Internet of vehicles track privacy protection method based on diffusion model and location sensitivity hierarchical disturbance

Abstract

The invention discloses an Internet of Vehicles trajectory privacy protection method based on a diffusion model and location sensitivity hierarchical disturbance, which comprises the following steps: preprocessing vehicle trajectory data, including removing abnormal points, filling missing values, separating time and spatial features, normalizing or standardizing the features, synchronizing time intervals, and carrying out feature extraction; adding extra condition information of the starting point; forward diffusion is carried out on the preprocessed time feature sequence and the preprocessed space feature sequence, and original data is disturbed by adding noise; when backward diffusion is carried out on the disturbed trajectory, a joint denoising network is adopted to predict noise, a noiseless time sequence and a noiseless space sequence are gradually restored, and the restored noiseless time sequence and the restored noiseless space sequence are fused to obtain a synthetic trajectory; clustering the synthesized track points by using a DBSCAN algorithm, defining and grading sensitive sites, distributing privacy budget according to the proportion of the sensitive sites, and disturbing sensitive position points in the track; and reconstructing the trajectory according to the trajectory spatial position and speed rationality.

Description

Technical Field

[0001] The present invention relates to the field of information security technology, and in particular to a method for protecting the privacy of vehicle-to-vehicle (IoV) trajectories based on a diffusion model and location sensitivity graded disturbance. Background Art

[0002] With the rapid development and widespread adoption of technologies such as the internet, cloud computing, 5G, and mobile applications, the boundaries of the Internet of Things (IoT) have expanded from local sensor networks to ubiquitous objects, including smart electronic devices, vehicles, and public infrastructure, ushering in a new era for the IoT. The Internet of Vehicles (IoV) connects vehicles to the internet through communication technologies, enabling data exchange and information transmission to meet users' daily needs. With the advancement of technology, an increasing number of vehicles and mobile devices are connected to the internet, enabling network connections between vehicles, users, and devices. This connection is mobile, open, uncertain, and social, enabling vehicles to receive, store, and transmit information in real time, including trajectory, location, user, and vehicle information. This information is both private and sensitive. Without privacy protection, massive amounts of aggregated data are vulnerable to attacks and even leakage. Therefore, the privacy protection of vehicle users has attracted widespread attention from researchers. To protect privacy, measures must be implemented throughout the entire process of trajectory data collection, processing, reporting, and dissemination.

[0003] The collection and analysis of personal location and trajectory information has become a major source of privacy leaks. Once this information is leaked, it can lead to user privacy exposure and personal security issues. Criminals or hackers can exploit trajectory data that has been exposed to the internet for a long time, or through key locations within the trajectory, to steal the vehicle user's data, identity, and whereabouts, and even infer other related information. Therefore, in this context, research on trajectory privacy protection technology has become particularly urgent. It requires privacy processing of vehicle trajectory data through various means and methods to protect the privacy and security of vehicle users and their companions.

[0004] Traditional methods for protecting the privacy of Internet of Vehicles trajectories include trajectory anonymization, trajectory encryption, fake trajectory generation, deep learning, and other methods, but they all have problems such as high computational cost and low efficiency, insufficient data availability, limited personalized protection capabilities, and high risks in sensitive locations. Summary of the Invention

[0005] To address the shortcomings of the existing technology, the present invention provides a method for protecting the privacy of vehicle-to-vehicle (IoV) trajectories based on a diffusion model and location sensitivity graded perturbations. This method improves computational efficiency by separating temporal and spatial features, and protects sensitive locations by adding a personalized differential privacy mechanism based on the sensitivity levels of user hotspots.

[0006] The present invention discloses a method for protecting the privacy of vehicle-to-vehicle trajectory based on a diffusion model and location sensitivity graded perturbation, comprising:

[0007] Step 1: Preprocess the vehicle trajectory data to obtain the time feature sequence and spatial feature sequence of the vehicle trajectory;

[0008] Step 2: forward-diffuse the time feature sequence and the spatial feature sequence pre-processed in step 1 to disturb the time feature sequence and the spatial feature sequence by adding noise;

[0009] Step 3: Backward diffusion is performed on the time feature sequence and spatial feature sequence disturbed by the noise in step 2. A joint denoising network is used to predict the noise, and the noise-free time and space sequences are gradually restored. The restored noise-free time and space sequences are then fused to obtain a synthetic trajectory.

[0010] Step 4: Extract user hotspot areas from the synthesized trajectory;

[0011] Step 5: The user defines a set of sensitive locations and classifies them according to their sensitivity. After preprocessing, the number of sensitive locations in each user's hotspot area is counted.

[0012] Step 6: Determine the total privacy budget based on the distribution of sensitive locations across all clusters, and dynamically allocate the corresponding privacy budget based on the number of sensitive locations within each cluster. For each sensitive location within each cluster, use the allocated privacy budget to add noise.

[0013] Step 7: Reconstruct and verify the synthetic trajectory with noise disturbance.

[0014] As a further improvement of the present invention, step 1 includes:

[0015] Remove outliers from vehicle trajectory data, fill in missing values, complete trajectory segments, synchronize time intervals, and add additional conditional information about the trajectory starting point;

[0016] Separate the temporal and spatial features of the trajectory data, normalize or standardize the temporal and spatial features respectively, and obtain clear temporal and spatial feature sequences.

[0017] As a further improvement of the present invention, step 2 includes:

[0018] Define the total number of diffusion steps and design independent noise scheduling parameters for the time branch and the space branch according to the different characteristics of the time feature sequence and the space feature sequence;

[0019] The time feature sequence and the spatial feature sequence are forward diffused, and the cumulative retention rate of the time branch and the spatial branch is calculated, that is, the proportion of the signal in the original data that is retained after multiple steps of diffusion.

[0020] As a further improvement of the present invention, for a true trajectory sample T, the forward diffusion process is expressed as (T1, T2, T3...T N ), N represents the maximum number of diffusion steps. The process aims to gradually add noise to the original trajectory, thereby gradually destroying the spatiotemporal characteristics of the original trajectory, and finally making the trajectory follow the standard Gaussian distribution

[0021] As a further improvement of the present invention, step 3 includes:

[0022] The initial noise-perturbed temporal and spatial feature sequences are input, and the total number of inverse diffusion steps is defined. The noise components of the temporal and spatial branches are predicted separately using a joint denoising network, and the corresponding denoising means are calculated based on their respective noise scheduling parameters and cumulative retention rates. By iteratively sampling from the current noisy trajectory and continuously updating it to a trajectory with lower noise, the noise-free temporal and spatial feature sequences are gradually restored, thus achieving trajectory generation and recovery. Finally, the restored noise-free temporal and spatial sequences are fused to obtain a synthetic trajectory.

[0023] As a further improvement of the present invention, the joint denoising network includes a time branch network and a space branch network. The time branch network focuses on capturing the dynamic and correlation characteristics in the time series and predicting the noise components in the time characteristics; the space branch network focuses on extracting spatial trajectory morphological information and predicting the noise components in the spatial characteristics.

[0024] As a further improvement of the present invention, the initial layer of the joint denoising network uses a CNN neural network to quickly extract local patterns and detailed features of the trajectory, ensuring the authenticity of the local trajectory; the subsequent use of Transformer enhances the deep understanding of the global trajectory sequence and space, so that the generated trajectory globally conforms to the distribution and dynamic laws of the real trajectory.

[0025] As a further improvement of the present invention, step 4 includes: clustering the trajectory points of the synthetic trajectory using the DBSCAN algorithm, and during clustering, using the Euclidean distance to calculate the similarity between the trajectory points to determine the center of mass of each cluster; selecting the top K clusters with the largest number of trajectory points, and defining the selected clusters as user hotspot areas.

[0026] As a further improvement of the present invention, step 7 includes:

[0027] First, verify whether each trajectory point is located within the actual road range to ensure its spatial location is legal; second, evaluate the rationality of the moving speed based on the movement patterns between trajectory points, and eliminate trajectory points with abnormal speeds; smooth the trajectory points with abnormal positions; and delete the trajectory points with abnormal speeds, and insert new trajectory points between the previous and next trajectory points to maintain trajectory continuity.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] This paper combines a diffusion model with location-sensitivity graded perturbation technology to protect trajectory privacy. The diffusion model is used to generate synthetic trajectories with global characteristics, preserving the overall characteristics of the trajectory while effectively ensuring the privacy security of the entire trajectory. At the same time, location-sensitivity graded perturbation provides fine-grained protection for sensitive local locations within the trajectory, achieving an organic combination of privacy protection and data practicality.

[0030] In the forward diffusion process of the present invention, the time and space branches are diffused separately, and different noise parameters are designed, which can more effectively maintain the true characteristics of time and space information and improve the quality of denoising and trajectory generation.

[0031] In the backward diffusion process of the present invention, CNN+Transformer is used as the denoising network, which realizes the organic integration of local details and global patterns of trajectory data, and improves the denoising network's ability to model trajectory authenticity and dynamic consistency, thereby effectively protecting user trajectory privacy in the Internet of Vehicles environment while ensuring data availability and rationality.

[0032] The personalized differential privacy feature dynamically adjusts the privacy budget based on the sensitivity of the location, assigning different protection levels to different locations. For highly sensitive and critical locations, stricter privacy protection measures are applied, increasing noise perturbation; for non-sensitive or general locations, the perturbation level is appropriately reduced. This approach maximizes data validity and usability while ensuring privacy, achieving precise and flexible privacy protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flow chart of the vehicle network trajectory privacy protection method based on the diffusion model and location sensitivity graded perturbation disclosed in the present invention;

[0034] Figure 2 This is the forward diffusion process in step 2 of the present invention;

[0035] Figure 3 This is the backward diffusion process in step 3 of the present invention;

[0036] Figure 4This is the location sensitivity graded disturbance process in steps 4-6 of the present invention. DETAILED DESCRIPTION

[0037] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0038] The present invention will be described in further detail below with reference to the accompanying drawings:

[0039] The present invention provides a privacy protection method for vehicle-to-vehicle (IoV) trajectories based on a diffusion model and hierarchical perturbation based on location sensitivity. The method first preprocesses vehicle trajectory data, including removing outliers, filling missing values, separating temporal and spatial features, normalizing or standardizing features, synchronizing time intervals, and adding additional conditional information about the starting point. The preprocessed temporal and spatial feature sequences are then forward diffused, and the original data is perturbed by adding noise. When the perturbed trajectory is backward diffused, a joint denoising network is used to predict noise, and the noise-free temporal and spatial sequences are gradually restored. The restored noise-free temporal and spatial sequences are then fused to obtain a synthetic trajectory. The synthesized trajectory points are clustered using the DBSCAN algorithm, sensitive locations are defined and classified, a privacy budget is allocated based on the proportion of sensitive locations, and sensitive locations in the trajectory are perturbed. Finally, the trajectory is reconstructed based on the rationality of the trajectory's spatial position and speed.

[0040] Example:

[0041] like Figures 1 to 4 The specific steps are as follows:

[0042] Step 1: Preprocess the vehicle trajectory data to obtain the time feature sequence and spatial feature sequence;

[0043] Specifically include:

[0044] The T-Drive dataset, containing GPS trajectories of 10,357 taxis in Beijing between February 2 and February 8, 2008, was used. Outliers were removed from the vehicle trajectory data, missing values ​​were filled, and trajectory segments were completed. Time intervals were synchronized, and additional conditional information about the trajectory starting point was added. The temporal and spatial features of the trajectory data were separated and normalized to obtain clear temporal and spatial feature sequences.

[0045] Step 2: forward diffuse the time feature sequence and spatial feature sequence preprocessed in step 1 to disturb the original data (time feature sequence and spatial feature sequence) by adding noise;

[0046] Specifically include:

[0047] Define the total number of diffusion steps N, and design independent noise scheduling parameters for the time branch and the space branch according to their different characteristics. Then, perform forward diffusion on the time and space feature sequences, and calculate the cumulative retention rate of the time branch and the space branch, that is, the proportion of the signal in the original data that is retained after N steps of diffusion. Specifically, for a real trajectory sample T, the forward diffusion process is expressed as (T1, T2, T3...T N ), N represents the maximum number of diffusion steps. The process aims to gradually add noise to the original trajectory, thereby gradually destroying the spatiotemporal characteristics of the original trajectory, and finally making the trajectory follow the standard Gaussian distribution

[0048] Step 3: Backward diffusion is performed on the time feature sequence and spatial feature sequence disturbed by the noise in step 2. A joint denoising network is used to predict the noise, and the noise-free time and space sequences are gradually restored. The restored noise-free time and space sequences are then fused to obtain a synthetic trajectory.

[0049] Specifically include:

[0050] The initial noisy trajectories of the time and space branches are input, the total number of inverse diffusion steps is defined, and the noise components of the time and space branches are predicted separately using a joint denoising network. The corresponding denoising means are calculated based on their respective noise scheduling parameters and cumulative retention rates. By iteratively sampling from the current noisy trajectory and continuously updating it to a trajectory with lower noise, the noise-free time and space feature sequences are gradually restored to achieve trajectory generation and recovery. Finally, the restored noise-free time and space sequences are fused to obtain a synthetic trajectory.

[0051] The joint denoising network gradually removes noise from the time branch and the space branch respectively, restoring clean data close to the true trajectory. The joint denoising network includes a time branch network and a space branch network. The time branch network focuses on capturing the dynamic and correlation characteristics in the time series and predicting the noise component in the time characteristics; the space branch network focuses on extracting the spatial trajectory morphological information and predicting the noise component in the spatial characteristics.

[0052] The initial layer of the joint denoising network uses a CNN neural network to quickly extract local patterns and detailed features of trajectories, ensuring the authenticity of local trajectories. The subsequent use of the Transformer improves the deep understanding of the global trajectory sequence and spatial information, ensuring that the generated trajectories globally conform to the distribution and dynamic laws of real trajectories.

[0053] Step 4: Extract user hotspot areas from the synthesized trajectory;

[0054] Specifically include:

[0055] The DBSCAN algorithm is used to cluster the trajectory points of the synthetic trajectory based on the time trajectory data. During clustering, the Euclidean distance is used to calculate the similarity between the trajectory points to determine the centroid of each cluster. The top K clusters with the largest number of trajectory points are selected and defined as user hotspots, that is, areas where users frequently stay.

[0056] Step 5: The user defines a set of sensitive locations and classifies them into high-sensitivity and low-sensitivity categories based on their sensitivity. After preprocessing, the synthesized trajectories are clustered using the DBSCAN algorithm to obtain user hotspots. The number of sensitive locations contained in each user hotspot is counted to quantify the coverage of sensitive information by different clusters.

[0057] Step 6: Determine the total privacy budget based on the distribution of sensitive locations in all clusters, and dynamically allocate the corresponding privacy budget based on the number of sensitive locations in each cluster. For each sensitive location in each cluster, use the allocated privacy budget to add noise, thereby achieving effective disturbance protection for sensitive location information.

[0058] Step 7: Reconstruct and verify the synthetic trajectory with noise disturbance;

[0059] Specifically include:

[0060] First, verify whether each trajectory point is located within the actual road range to ensure its spatial location is legal; second, evaluate the rationality of the moving speed based on the movement patterns between trajectory points, and eliminate trajectory points with abnormal speeds; smooth the trajectory points with abnormal positions; and delete the trajectory points with abnormal speeds, and insert new trajectory points between the previous and next trajectory points to maintain trajectory continuity.

[0061] The accuracy of trajectory-user binding is used to measure the difficulty for attackers to identify users through their trajectories. This method outperforms random perturbations and directly adding differential privacy in terms of accuracy, fully demonstrating its excellent privacy protection capabilities.

[0062] Through temporal feature similarity analysis, we calculated the visit probability and frequency of different location categories during different time periods. The results show that the access patterns of the generated trajectories for mainstream categories and key time periods are highly consistent with the real trajectories, effectively preserving the time series characteristics and ensuring the generated data has high practical value.

[0063] Spatial feature similarity analysis measures the spatial differences between generated and original trajectory points based on the Euclidean distance, the Hausdorff distance. Although the average Hausdorff distance of the generated trajectories is slightly higher than that of the simple perturbation method, this difference is still within an acceptable range given its excellent privacy protection.

[0064] In addition, as training continues, the privacy protection effect of the model continues to improve, while the temporal and spatial similarity indicators remain basically stable, indicating that while the model strengthens privacy protection, it does not cause significant loss to the temporal and spatial characteristics of the data.

[0065] like Figure 2 As shown in Figure 2, the temporal and spatial features of a trajectory have different properties and statistical characteristics. Temporal features reflect the dynamic changes of the trajectory over time, while spatial features focus on the spatial distribution and structure of trajectory points. By independently designing the noise scheduling parameters and cumulative retention rates for the temporal and spatial branches, the model can more accurately capture and express the noise patterns and dynamic characteristics of each branch, avoiding the potential loss of information or degradation of model performance caused by mixing different features. This approach more effectively preserves the true characteristics of temporal and spatial information, improving the quality of denoising and trajectory generation.

[0066] Figure 3 This is a schematic diagram of the backward diffusion of noisy trajectories in step 1 of an embodiment of the present invention; trajectory data is usually a spatiotemporal sequence, which contains both local details (such as instantaneous speed and direction changes) and overall dynamic laws and spatial dependencies. Therefore, the present invention uses CNN+Transformer as a denoising network. CNN uses rapid extraction of local spatial and temporal features of the trajectory in the initial layer, which can accurately capture details such as speed changes and turns, ensure the authenticity of the trajectory structure, and avoid loss of details due to denoising distortion. Transformer captures long-term temporal dependencies and spatial relationships in the trajectory sequence through the self-attention mechanism, comprehensively understands the overall motion trend and dynamic laws of the trajectory, and makes the generated results conform to the distribution characteristics of the real trajectory at a macro level. The combined use can effectively learn the complex spatiotemporal dynamics of the trajectory, avoid unreasonable jumps or anomalies in the generated trajectory through global and local synchronous modeling, and ensure the dynamic coherence of the trajectory data.

[0067] The convolutional layer of the denoising network uses multiple layers of one-dimensional convolution operations with a convolution kernel of 4 to capture the spatial and temporal features of local continuous points in the trajectory. The ReLU activation function is used after each convolution layer to enhance the nonlinear expression capability of the model. Some layers combine average pooling for dimensionality reduction and feature extraction. At the same time, a batch normalization layer is introduced to speed up training and improve stability.

[0068] The Transformer layer of the denoising network uses a self-attention layer to focus on the relationship between different positions in the trajectory sequence, uses a feedforward neural network for more complex linear changes, and uses layer normalization and residual connections to ensure smooth information flow and gradient stability.

[0069] Figure 4 This figure is a schematic diagram of the location sensitivity graded perturbation of the synthetic trajectory in steps 4-6 of an embodiment of the present invention. Although the synthetic trajectory generated based on the diffusion model can reduce direct leakage, potential privacy information remains. In particular, hot spots and sensitive areas in the synthetic trajectory may be identified and associated.

[0070] First, the user defines a set of sensitive locations and classifies them according to their sensitivity (high sensitivity and low sensitivity). After preprocessing, the synthesized trajectories are clustered using the DBSCAN algorithm to obtain user hotspots. The number of sensitive locations contained in each user hotspot is counted to quantify the coverage of sensitive information by different clusters.

[0071] Based on the distribution of sensitive locations across all clusters, the total privacy budget is determined and dynamically allocated based on the number of sensitive locations within each cluster. For each sensitive location within each cluster, noise is added using the allocated privacy budget, effectively protecting sensitive location information from disturbances.

[0072] The overall privacy budget is defined as:

[0073] ε total

[0074] User hotspot area determination is based on the ratio of trajectory points:

[0075] like Then C i Defined as user hotspot area.

[0076] Among them, |C i | is the number of trajectory points in cluster i, N is the total number of all trajectory points, and τ is the set threshold.

[0077] Definition of sensitivity of user hotspot area:

[0078] S(C i )=S1(C i )+S2(C i )

[0079] Among them, S1(C i ) and S2(C i ) represent clusters C i Sensitivity on a high sensitivity level and a low sensitivity level.

[0080] The privacy budget allocation formula for each cluster is:

[0081]

[0082] in,

[0083] ε i is the privacy budget allocated to the cluster.

[0084] ε total is the overall privacy budget.

[0085] S(C i ) is the sensitivity of cluster i.

[0086] ∑ j S(C j ) is the sum of all cluster sensitivities.

[0087] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for protecting the privacy of vehicle-to-vehicle trajectory based on a diffusion model and location-sensitivity graded perturbation, characterized by: include: Step 1: Preprocess the vehicle trajectory data to obtain the time feature sequence and spatial feature sequence of the vehicle trajectory; Step 2: forward-diffuse the time feature sequence and the spatial feature sequence pre-processed in step 1 to disturb the time feature sequence and the spatial feature sequence by adding noise; Step 3: Backward diffusion is performed on the time feature sequence and spatial feature sequence disturbed by the noise in step 2. A joint denoising network is used to predict the noise, and the noise-free time and space sequences are gradually restored. The restored noise-free time and space sequences are then fused to obtain a synthetic trajectory. Step 4: Extract user hotspot areas from the synthesized trajectory; Step 5: The user defines a set of sensitive locations and classifies them according to their sensitivity. After preprocessing, the number of sensitive locations in each user's hotspot area is counted. Step 6: Determine the total privacy budget based on the distribution of sensitive locations across all clusters, and dynamically allocate the corresponding privacy budget based on the number of sensitive locations within each cluster. For each sensitive location within each cluster, use the allocated privacy budget to add noise. Step 7: Reconstruct and verify the synthetic trajectory with noise disturbance.

2. The method for protecting the privacy of vehicle-to-vehicle trajectory based on the diffusion model and location-sensitivity graded perturbation according to claim 1, characterized in that: The step 1 comprises: Remove outliers from vehicle trajectory data, fill in missing values, complete trajectory segments, synchronize time intervals, and add additional conditional information about the trajectory starting point; Separate the temporal and spatial features of the trajectory data, normalize or standardize the temporal and spatial features respectively, and obtain clear temporal and spatial feature sequences.

3. The method for protecting the privacy of vehicle-to-vehicle trajectory based on the diffusion model and location-sensitivity graded perturbation according to claim 1, characterized in that: The step 2 includes: Define the total number of diffusion steps and design independent noise scheduling parameters for the time branch and the space branch according to the different characteristics of the time feature sequence and the space feature sequence; The time feature sequence and the spatial feature sequence are forward diffused, and the cumulative retention rate of the time branch and the spatial branch is calculated, that is, the proportion of the signal in the original data that is retained after multiple steps of diffusion.

4. The method for protecting the privacy of vehicle-to-vehicle trajectory based on a diffusion model and location-sensitivity graded perturbations according to claim 3, characterized in that: For the true trajectory sample T, the forward diffusion process is expressed as (T1, T2, T3...T N ), N represents the maximum number of diffusion steps. The process aims to gradually add noise to the original trajectory, thereby gradually destroying the spatiotemporal characteristics of the original trajectory, and finally making the trajectory follow the standard Gaussian distribution 5. The method for protecting the privacy of vehicle-to-vehicle trajectory based on a diffusion model and location-sensitivity graded perturbations according to claim 1, characterized in that: The step 3 comprises: The initial noise-perturbed temporal and spatial feature sequences are input, and the total number of inverse diffusion steps is defined. The noise components of the temporal and spatial branches are predicted separately using a joint denoising network, and the corresponding denoising means are calculated based on their respective noise scheduling parameters and cumulative retention rates. By iteratively sampling from the current noisy trajectory and continuously updating it to a trajectory with lower noise, the noise-free temporal and spatial feature sequences are gradually restored, thus achieving trajectory generation and recovery. Finally, the restored noise-free temporal and spatial sequences are fused to obtain a synthetic trajectory.

6. The method for protecting the privacy of vehicle-to-vehicle trajectory based on a diffusion model and location-sensitivity graded perturbations according to claim 5, characterized in that: The joint denoising network includes a time branch network and a space branch network. The time branch network focuses on capturing the dynamic and correlation characteristics in the time series and predicting the noise components in the time characteristics; the space branch network focuses on extracting spatial trajectory morphological information and predicting the noise components in the spatial characteristics.

7. The method for protecting the privacy of vehicle-to-vehicle trajectory based on a diffusion model and location-sensitivity graded perturbations according to claim 5, characterized in that: The initial layer of the joint denoising network uses a CNN neural network, and the subsequent layer uses a Transformer.

8. The method for protecting the privacy of vehicle-to-vehicle trajectory based on a diffusion model and location-sensitivity graded perturbation according to claim 1, characterized in that: Step 4 includes: clustering the trajectory points of the synthetic trajectory using the DBSCAN algorithm, calculating the similarity between the trajectory points using the Euclidean distance during clustering to determine the centroid of each cluster; selecting the top K clusters with the largest number of trajectory points, and defining the selected clusters as user hotspot areas.

9. The method for protecting the privacy of vehicle-to-vehicle trajectory based on a diffusion model and location-sensitivity graded perturbation according to claim 1, characterized in that: The step 7 comprises: First, verify whether each trajectory point is located within the actual road range to ensure its spatial location is legal; second, evaluate the rationality of the moving speed based on the movement patterns between trajectory points, and eliminate trajectory points with abnormal speeds; smooth the trajectory points with abnormal positions; and delete the trajectory points with abnormal speeds, and insert new trajectory points between the previous and next trajectory points to maintain trajectory continuity.

Images