is limited or the number of candidate points is insufficient, the bottom regularization is used to ensure that the output fully supports the randomization mechanism and maintain its effectiveness and stability.

[0049] (5) The present invention designs and develops a local differential privacy trajectory publishing algorithm for regions of interest. The algorithm forms a spatiotemporal constraint mechanism by using a dwell structure constrained by duration and spatial perturbation of the safe area. Under the premise of satisfying local differential privacy, it improves the availability of the trajectory and enhances the ability to maintain the dwell structure. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the local differential privacy trajectory publishing algorithm for regions of interest described in this invention.

[0051] Figure 2 This is a bar chart illustrating the distribution of AoI numbers in the GeoLife dataset described in this invention.

[0052] Figure 3 This is a bar chart illustrating the distribution of AoI numbers in the Rome Taxi dataset described in this invention.

[0053] Figure 4 This is a schematic diagram showing the comparison results of DTWS with different privacy budgets under the GeoLife dataset AoI=6 described in this invention.

[0054] Figure 5 This is a schematic diagram showing the comparison results of DTWS with different privacy budgets under the GeoLife dataset AoI=7 described in this invention.

[0055] Figure 6 This is a schematic diagram showing the comparison results of DTWS with different privacy budgets under the GeoLife dataset AoI=8 described in this invention.

[0056] Figure 7This is a schematic diagram showing the comparison results of DTWS with different privacy budgets under the Rome Taxi dataset AoI=3 described in this invention.

[0057] Figure 8 This is a schematic diagram showing the comparison results of DTWS with different privacy budgets under the Rome Taxi dataset AoI=4 described in this invention.

[0058] Figure 9 This is a schematic diagram showing the comparison results of DTWS with different privacy budgets under the Rome Taxi dataset AoI=5 described in this invention. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0060] like Figure 1 As shown, the local differential privacy trajectory publishing algorithm (ST-LDPAoI) for regions of interest provided by this invention includes:

[0061] Step 1: Collect raw user trajectories including location coordinates and corresponding timestamps;

[0062] In this embodiment, the original user trajectory is collected through a positioning device or positioning service.

[0063] Step 2: From the original user trajectory Extract the set of dwell intervals Interest Area Sets and safe zone collection Specifically, it includes:

[0064] Step 1, for Preprocessing:

[0065] Will Sort and clean by time. For the first The position of each trajectory point For the first Timestamps of each trajectory point;

[0066] In this embodiment, the cleaning is to delete abnormal speed points and timestamp inconsistencies. Abnormal speed points refer to points with abnormal speeds greater than an abnormal speed threshold, and timestamp inconsistencies refer to trajectory points whose timestamp intervals with adjacent trajectory points exceed a time break threshold.

[0067] In this embodiment, the abnormal speed threshold is 180 km / h, and the time break threshold is 30 min;

[0068] Step 2: Process the preprocessed user trajectory using a Hidden Semi-Markov Model (HSMM) to obtain a state sequence. Specifically:

[0069] Let the set of hidden states be the set of trajectory states, which is initially empty;

[0070] Hidden semi-Markov models are typically characterized by three types of parameters:

[0071] 1) State transition probability ;

[0072] 2) Launch probability ;

[0073] 3) Explicit duration distribution ;

[0074] The observation sequence is obtained by extracting trajectory point information in chronological order from the preprocessed raw user trajectory. The hidden semi-Markov model models the trajectory state through state transition probabilities, emission probabilities, and explicit duration distributions, and decodes the state of the observation sequence to obtain the first... The state corresponding to each trajectory point ;

[0075] in, Represents the first intermediate state variable. Represents the second intermediate state variable, and , Indicates the first The trajectory state of each trajectory point is determined by... Transfer to The probability, i.e. the transition relationship between trajectory states; Indicates that the i-th trajectory state is At that time, its corresponding observation item The probability of occurrence, i.e., the relationship between trajectory state and corresponding observation; Indicates trajectory state continued The probability of a continuous trajectory point, i.e. the duration feature of the trajectory state; Represents the observation sequence The Middle One observation item, For the first Instantaneous velocity (m / s) of each trajectory point. For the first Spatial distance (m) between each trajectory point and the previous trajectory point. Indicates a stationary state. Indicates the movement status. This indicates the duration of a state, that is, the number of trajectory points through which the state is maintained continuously. For the first The state of each trajectory point For the first The instantaneous velocity of each trajectory point For the first The time interval between each trajectory point and the previous trajectory point;

[0076] In this embodiment, the state transition probability is preset as follows:

[0077]

[0078] In the launch probability, the instantaneous velocity and spatial displacement are relatively small in the preset stationary state and relatively large in the moving state, which can be expressed as:

[0079]

[0080] In the formula, It follows a normal distribution. This represents the mean, and its unit is consistent with the corresponding observed variable. It represents the standard deviation, and its unit is consistent with the corresponding observed variable; It represents the variance, and its unit is the square of the corresponding observed variable unit.

[0081] Step 3: Scan the state sequence in order. , to make continuous Interval extraction as dwell interval The set of dwell intervals is constructed by using the dwell intervals that satisfy the dwell constraints as elements. ;

[0082] in, and They represent the first The start and end indices of each dwell interval. , , For the first The state sequence corresponding to each trajectory point ,here Indicates the total number of time intervals;

[0083] The dwell constraint is: and ,in, The duration of the stay interval. As a minimum stay duration constraint, The spatial radius of the trajectory points within the dwell interval. Maximum dwell radius constraint;

[0084] The minimum stay duration constraint and the maximum stay radius constraint are set based on experience. In this embodiment, in scenarios such as residential communities and office parks, they can be... Set the time to 5-10 minutes; for larger areas such as shopping malls and hospital campuses, the time spent in these areas should be limited to 5-10 minutes. A range of 80m to 150m can be taken;

[0085] Step 4: Set up the dwelling intervals The dwell interval in the middle is compressed into a binary tuple. Compact sets are formed using pairs as elements. ;

[0086] in, For the first Each rest area The representative position is obtained by averaging the position coordinates of all trajectory points within that stopping interval. Duration;

[0087] Step 5: Compress the set DBSCAN is used for density clustering, grouping spatially adjacent dwelling locations into regions of interest, and using these regions of interest as elements to form a set of regions of interest. ;

[0088] Step 6: Buffer each region of interest outwards to obtain a set of safe regions. :

[0089] For each region of interest, the boundary is expanded outwards by a preset distance to obtain the corresponding safe zone. , to secure the area As elements, they constitute a set of safe zones. Among them, the larger the preset distance, the stronger the privacy protection but the lower the data accuracy; the smaller the preset distance, the higher the data accuracy but the weaker the privacy protection.

[0090] In this embodiment, the preset distance is 50m to 120m.

[0091] Step 3: Traverse each trajectory point in the original user trajectory in chronological order, and further divide the trajectory points that fall into the safe area into regions to obtain discrete candidate domains and nearest candidate points;

[0092] If a trajectory point falls into multiple safe zones simultaneously, the unique zone of belonging is determined according to the principle of closest distance to ensure consistent subsequent processing.

[0093] The trajectory points that fall into the safe zone will be further divided into regions, specifically including:

[0094] For trajectory points that fall into the safe zone, construct discrete candidate domains within the corresponding safe zone. Discrete candidate domain The trajectory points in the data are candidate points. For the current trajectory point, calculate the spatial distance between it and each candidate point, and determine the candidate point with the smallest distance as the nearest candidate point after mapping the trajectory point.

[0095] The discrete candidate domain It is a subset of the safe zone where the current location is located, consisting of the elements within that safe zone and the first element. Candidate points whose spatial distance between trajectory points does not exceed a preset threshold are constituted;

[0096] The safe region is used to define the upper bound of the disturbance range, and the discrete candidate domain is used for other purposes. Used to further define the actual selectable local disturbance locations within the safe area, the preset threshold is determined based on the positioning error level and the local reachable range;

[0097] In this embodiment, the preset threshold is set to 30m to 100m.

[0098] Step 4: Perform distance-stratified constrained random response within the discrete candidate domain to generate perturbation points. These perturbation points, along with trajectory points that do not fall into any safe zone, constitute the published trajectory. Specifically, it includes the following steps:

[0099] Step a: Using the nearest candidate point as the center, divide the discrete candidate domain into sections according to spatial distance from nearest to farthest. A ring-shaped distance layer, denoted as , The preset number of layers, and The radial spacing between each annular spacer layer is equal;

[0100] In this embodiment, the preset number of layers is 3 to 5;

[0101] Step b, Generation of disturbance points:

[0102] If the initial random number Then directly in the discrete candidate domain The disturbance point is obtained by uniform random sampling in the middle; otherwise, random numbers are generated again.

[0103] If the random number is generated again If so, the nearest candidate point is selected as the perturbation point;

[0104] If the generated random number and Then, arbitrarily select one of the other layers and uniformly sample within the layer to obtain the perturbation point;

[0105] in, As the probability of being at the bottom, The value should be a small constant that is greater than 0 and much less than 1. Its purpose is to ensure that all candidate points have a non-zero output probability while minimizing the impact on the nearest neighbor preference distribution.

[0106] Choose the probability for the nearest layer, and satisfy:

[0107] ;

[0108] In the formula, Budget for privacy.

[0109] The specific flow of the local differential privacy trajectory publishing algorithm for regions of interest described in this invention is as follows:

[0110]

[0111] As shown in Algorithm 1, the end-to-end publishing process of ST-LDPAoI is given. First, Algorithm 2 is called to extract the set of dwell intervals S, the set of regions of interest A, and the set of safe regions R from the original trajectory T, and the published trajectory is initialized to an empty set (lines 1 and 2). Then, each trajectory point in T is traversed in chronological order. Determine its location Whether a point falls into a safe region SR; if it falls into multiple regions simultaneously, a unique region is determined according to preset rules to ensure consistency in subsequent processing (lines 3-5); for points falling into safe regions, a discrete candidate region is constructed within SR. Then Mapped to the nearest candidate point in the candidate domain As input to Algorithm 3, the continuous spatial perturbation is transformed into randomization over a discrete candidate domain (lines 6 and 7); then Algorithm 3 is called to address the privacy budget. Stratification threshold and the probability of being at the bottom Generating perturbation points under constraints (Line 8); Finally, Write to the release history; if If it is not within any safe zone, then retain the original point. Repeat the above process until the traversal is complete, and output the publishing trajectory. (Lines 9-14).

[0112]

[0113] As shown in Algorithm 2, it is used to extract the dwell structure from the original trajectory and generate the region of interest and the safe region. First, the trajectory is processed... Perform sorting and anomaly cleaning, then construct the observation sequence. , The original trajectory T was extracted in chronological order. Each observation was used to characterize the motion features of the trajectory points, and the results were analyzed under HSMM based on the observation sequence. Determine the state of each sampling point to obtain the state sequence. (Lines 1-3); then scan the state sequence sequentially, putting consecutive... Interval extraction as dwell interval (Lines 4-9); For each dwell interval, use the duration and the radius within the segment The process involves filtering out short pauses and pseudo-pauses with excessively large spatial dispersion. Dwell intervals that meet the criteria are retained in the set. (Lines 10-17); in obtaining Then, each dwell interval is compressed into a binary tuple. , forming a set This compression step transforms the dwell interval from a boundary representation to a location and duration representation, reducing the complexity of subsequent clustering processing while preserving the main dwell features. These features serve as input for subsequent clustering to construct regions of interest (line 18); subsequently, ... The representative locations in the data are clustered using DBSCAN density clustering, grouping spatially adjacent locations into regions of interest to obtain a set of regions of interest. Then for each Buffer outwards to obtain a set of safe zones. This is provided for subsequent disturbance modules to use (lines 19-21).

[0114]

[0115] As shown in Algorithm 3, a restricted random response with distance hierarchical structure is used to perform the response within the candidate domain, given candidate points. Candidate domain Privacy Budget and the probability of being at the bottom Then, the algorithm first uses For reference, according to distance Divided into several distance layers (Line 1); then perform the bottom-laying step first: if Then directly in the entire candidate domain uniformly selected disturbance points (Lines 2-4); If the bottom layer is not triggered, a random response is performed on the layer index (line 5); then, the nearest layer is selected with probability α, otherwise a layer number sel is selected uniformly from the remaining layers, and in the corresponding layer... Perturbation points are obtained by uniform sampling within the inner area. (Lines 6-12) This can both restrict long-distance jumps and meet the randomization requirements of local differential privacy.

[0116] To evaluate the performance of the algorithm, this invention selected two publicly available GPS trajectory datasets, GeoLife and Rome Taxi, for comparative experiments. GeoLife contains personal mobile trajectory data, covering 182 users and including various modes of travel such as walking, cycling, and driving. The tracks show significant stops, short pauses, and discontinuities. Rome Taxi contains Rome city taxi tracks, covering approximately 320 taxis, reflecting relatively continuous motorized travel characteristics under road network constraints. Both datasets use WGS-84 latitude and longitude coordinates and timestamps to construct point sequences, suitable for evaluating the shape preservation and stability of stop structures of the tracks after disturbances from the perspective of terminal-side distribution.

[0117] To ensure the reproducibility and scalability of the experimental process, this invention employs a unified preprocessing strategy for both datasets. First, trajectory segments are filtered by user or vehicle. The original trajectories are deduplicated by timestamp and illegal points and abnormal records are deleted. Then, instantaneous speeds are calculated based on the spatial distance and time difference between adjacent points. Jump segments exceeding 180 km / h are broken at these points to mitigate the impact of non-realistic long jumps caused by positioning drift, missing sampling, or abnormal reporting on subsequent evaluations. Furthermore, a time break threshold of 30 minutes is set to divide the long sequence into relatively continuous trajectory segments, making each segment more consistent with a single trip or continuous driving process in terms of time sequence, facilitating subsequent unified calculation of indicators and group comparison.

[0118] Based on this, the present invention represents the main activity area of ​​each dataset in a gridded manner to reduce the noise sensitivity caused by continuous latitude and longitude space and to unify the distance metric. Specifically, the bounding box is determined by the latitude and longitude [1%, 99%] quantiles, a square grid with a side length of 200m is constructed, and each sampling point is mapped to the corresponding grid to obtain a grid sequence representation. At the same time, the number of accessed grids and the number of grid jumps are counted to characterize the spatial coverage and movement complexity of the trajectory, and to serve as the basis for subsequent screening and grouping. In order to avoid the bias of the index caused by insufficient information or near-static segments, the present invention sets that the effective trajectory segment must meet the requirement of a duration of more than 10 minutes and meet the minimum spatial coverage requirement to eliminate segments that are too short, have too small a coverage area, or fluctuate only within a few grids. At the same time, the effective users are required to have more than 10 effective trajectory segments to ensure that each subject has enough samples for stable evaluation. The key parameters of the final filtered dataset are shown in Table 1.

[0119] Table 1 Dataset Information

[0120]

[0121] The effectiveness of ST-LDPAoI in adding noisy trajectories was evaluated on the GeoLife and Rome Taxi datasets. To select suitable trajectory points for protection, the Areas of Interest (AoI) of the trajectory, based on the HSMM proposed earlier, were used to extract the AoI. To verify the applicability and advantages of the method, four representative baselines were selected for comparison: point-level randomization mechanism SDD, differential privacy method TPRSDP based on random sampling, local random response mechanism L-SRR, and randomization mechanism LDPTPM-AoI within the candidate domain. The evaluation metric used was Dynamic Time Warping (DTW) to measure the difference in overall shape between the published trajectory and the original trajectory, and a normalized metric DTWS (DTW similarity measure) was further introduced, defined as:

[0122] ;

[0123] in, and These are the original trajectory and the published trajectory, respectively. and This indicates the number of points on the corresponding trajectory. The smaller the DTWS value, the more similar the two trajectories are, and the better the overall shape is preserved.

[0124] First, the distribution of AoI (Aspect of Interest) counts in the two datasets after processing by Algorithm 2 is statistically analyzed to characterize the complexity of the trajectory dwelling structure. Based on this, the main coverage area and sampling strategy for subsequent DTWS curve comparison are determined. The number of AoIs characterizes the complexity of the trajectory dwelling structure. Generally speaking, a larger number of AoIs means a longer activity chain, more frequent switching between dwelling and moving states, and richer relative positional relationships between dwelling points. In privacy releases, such trajectories are more susceptible to the cumulative effects of random disturbances, manifesting as shape shifts, cross-regional jumps, and other problems. Therefore, grouping by AoI level can stratify the samples according to difficulty, providing a clearer reference for comparing the differences in shape preservation and stability of different mechanisms.

[0125] As shown, the GeoLife dataset exhibits a wider distribution of AoIs, indicating more diverse individual travel patterns and relatively complex stop structures. Based on this distribution characteristic, subsequent DTWS comparisons on GeoLife primarily cover the typical AoI range of 6–8. Within each AoI level, samples and statistical results are organized according to uniform rules to ensure that different methods are evaluated under comparable conditions and to avoid additional bias introduced by differences in sample size or structure. As shown, the Rome Taxi dataset shows a more concentrated distribution of AoIs with a shorter long tail, reflecting that taxi trajectories are more constrained by road networks and operating patterns, with a relatively simple stop structure and fewer cross-regional stop chains. Based on this distribution characteristic, subsequent DTWS curve comparisons on Rome Taxi focus on the main AoI range of 3–5, and similarly employ consistent statistical methods and display standards within each level to objectively reflect the impact of changes in privacy budgets on trajectory utility and to make the comparison conclusions of the two datasets more targeted in complex stop structure scenarios.

[0126] Under the aforementioned AoI grouping settings, the DTWS comparison results of the GeoLife dataset under different privacy budget sequences are shown in the figure. Overall, as the privacy budget increases, the DTWS of all methods gradually decreases, which is a common trend brought about by the weakening of perturbation amplitude. What distinguishes the methods more clearly is the change in AoI level: when the AoI is low, the distance between the curves of each method is relatively limited; when the AoI increases to the mid-to-high level, the DTWS levels of baseline methods such as SDD and TPRSDP shift upwards significantly, indicating that in data with relatively complex dwell structures like GeoLife, conventional randomization is more likely to disrupt the relative stability across AoIs. Positional relationships accumulate into more pronounced global shape shifts. The overall DTWS of L-SRR is lower than that of SDD and TPRSDP, indicating that local random responses can partially alleviate disturbances, but shape shifts across AoIs are still difficult to avoid. In contrast, LDPTPM-AoI introduces an AoI protection mechanism, thus maintaining a lower overall DTWS across all AoI levels. ST-LDPAoI performs best and is smoother with changes in privacy budget, demonstrating that limiting the candidate domain within a safe area and performing distance-layered random responses can effectively suppress far-jump deformation, enabling the published trajectory to maintain stable shape fidelity even under complex dwell structures.

[0127] For the Rome Taxi dataset, the DTWS comparison results in the main interval of AoI=3 to 5 are shown in the figure. RomeTaxi's error is more biased towards local road segment continuity disturbances rather than cross-regional deformation amplification. Under these circumstances, ST-LDPAoI still maintains the lowest DTWS in each AoI level, indicating that the candidate domain restriction and hierarchical randomization under AoI constraints can also reduce the cumulative effect of local offsets in scenarios with strong road network consistency.

[0128] To facilitate the generalization of the overall performance of different methods within the budget range at the dataset level, the DTWS results of the main AoI levels were macro-averaged, and the mean and standard deviation were calculated on the privacy budget sequence. The mean was used to characterize the overall shape distortion level of the method across the entire budget range, while the standard deviation reflected its sensitivity and stability to changes in the privacy budget. The results showed that ST-LDPAoI had the lowest average DTWS and less fluctuation, indicating that limiting the candidate domain by AoI and introducing hierarchical perturbations of nearest neighbor preference can effectively suppress unreasonable jumps, thereby achieving a more stable shape preservation effect. Compared with the baseline LDPTPM-AoI, ST-LDPAoI reduced the average DTWS by 16.61% and 15.50% on GeoLife and Rome Taxi, respectively, indicating that this advantage can be stably reflected under different trajectory structures and scene constraints.

[0129] Table 2 Mean and standard deviation of DTWS for each method

[0130]

[0131] This invention presents a local differential privacy trajectory publishing algorithm oriented towards regions of interest (ROIs). Based on the structured characterization of travel behavior at the terminal side, it first identifies and segments dwelling behaviors within the trajectory, abstracting the semantics of dwelling and the individual's activity range into ROIs. Subsequently, the privacy mechanism explicitly introduces accessibility constraints and candidate domain restrictions based on ROIs. While ensuring that the perturbation output satisfies local differential privacy constraints, it makes the publishing results more consistent with the spatiotemporal continuity and accessibility patterns of real movement, thereby effectively reducing the impact of abnormal jumps on the overall trajectory morphology, dwelling structure, and downstream task analysis (such as dwell point detection, activity range estimation, and statistical inference). Furthermore, the method achieves a finer-grained trade-off between privacy strength and trajectory availability through a hierarchical randomization strategy: it can still suppress excessive deformation when the privacy budget is small, and maintain a more stable utility improvement when the budget increases. Experimental results show that, under the premise of meeting privacy protection requirements, the method described in this invention outperforms the baseline in both stop point detection performance and trajectory shape preservation, and exhibits more stable and consistent utility performance under different privacy budgets. Future work can further expand the modeling of regions of interest and the expression of accessibility constraints for more complex travel patterns and multi-source scenario constraints, explore the impact of trajectory heterogeneity and individual differences on privacy mechanisms, and improve the ability to balance privacy protection and trajectory utility.

[0132] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and embodiments shown and described herein.

Claims

1. A local differential privacy trajectory publishing algorithm for regions of interest, characterized in that, Includes the following steps: Step 1: Collect raw user trajectories including location coordinates and corresponding timestamps; Step 2: Extract a set of safe zones from the original user trajectory; Step 3: Traverse each trajectory point in the original user trajectory in chronological order, and further divide the trajectory points that fall into the safe area into regions to obtain discrete candidate domains and nearest candidate points; Step 4: Perform distance-stratified restricted random response within the discrete candidate domain to generate perturbation points. These perturbation points and trajectory points that do not fall into any safe zone constitute the published trajectory.

2. The local differential privacy trajectory publishing algorithm for regions of interest as described in claim 1, characterized in that, Step two specifically includes: Extract the set of dwell intervals, the set of interest areas, and the set of safe areas sequentially from the original user trajectory.

3. The local differential privacy trajectory publishing algorithm for regions of interest as described in claim 2, characterized in that, The extraction of the set of safe zones specifically includes the following steps: Step 1: Preprocess the original user trajectory; Step 2: Process the preprocessed user trajectory using a hidden semi-Markov model to obtain a state sequence; Step 3: Scan the state sequence in order, extract the continuous dwell state intervals as dwell intervals, and use the dwell intervals that satisfy the dwell constraints as elements to form a set of dwell intervals; Step 4: Compress the dwell intervals in the dwell interval set into pairs, and use the pairs as elements to form a compressed set; Step 5: Perform density clustering on the compressed set using DBSCAN, grouping spatially adjacent dwelling positions into regions of interest, and constructing a set of regions of interest using these regions as elements. Step 6: Buffer each region of interest outward to obtain a set of safe regions.

4. The local differential privacy trajectory publishing algorithm for regions of interest as described in claim 3, characterized in that, The preprocessing includes: The original user trajectories are sorted by time and then cleaned.

5. The local differential privacy trajectory publishing algorithm for regions of interest as described in claim 4, characterized in that, The hidden semi-Markov model specifically includes: Hidden state set: The set of user trajectory states; State transition probability: ; Launch probability: ; Display duration distribution: ; The observation sequence is obtained by extracting trajectory point information in chronological order from the preprocessed user trajectory. The hidden semi-Markov model models the trajectory state through state transition probabilities, emission probabilities, and explicit duration distributions, and decodes the state of the observation sequence to obtain the first... The state corresponding to each trajectory point ; In the formula, Represents the first intermediate state variable. Represents the second intermediate state variable, and , Indicates the first The trajectory state of each trajectory point is determined by... Transfer to The probability of; Indicates that the i-th trajectory state is When its corresponding observation item The probability of occurrence Indicates trajectory state continued The probability of a series of consecutive trajectory points Represents the observation sequence The Middle One observation item, For the first The instantaneous velocity of each trajectory point For the first The spatial distance between each trajectory point and the previous trajectory point Indicates a stationary state. Indicates the movement status. Indicates the duration of the state. For the first The state of each trajectory point For the first The instantaneous velocity of each trajectory point.

6. The local differential privacy trajectory publishing algorithm for regions of interest as described in claim 5, characterized in that, The dwell constraint is: and ; in, The duration of the stay interval. As a minimum stay duration constraint, The spatial radius of the trajectory points within the dwell interval. This is a constraint on the maximum dwell radius.

7. The local differential privacy trajectory publishing algorithm for regions of interest as described in claim 6, characterized in that, Step three specifically includes: For trajectory points that fall into the safe zone, construct discrete candidate domains within the corresponding safe zone. Discrete candidate domain The candidate point with the smallest spatial distance from the current trajectory point is the nearest candidate point; Wherein, the discrete candidate domain Within its corresponding security zone and the first Candidate points are formed by the spatial distance between trajectory points not exceeding a preset threshold.

8. The local differential privacy trajectory publishing algorithm for regions of interest as described in claim 7, characterized in that, The generation of the disturbance point in step four includes the following steps: Step a: Divide the discrete candidate domain into A ring-shaped distance layer ; Step b: Obtain perturbation points by sampling in different distance layers according to the sampling rules.

9. The local differential privacy trajectory publishing algorithm for regions of interest as described in claim 8, characterized in that, In step a The radial spacing between the annular distance layers is equal.

10. The local differential privacy trajectory publishing algorithm for regions of interest as described in claim 9, characterized in that, The sampling rule is as follows: If the initial random number Then in the discrete candidate domain The disturbance point is obtained by uniform random sampling in the middle; otherwise, random numbers are generated again. If the random number is generated again If so, the nearest candidate point is selected as the perturbation point; If the random number is generated again and Then, any layer in the other annular distance layers can be selected and the perturbation points are obtained by uniformly sampling within the layer; in, As the probability of being at the bottom, Select the probability for the nearest layer.

Images