A data completion method based on spatiotemporal privacy protection

Abstract

This invention discloses a data completion method based on spatiotemporal privacy protection, which relates to the field of data completion technology. The method comprises the following steps: obtaining spatiotemporal data to be completed; obfuscating the spatiotemporal data to be completed; mapping the obfuscated spatiotemporal data onto a three-dimensional Hilbert curve, assigning a unique Hilbert value to each position, and constructing a location privacy set; generating a confusion probability matrix by permuting and flipping the location privacy set; and inputting the confusion probability matrix, the obfuscated spatiotemporal data, and the spatiotemporal data to be completed into a completion model to obtain the completed spatiotemporal data. The completion model includes a diffusion process and an inverse process. The present invention improves completion accuracy in privacy protection scenarios.

Description

Technical Field

[0001] The present invention belongs to the technical field of data completion, and in particular relates to a data completion method based on spatiotemporal privacy protection. Background Art

[0002] The current mainstream privacy protection methods are mainly divided into two categories: anonymization-based methods and obfuscation-based methods. Among them, k-anonymity technology is a typical anonymization method. By generating k-1 virtual location points and the real location to form an anonymous set, it makes it difficult for attackers to identify the real location with a probability higher than 1 / k. In terms of obfuscation methods, differential privacy has become a research hotspot due to its rigorous mathematical proof and strong privacy protection capabilities. This model can provide quantifiable privacy protection even when the attacker has background knowledge. In particular, geographic indistinguishability, as an extended application of differential privacy in the field of location services, has become a widely recognized privacy protection standard in this field by ensuring that the geographic location produces statistically indistinguishable output within a specific radius. Existing research on data inference in uncovered areas is mainly based on the following two methods: (1) Traditional statistical methods: matrix decomposition technology realizes data reconstruction through low-rank approximation, and compressed sensing technology uses spatial correlation to restore the complete data distribution from sparse sampling points in scenarios such as urban noise monitoring. Interpolation methods based on spatiotemporal correlation are widely used in fields such as traffic monitoring. (2) Deep learning methods: Deep matrix factorization enhances the feature extraction capability of traditional MF through neural networks. Diffusion models show great potential in spatiotemporal data modeling. Specialized models such as CSDI and PriSTI are optimized for time series data and spatiotemporal data completion, respectively.

[0003] Existing location privacy protection techniques (k-anonymity, differential privacy, and geographic indistinguishability) primarily focus on protecting privacy in the spatial dimension, namely, hiding or obfuscating a user's location information. However, these methods often overlook privacy requirements in the temporal dimension, allowing attackers to infer a user's movement patterns, behavioral habits, and even identity information from timestamps. For example, even if spatial location is obfuscated, consecutive timestamps may still reveal the user's trajectory patterns, thereby reducing the effectiveness of privacy protection. Furthermore, existing privacy protection methods are often independent of downstream data analysis tasks. As a result, data that has undergone privacy protection (such as noise addition or location generalization) may introduce significant errors when subsequently used (such as in data completion, prediction, or visualization), affecting data usability.

[0004] Existing data completion methods (such as matrix factorization, compressed sensing, and deep learning models) primarily target raw spatiotemporal data, assuming it is unobfuscated or unperturbed. However, in privacy-preserving scenarios, data may be subjected to noise injection, generalization, or differential privacy processing, resulting in performance degradation for traditional completion methods. For example, traditional methods rely on assumptions of low rank or sparsity, but noise can destroy the data's structural characteristics. Deep learning methods (such as deep matrix factorization and diffusion models) typically require high-quality training data, and privacy-preserving data can affect the model's generalization capabilities. Existing completion models fail to consider the characteristics of privacy-preserving data, resulting in limited completion accuracy.

[0005] Combining privacy protection and data quality maintenance faces the following challenges:

[0006] (1) Collaborative protection of spatiotemporal privacy: Existing methods mainly focus on spatial privacy, while the protection of the temporal dimension (such as timestamp generalization and temporal differential privacy) has not been fully studied.

[0007] (2) Availability of privacy-preserving data: Noise or generalization may reduce data quality. How to minimize data distortion while protecting privacy is a key issue.

[0008] Completion methods for obfuscated data: Existing completion models are not optimized for handling privacy-protected data. Robust completion algorithms need to be designed to adapt to noisy or generalized data distributions. Therefore, a data completion method based on spatiotemporal privacy protection is urgently needed. Summary of the Invention

[0009] To solve the above technical problems, the present invention proposes a data completion method based on spatiotemporal privacy protection, designs an end-to-end privacy protection and data completion framework, applies privacy protection mechanisms in both spatial and temporal dimensions, reduces the impact of privacy noise on data availability through data adjustment functions, optimizes the completion algorithm for obfuscated spatiotemporal data, and improves the completion accuracy in privacy protection scenarios.

[0010] To achieve the above objectives, the present invention provides a data completion method based on spatiotemporal privacy protection, comprising: obtaining spatiotemporal data to be completed, obfuscating the spatiotemporal data to be completed, mapping the obfuscated spatiotemporal data onto a three-dimensional Hilbert curve, assigning a unique Hilbert value to each position, and constructing a location privacy set; generating a confusion probability matrix by permuting and flipping the location privacy set;

[0011] The confusion probability matrix, the confused spatiotemporal data, and the spatiotemporal data to be completed are input into a completion model to obtain the completed spatiotemporal data, wherein the completion model includes a diffusion process and an inverse process.

[0012] Optionally, constructing the location privacy set includes:

[0013] The spatiotemporal position is mapped to a one-dimensional value through a three-dimensional Hilbert curve, the spatiotemporal distance sensitivity of the position in the search neighborhood is calculated, the expected inference error is constrained, and a location privacy set is constructed.

[0014] Optionally, the expected inference error of the location privacy set includes:

[0015] ;

[0016] Where, is the expected inference error of the location privacy set; is the real space-time position; is the inferred space-time location; Protect collections for locations; for The index of the position; For prior knowledge; is the space-time distance.

[0017] Optionally, during the diffusion process of the completion model, Gaussian noise is added to the spatiotemporal data to be completed, which is expressed as follows:

[0018] ;

[0019] ;

[0020] in, To generate the joint probability distribution of the noise sequence given the original data; is the noise data sequence from step 1 to step T; is the original data; is the total step length; For a certain step; is the noise data of step t; for arrive The single-step diffusion conditional probability distribution of ; is a Gaussian distribution; is the noise variance at step t; is the mean of the Gaussian distribution, which represents the scaling of the previous step data; is the variance.

[0021] Optionally, the inverse process of the completion model is to gradually convert random noise into spatiotemporally consistent missing values ​​based on conditional information. The inverse process is expressed as follows:

[0022] ;

[0023] ;

[0024] in, is the denoised data sequence from step 0 to step T−1; is the pure noise data of the Tth step; is conditional information; is the geographic adjacency matrix; To generate the joint probability distribution of denoised sequence given the noise data and conditional information; is the total step length; For a certain step; is the Markov property of the reverse process, indicating that the joint distribution can be decomposed into the product of the conditional probabilities at each step; To predict the denoised data of the previous step; is a Gaussian distribution; is the mean value predicted by the neural network; is the variance matrix.

[0025] Optionally, the completion model further includes a conditional feature extraction module and a noise estimation module;

[0026] The conditional feature extraction module is used to extract features from the confusion probability matrix, the obfuscated spatiotemporal data, and the spatiotemporal data to be completed to obtain extracted conditional features;

[0027] The noise estimation module is used to perform noise prediction on the real data with noise, the obfuscated spatiotemporal data and the extracted conditional features to obtain a noise prediction result.

[0028] Optionally, the conditional feature extraction module includes: a spatial attention unit, a temporal attention unit and a graph convolutional network unit;

[0029] The spatial attention unit is used to extract global spatial correlation from the confusion probability matrix, the obfuscated spatiotemporal data, and the spatiotemporal data to be completed;

[0030] The temporal attention unit is configured to adjust and model a time series dynamic pattern of the confusion probability matrix, the confused spatiotemporal data, and the spatiotemporal data to be completed;

[0031] The graph convolutional network unit is used to fuse geographic adjacency information.

[0032] Optionally, the extracted conditional features are:

[0033] ;

[0034] ;

[0035] ;

[0036] ;

[0037] ;

[0038] in, is the global spatiotemporal feature; The feature representation of the interpolated observation data after 1×1 convolution; is the geographic adjacency matrix; () is a multi-layer perceptron, used to fuse the three features; for spatial attention; For time attention; It is a graph neural network; () is the normalization layer; Multi-head self-attention mechanism for spatial dimension; It is a multi-head self-attention mechanism in the time dimension; for message passing neural networks; is the space-time data after confusion; is the confusion probability matrix; It is a convolutional network; is the layer; Att() is the global attention mechanism; spa is the spatial attention; tem is the temporal attention; Conv is the convolutional layer.

[0039] Technical effect of the present invention: The present invention discloses a data completion method based on spatiotemporal privacy protection, which introduces the confusion probability matrix into the completion model. This design minimizes the impact of the privacy protection mechanism on the completion accuracy. The embedded design of the confusion probability matrix P improves the adaptability of the model to the privacy protection mechanism. Combined with spatiotemporal attention, graph convolution and diffusion models, it takes into account both global dependencies and local features. The gradient vanishing is alleviated by residual and jump connections to ensure the effective modeling of complex spatiotemporal patterns. Temporal information is also protected, and the reasoning error is limited, thereby enhancing robustness. The obfuscation method of the present invention can resist long-term reasoning attacks. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0041] Figure 1 This is a schematic diagram of the overall framework of a data completion method based on spatiotemporal privacy protection according to an embodiment of the present invention;

[0042] Figure 2 This is a complete model diagram for an embodiment of the present invention. DETAILED DESCRIPTION

[0043] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0044] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0045] like Figure 1 As shown, this embodiment provides a data completion method based on spatiotemporal privacy protection, including:

[0046] Obtain the spatiotemporal data to be completed, obfuscate the spatiotemporal data to be completed, map the obfuscated spatiotemporal data onto a three-dimensional Hilbert curve, assign a unique Hilbert value to each position, and construct a location privacy set; generate a confusion probability matrix by permuting and flipping the location privacy set;

[0047] The confusion probability matrix, the confused spatiotemporal data, and the spatiotemporal data to be completed are input into a completion model to obtain the completed spatiotemporal data, wherein the completion model includes a diffusion process and an inverse process.

[0048] Specifically, the overall framework proposed in this embodiment includes two key components: a spatiotemporal privacy protection mechanism and a data completion model. The spatiotemporal privacy protection mechanism first maps all possible spatiotemporal locations onto a three-dimensional Hilbert curve, assigning each location a unique Hilbert value. Along the Hilbert curve, each spatiotemporal location is given its own identification location privacy set to effectively hide its true spatiotemporal location. Subsequently, a confusion probability matrix is ​​generated through a permutation and flipping mechanism to produce confused locations that satisfy differential privacy, and the mechanism is able to resist Bayesian attacks and optimal inference attacks. This spatiotemporal privacy protection mechanism ensures the simultaneous protection of temporal and spatial attributes. On this basis, a diffusion model is used to complete missing data. The model is guided by the confusion probability matrix, the obfuscated spatiotemporal data, and geographic information. This approach minimizes the adverse effects of privacy protection on the accuracy of the data completion process.

[0049] Specifically, the spatiotemporal privacy protection mechanism includes:

[0050] In order to protect the temporal and spatial attributes of spatiotemporal data, we first connect the possible spatiotemporal locations using a three-dimensional Hilbert curve. The three-dimensional Hilbert curve maps the three-dimensional spatiotemporal location x to a one-dimensional value, called the Hilbert value of location x. By sorting the Hilbert values ​​of all possible access locations, we can obtain the rank of each location x in the set X. Specifically, let is the sequence of locations within the search neighborhood of x along the Hilbert curve, sorted by their Hilbert values. After determining the possible output set, another question is how to effectively determine the PLS, because the PLS cannot be infinitely large or infinitely small, because the quality of service must be guaranteed, that is, to reduce the damage to data quality, while also ensuring the expected inference error. The expected inference error of the location privacy set is calculated using the following formula:

[0051] ;

[0052] in, is the expected inference error of the location privacy set; is the real space-time position; is the inferred space-time location; Protect collections for locations; for The index of the position; For prior knowledge; is the space-time distance.

[0053] Introducing a privacy parameter , in order to ensure a lower bound on the expected inference error, the PLS search must satisfy .

[0054] Given a location privacy set, differential privacy is achieved by publishing obfuscated spatiotemporal locations using a permutation and flipping mechanism. This mechanism is initially designed to protect privacy during data publishing and generates obfuscated locations from the location privacy set determined in the first phase. The utility of an obfuscated location is given by the distance between the obfuscated location and the location x in the location privacy set. The smaller the distance, the higher the utility. The location privacy set identifies the neighboring locations of the real location, and the sensitivity of the utility function for:

[0055] ;

[0056] in, is the real space-time position; is the space-time position after confusion; is the position in the position set; Protect collections for locations; is the space-time distance; is the space-time distance.

[0057] Given the location In three-dimensional space, the triangle inequality can be used to derive the following conclusions: , , is the maximum spatiotemporal distance between two locations in PLS. For a given location and location protection set, the probability of outputting a confused location is proportional to the following, where is the normalization factor of the probability distribution :

[0058] ;

[0059] in, is the normalization factor; () is the power of e; budget for privacy; is the space-time distance; is the maximum space-time distance; is the real location; is the position after confusion; The maximum spatiotemporal distance value for the location protection set.

[0060] Specifically, such as Figure 2 The spatiotemporal data shown are complete with:

[0061] In order to apply the diffusion model to space-time completion, the space-time completion problem is regarded as a conditional generation task. The existing technology has demonstrated the ability of conditional diffusion probability models for multivariate time series completion. The space-time completion task is regarded as computing the conditional probability distribution ,in The complement of the observed value However, the existing technology does not consider the spatial relationship for completion, but simply uses the observation value as conditional information. To represent the diffusion steps. The diffusion probability model includes the diffusion process and the inverse process. The diffusion process of spatiotemporal completion is independent of the conditional information. Gaussian noise is added to the original data of the completion part, which is formalized as:

[0062] ;

[0063] ;

[0064] in, To generate the joint probability distribution of the noise sequence given the original data; is the noise data sequence from step 1 to step T; is the original data; is the total step length; For a certain step; is the noise data of step t; for arrive The single-step diffusion conditional probability distribution of is Gaussian distribution; is the noise variance at step t (controlling noise intensity), which is a pre-set hyperparameter; is the mean of the Gaussian distribution, which represents the scaling of the previous step data; is the variance.

[0065] A PriSTI-based spatiotemporal data completion framework was constructed to mitigate the data quality degradation caused by obfuscation. The framework uses a conditional diffusion model, utilizing the obfuscated spatiotemporal data and the confusion probability matrix as conditional constraints to optimize the completion process. Furthermore, geographic information is incorporated to further guide the model in accurately filling in the obfuscated spatiotemporal data.

[0066] in is a small constant hyperparameter that controls the variance of the added noise. The sampling formula is ,in , is the sampled standard Gaussian noise. When T is large enough, Close to the standard normal distribution. The inverse process of spatiotemporal completion is to gradually convert random noise into spatiotemporally consistent missing values ​​based on conditional information. The inverse process is conditioned on the interpolation condition information X of the enhanced observations and the geographic information A. The reverse process can be formalized as:

[0067] ;

[0068] ;

[0069] in, is the denoised data sequence from step 0 to step T−1; is the pure noise data of the Tth step; is conditional information; is the geographic adjacency matrix; To generate the joint probability distribution of denoised sequence given the noise data and conditional information; is the total step length; For a certain step; is the Markov property of the reverse process, indicating that the joint distribution can be decomposed into the product of the conditional probabilities at each step; To predict the denoised data of the previous step; is a Gaussian distribution; is the mean value predicted by the neural network; is the variance matrix.

[0070] in, and A valid parameterization of is defined as:

[0071] ;

[0072] .

[0073] Noise prediction model design:

[0074] The conditional feature module inputs the confusion probability matrix P, the interpolated obfuscated data X', and the geographic adjacency matrix A into the completion model. Graph convolution is used to enhance neighborhood node interactions and capture deep spatiotemporal dependencies. Specifically, it includes: spatial attention: extracting global spatial correlations; temporal attention: modeling dynamic patterns in time series; graph convolutional network: integrating geographic adjacency information. The final output is the conditional feature, which is used in the diffusion denoising process to ensure global context consistency. The conditional features are specifically:

[0075] ;

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] in, is the global spatiotemporal feature; The feature representation of the interpolated observation data after 1×1 convolution; is the geographic adjacency matrix; () is a multi-layer perceptron, used to fuse the three features; for spatial attention; For time attention; It is a graph neural network; () is the normalization layer; Multi-head self-attention mechanism for spatial dimension; It is a multi-head self-attention mechanism in the time dimension; for message passing neural networks; is the space-time data after confusion; is the confusion probability matrix; It is a convolutional network; is the layer; Att() is the global attention mechanism; spa is the spatial attention; tem is the temporal attention; Conv is the convolutional layer.

[0081] The noise estimation module takes in noisy real data, interpolated obfuscated data Y', conditional features, and the geographic adjacency matrix A. Through a multi-layer cascade structure, each layer generates spatial features, and information flow is optimized through residual and skip connections. An embedded representation of the confusion probability matrix P is introduced to enhance the modeling capabilities of conditional constraints. Finally, the noisy prediction result is output, retaining only the predicted value for the target portion.

[0082] Training objective (loss function) :

[0083] ;

[0084] in, is the expected value; is the real Gaussian noise; is the predicted noise; is the noise data after diffusion in step t; is conditional information; is the geographic adjacency matrix; is the confusion probability matrix; is the diffusion step.

[0085] A specific application example of the present invention is as follows:

[0086] In intelligent transportation systems, traffic management departments need to collect and analyze vehicle spatiotemporal trajectory data (such as GPS location, timestamp, and speed) to optimize traffic flow, predict congestion, and plan routes. However, this data contains sensitive information, and directly publishing or using it could leak user privacy (such as home addresses and frequented locations). Furthermore, due to signal loss or equipment failure, raw data may contain missing values, requiring data interpolation. The method of the present invention can achieve the dual goals of privacy protection and data interpolation in this scenario.

[0087] This paper proposes a unified framework that integrates spatiotemporal privacy protection and data completion. This framework addresses the dual challenge of protecting privacy while maintaining the practicality of spatiotemporal data in downstream tasks. This paper proposes a spatiotemporal obfuscation mechanism based on differential privacy. This mechanism aims to effectively protect the spatial and temporal attributes of data while ensuring privacy. A rigorous theoretical analysis is provided to demonstrate the differential privacy guarantees of the proposed spatiotemporal privacy protection mechanism and its robustness against inference attacks (including expected error bounds and posterior probability limits). This paper develops a diffusion model for obfuscated spatiotemporal data completion, which is conditioned on a confusion probability matrix, obfuscated data, and geographic information. This design minimizes the impact of the privacy protection mechanism on completion accuracy.

[0088] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A data completion method based on spatiotemporal privacy protection, characterized in that: include: Obtain the spatiotemporal data to be completed, obfuscate the spatiotemporal data, map the obfuscated spatiotemporal data onto a three-dimensional Hilbert curve, assign a unique Hilbert value to each position, and construct a location privacy set; generate a confusion probability matrix by permuting and flipping the location privacy set; the spatiotemporal data to be completed is the spatiotemporal trajectory data of the vehicle, including GPS position, timestamp, and speed; Inputting the confusion probability matrix, the obfuscated spatiotemporal data, and the spatiotemporal data to be completed into a completion model to obtain completed spatiotemporal data, wherein the completion model includes a diffusion process and an inverse process; Constructing the location privacy set includes: Mapping spatiotemporal locations to one-dimensional values ​​through a three-dimensional Hilbert curve, calculating the spatiotemporal distance sensitivity of locations within the search neighborhood, constraining the expected inference error, and constructing a location privacy set; The expected inference error of the location privacy set includes: ； in, is the expected inference error of the location privacy set; is the real space-time position; is the inferred space-time location; Protect collections for locations; for The index of the position; For prior knowledge; is the space-time distance.

2. The data completion method based on spatiotemporal privacy protection according to claim 1, characterized in that: During the diffusion process of the completion model, Gaussian noise is added to the spatiotemporal data to be completed, which is expressed as follows: ； ； in, To generate the joint probability distribution of the noise sequence given the original data; is the noise data sequence from step 1 to step T; is the original data; is the total step length; For a certain step; is the noise data of step t; for arrive The single-step diffusion conditional probability distribution of ; is a Gaussian distribution; is the noise variance at step t; is the mean of the Gaussian distribution, which represents the scaling of the previous step data; is the variance.

3. The data completion method based on spatiotemporal privacy protection according to claim 1, characterized in that: The inverse process of the completion model is to gradually convert random noise into time-space consistent missing values ​​based on conditional information. The inverse process is expressed as follows: ； ； in, is the denoised data sequence from step 0 to step T-1; is the pure noise data of the Tth step; is conditional information; is the geographic adjacency matrix; To generate the joint probability distribution of denoised sequence given the noise data and conditional information; is the total step length; For a certain step; is the Markov property of the reverse process, indicating that the joint distribution can be decomposed into the product of the conditional probabilities at each step; To predict the denoised data of the previous step; is a Gaussian distribution; is the mean value predicted by the neural network; is the variance matrix.

4. The data completion method based on spatiotemporal privacy protection according to claim 1, characterized in that: The completion model also includes a conditional feature extraction module and a noise estimation module; The conditional feature extraction module is used to extract features from the confusion probability matrix, the obfuscated spatiotemporal data, and the spatiotemporal data to be completed to obtain extracted conditional features; The noise estimation module is used to perform noise prediction on the real data with noise, the obfuscated spatiotemporal data and the extracted conditional features to obtain a noise prediction result.

5. The data completion method based on spatiotemporal privacy protection according to claim 4, characterized in that: The conditional feature extraction module includes: a spatial attention unit, a temporal attention unit and a graph convolutional network unit; The spatial attention unit is used to extract global spatial correlation from the confusion probability matrix, the obfuscated spatiotemporal data, and the spatiotemporal data to be completed; The temporal attention unit is configured to adjust and model a time series dynamic pattern of the confusion probability matrix, the confused spatiotemporal data, and the spatiotemporal data to be completed; The graph convolutional network unit is used to fuse geographic adjacency information.

6. The data completion method based on spatiotemporal privacy protection according to claim 4, characterized in that: The extracted conditional features are: ； ； ； ； ； in, is the global spatiotemporal feature; The feature representation of the interpolated observation data after 1×1 convolution; is the geographic adjacency matrix; () is a multi-layer perceptron, used to fuse the three features; for spatial attention; For time attention; It is a graph neural network; () is the normalization layer; Multi-head self-attention mechanism for spatial dimension; It is a multi-head self-attention mechanism in the time dimension; for message passing neural networks; is the space-time data after confusion; is the confusion probability matrix; It is a convolutional network; is the layer; Att() is the global attention mechanism; spa is the spatial attention; tem is the temporal attention; Conv is the convolutional layer.

Images