Local differential privacy method and system for internet of things time series data publishing

Abstract

The application provides a local differential privacy method and system for Internet of Things time series data publishing, and belongs to the technical field of data privacy protection. The scheme breaks the limitation of traditional independent disturbance values or time stamps, and protects the "time-value" pair as an indivisible privacy unit by co-disturbing the time stamp and the value. Secondly, the proportional-integral-derivative (PID) control framework is innovatively introduced to model the dynamic behavior of time series data, realizing fine-grained adaptive disturbance intensity allocation based on the data change mode. Finally, by designing a comprehensive utility function combining time offset, value offset and time-value correlation offset, and using the efficient sampling strategy of tournament selection in the candidate set, the key dynamic patterns and correlation structures in the original time series data are maximized while ensuring strict and provable local differential privacy protection.

Description

Technical Field

[0001] This invention belongs to the field of data privacy protection technology, and in particular relates to a local differential privacy method and system for publishing time-series data in the Internet of Things. Background Technology

[0002] With the rapid development and widespread adoption of IoT technology, time-series data has experienced explosive growth and continuous generation. This data, continuously collected by various sensors and smart devices, holds immense value and broad application prospects, becoming an indispensable key factor in fields such as smart cities, smart healthcare, and the Industrial Internet of Things, paving the way for applications such as intelligent transportation planning, behavioral pattern analysis, and personalized services.

[0003] However, mining time-series data also carries serious risks of privacy breaches. Unlike independent, static data points, time-series data records the continuous state of an individual or system over time. Its strong continuity and high correlation are precisely the main reasons for privacy leaks. Attackers can not only obtain sensitive values ​​at individual points in time (such as real-time location, household energy consumption, and physiological indicators), but also infer deeper levels of privacy information such as user behavior, lifestyle patterns, and even identity by analyzing its evolution trends.

[0004] To address the aforementioned risks, existing research largely employs Differential Privacy (DP) techniques for privacy protection. Traditional methods typically assume the existence of a trusted data manager that centrally processes user data and injects noise. However, this assumption carries credibility risks in practical applications; service providers may leak data due to commercial interests or system vulnerabilities. Therefore, Local Differential Privacy (LDP), as a privacy protection paradigm with a solid mathematical foundation, is widely used in distributed data collection scenarios. It allows users to perturb data before it leaves the local device, thereby avoiding reliance on a trusted third party and providing robust privacy protection.

[0005] However, directly applying LDP to time-series data protection still faces significant challenges. Existing research generally adopts the approach of "independent perturbation," that is, perturbing the data values ​​or timestamps separately. In terms of specific methods, value perturbation mainly injects noise into the observations at each timestamp to reduce the risk of attackers inferring the true data with high confidence. However, this method of directly perturbing the values ​​inevitably leads to data distortion, thus affecting the reliability of the analysis results in application scenarios where high data accuracy is required. To address this, some researchers have turned to timestamp perturbation, that is, adjusting the time-series information of the event to avoid directly altering the original values, thereby better preserving the authenticity of the values. Although these methods have achieved certain results in their respective protection dimensions, they generally overlook a key fact: the "time-value" pair is an indivisible privacy unit. In real-world scenarios, specific values ​​appearing at specific times (such as zero-energy readings at night or regular location movements during commuting hours) are themselves highly identifiable sensitive patterns. Independent perturbation schemes destroy information in one dimension but retain another dimension as an anchor point for attackers to perform correlation reasoning, making them difficult to effectively defend against complex correlation attacks.

[0006] In addition, most existing privacy protection methods based on LDP adopt a strategy of perturbing each data point independently, failing to fully consider the inherent dynamic correlations and dependencies between consecutive points in time series data. This approach not only easily destroys the key time series patterns (such as trends and periods) contained in the sequence, but also weakens the actual effect of privacy protection to some extent, while reducing the analytical and application value of the perturbed data. Summary of the Invention

[0007] This invention provides a local differential privacy method and system for publishing time-series data in the Internet of Things (IoT).

[0008] According to a first aspect of the present invention, a local differential privacy method for publishing Internet of Things (IoT) time-series data is provided, comprising: The timestamps and values ​​in the original time series data are discretized respectively, and continuous time and values ​​are mapped to a discrete index space. The change pattern of each data point is modeled based on the proportional-integral-derivative control framework, and normalized feature indicators representing the dynamic behavior of the data are extracted. Based on the extracted normalized feature indices, the adaptive perturbation radius is calculated for each data point in both the time and numerical dimensions. Based on the adaptive perturbation radius of the time and numerical dimensions, a joint candidate output set containing time and numerical index pairs is constructed in the discrete index space; In the joint candidate output set, probability sampling is performed through a preset filtering strategy to output perturbed index pairs; The perturbed index pairs are restored to continuous timestamps and values ​​through inverse mapping, and then sorted in chronological order to generate the final published data.

[0009] Furthermore, the extraction of the normalized feature index specifically includes: Calculate the slope error between the current time and the previous time. Construct a weighted combination of proportional error term, integral error term, and differential error term; The combined results are normalized and used as a dynamic importance measure.

[0010] Furthermore, the proportional-integral-derivative control framework is used to model the change pattern of each data point, and normalized feature indicators representing the dynamic behavior of the data are extracted, specifically as follows: ; ; in, These are the proportional, integral, and differential coefficients, respectively. For feedback error, This is the length of the time window for integration. This indicates the dynamic proportional error of the current data; Indicates the integral error; This represents the differential error in predicting the current change. For the first Normalized characteristic index of observations, It is a preset normalization baseline parameter used to normalize the original... Eigenvalues ​​mapped to standard intervals For time indexing, For the first The specific timestamp of each observation.

[0011] Furthermore, based on the extracted normalized feature indices, the adaptive perturbation radius is calculated for each data point in both the time and numerical dimensions, as specifically shown below: ; ; in, For the first The adaptive perturbation radius of each observation in the time dimension For the first The adaptive perturbation radius of each observation in the numerical dimension. and They are respectively The corresponding acceptable minimum and maximum perturbations, For the first Normalized characteristic index of observations, and They are respectively The corresponding acceptable minimum and maximum perturbations.

[0012] Furthermore, the probability sampling through a preset screening strategy specifically includes: Based on the obtained joint candidate output set, a utility score is calculated for each candidate output using a pre-constructed utility function; Based on the obtained utility scores, probability sampling is performed using an index mechanism combined with a tournament selection strategy.

[0013] Furthermore, the utility function is specifically expressed as follows: ; in, This is the time index after the perturbation. For the current number Time index of observations, For the first The adaptive perturbation radius of each observation in the time dimension This represents the normalized distance between the perturbed time index and the current observation time index, relative to the allowable radius. This is the index of the perturbed values. For the current number Numerical index of observations, For the first The adaptive perturbation radius of each observation in the numerical dimension. This represents the normalized distance between the perturbed numerical index and the currently observed numerical index, relative to the allowable radius. The normalized result is used to measure the degree of deviation of the point after the perturbation from the original linear trend; and To adjust the parameters.

[0014] Furthermore, the discretization processing of the timestamps and values ​​in the original time-series data specifically includes: Numerical discretization: Determine the numerical range of the original data, map the values ​​to the interval [0,1], divide them into M equal segments, and then convert them into integer indices in the range [0, M-1]. Timestamp discretization: Determine the baseline reference time, convert continuous timestamps into integer indices based on the preset time discretization granularity, and perform boundary pruning on the indices to limit them to the range of valid indices.

[0015] According to a second aspect of the present invention, a local differential privacy system for publishing IoT time-series data is provided, comprising: The data preprocessing unit is used to discretize the timestamps and values ​​in the original time series data, mapping continuous time and values ​​to a discrete index space. The dynamic feature extraction unit is used to model the change pattern of each data point based on the proportional-integral-derivative control framework and extract normalized feature indicators that represent the dynamic behavior of the data. The adaptive perturbation radius calculation unit is used to calculate the adaptive perturbation radius for each data point in both the time and numerical dimensions based on the extracted normalized feature indices. A joint candidate set construction unit is used to construct a joint candidate output set containing time and numerical index pairs in a discrete index space based on the adaptive perturbation radius of the time and numerical dimensions. A probability sampling unit is used to perform probability sampling in the joint candidate output set through a preset screening strategy and output perturbed index pairs. The data reconstruction unit is used to restore the perturbed index pairs to continuous timestamps and values ​​through inverse mapping, and then sort them in chronological order to generate the final published data.

[0016] According to a third aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the aforementioned local differential privacy method for publishing time-series data for the Internet of Things.

[0017] According to a fourth aspect of the present invention, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the aforementioned local differential privacy method for publishing time-series data for the Internet of Things.

[0018] The above one or more technical solutions have the following beneficial effects: (1) This invention provides a local differential privacy method and system for publishing IoT time-series data. The solution effectively solves three core problems of time-series data privacy protection in the prior art through a self-designed "time-value" joint perturbation framework and dynamic adaptive mechanism. First, it breaks the limitations of traditional independent perturbation of values ​​or timestamps. By coordinating the perturbation of timestamps and values, the "time-value" pair is protected as an indivisible privacy unit, fundamentally resisting attacks by attackers using a single dimension as an anchor point for correlation reasoning. Second, it innovatively introduces a proportional-integral-derivative (PID) control framework to model the dynamic behavior of time-series data, realizing fine-grained adaptive perturbation intensity allocation based on the data's own change patterns, rather than using a uniform or random perturbation strategy for the entire sequence. Finally, by designing a comprehensive utility function that combines time offset, numerical offset, and time-value correlation offset, and employing an efficient sampling strategy of tournament screening in the candidate set, the scheme of this invention preserves the key dynamic patterns and correlation structures in the original time series data to the maximum extent while ensuring strict and provable local differential privacy protection. This significantly enhances the analysis and application value of the perturbed data and provides reliable technical support for privacy-secure data sharing in high-value IoT applications such as smart cities and smart healthcare.

[0019] (2) Considering that the existing schemes independently perturb the timestamp or the value, which destroys the integrity of the "time-value" pair, and attackers can still use the other dimension of information that is not disturbed or slightly disturbed to infer the user's sensitive behavior patterns through correlation analysis, the present invention constructs a joint candidate set containing time and value index pairs for each data point in a discretized index space, and performs cooperative probability sampling on this set, ensuring that the perturbed output changes both time and value information. This method makes it impossible for attackers to reliably anchor the exact time of occurrence of the original data point and its corresponding specific value from the published data, thereby effectively blocking complex reasoning attacks based on time-series correlation and providing stronger privacy protection than the traditional independent perturbation method.

[0020] (3) Existing methods typically apply noise of the same intensity to all data points in a sequence or employ simple random strategies, ignoring the differences in the importance of the dynamic behavior of the data points themselves. This leads to excessive perturbation of stable or normal data, wasting the privacy budget, while failing to adequately protect key points of change. This invention introduces a PID control framework to calculate the proportional error (instantaneous fluctuation), integral error (cumulative trend), and differential error (acceleration of change) of each data point in real time, and normalizes them into a dynamic importance metric. Based on this metric, this invention adaptively allocates the perturbation radius of time and numerical dimensions to each data point. Through this "on-demand allocation" strategy, the limited privacy budget is utilized more efficiently. Under the same level of privacy protection, key dynamic patterns such as trends and inflection points in the original data are preserved to the greatest extent, significantly improving the analytical accuracy and usability of the published data.

[0021] (4) Considering that the traditional LDP method may lead to excessive computational complexity or difficulty in implementation when directly applying the exponential mechanism in continuous or high-dimensional space, this invention transforms the continuous and infinite output space into a finite and enumerable discrete index space through the initial discretization process, making it possible to construct a bounded joint candidate set. At the same time, the comprehensive utility function designed in this invention not only considers the offset of time and value, but also innovatively adds a "time-value correlation offset" term to punish those perturbation results that seriously deviate from the linear trend of the original data, thereby actively maintaining the shape correlation of the sequence during the perturbation process. Furthermore, the tournament screening strategy adopted in this invention replaces the traditional full candidate set weight calculation and sampling, and efficiently screens high-weight high-quality candidates through multiple rounds of small-scale competitions, greatly reducing the computational overhead, which is particularly suitable for resource-constrained IoT terminal devices.

[0022] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0023] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0024] Figure 1 This is a flowchart of a local differential privacy method for publishing IoT time-series data, as described in an embodiment of the present invention. Figure 2 This is a schematic diagram of a local differential privacy system architecture for publishing IoT time-series data, as described in an embodiment of the present invention. Figure 3 This is a flowchart of the tournament selection process described in an embodiment of the present invention. Detailed Implementation

[0025] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0026] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0027] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0028] Example 1

[0029] In one or more embodiments, such as Figure 1 As shown, a local differential privacy method for publishing IoT time-series data includes: The timestamps and values ​​in the original time series data are discretized respectively, and continuous time and values ​​are mapped to a discrete index space. The change pattern of each data point is modeled based on the proportional-integral-derivative control framework, and normalized feature indicators representing the dynamic behavior of the data are extracted. Based on the extracted normalized feature indices, the adaptive perturbation radius is calculated for each data point in both the time and numerical dimensions. Based on the adaptive perturbation radius of the time and numerical dimensions, a joint candidate output set containing time and numerical index pairs is constructed in the discrete index space; In the joint candidate output set, probability sampling is performed through a preset filtering strategy to output perturbed index pairs; The perturbed index pairs are restored to continuous timestamps and values ​​through inverse mapping, and then sorted in chronological order to generate the final published data.

[0030] In one or more embodiments, the discretization processing of the timestamps and values ​​in the original time-series data specifically includes: Numerical discretization: Determine the numerical range of the original data, map the values ​​to the interval [0,1], divide them into M equal segments, and then convert them into integer indices in the range [0, M-1]. Timestamp discretization: Determine the baseline reference time, convert continuous timestamps into integer indices based on the preset time discretization granularity, and perform boundary pruning on the indices to limit them to the range of valid indices.

[0031] Specifically, the process of discretizing the timestamps and values ​​in the original time-series data, mapping continuous time and values ​​to a discrete index space, includes the following steps: Data value discretization: First, a baseline is established for the discretization process by determining the numerical range of the original data. If historical data or prior distributions are lacking, the static scanning method or sliding window online estimation method can be used, depending on whether the data evolves over time and whether dynamic adjustment of the range is required. Based on this, the original numerical values ​​are discretized using the following formula. Mapped to Interval: ; in, and These represent the minimum and maximum values ​​of the data, respectively; then the continuous intervals... Divided into equal parts The width of each segment is defined by the following formula: ; Further discretization into integer indices can be expressed as follows: ; in, The preset discretization level controls the numerical granularity.

[0032] Timestamp discretization: First, determine the reference point for timestamp discretization, selecting the earliest timestamp of the entire sequence or midnight of each day as the reference time. This unifies the starting point for all timestamp measurements. Subsequently, based on the time precision requirements of actual application scenarios, the discretization granularity is set. This parameter can be set to units such as seconds, minutes, or hours as needed, and its value directly affects the accuracy of the time index and the complexity of subsequent perturbation calculations. Based on this, continuous timestamps are converted into regularized time indices using the following formula. : ; Next, to limit the index to a valid range, the entire publication period is calculated. Maximum time index within: ; Finally, the calculated time index Perform boundary clipping: if the calculated index Then cut ;like Then cut Through the above processing, all original timestamps are normalized into integer indices within a finite range, laying the foundation for subsequent joint perturbation in the discrete index space.

[0033] In one or more embodiments, the extraction of the normalized feature index specifically includes: Calculate the slope error between the current time and the previous time. Construct a weighted combination of proportional error term, integral error term, and differential error term; The combined results are normalized and used as a dynamic importance measure.

[0034] In one or more embodiments, the proportional-integral-derivative control framework is used to model the change pattern of each data point and extract normalized feature indicators representing the dynamic behavior of the data, specifically as follows: ; ; in, These are the proportional, integral, and differential coefficients, respectively. For feedback error, This is the length of the time window for integration. This indicates the dynamic proportional error of the current data; Indicates the integral error; This represents the differential error in predicting the current change. For the first Normalized characteristic index of observations, It is a preset normalization baseline parameter used to normalize the original... Eigenvalues ​​mapped to standard intervals For time indexing, For the first The specific timestamp of each observation.

[0035] In practical implementation, the proportional-integral-derivative control framework is used to model the change pattern of each data point and extract normalized feature indicators representing the dynamic behavior of the data. This process includes the following steps: To quantify the dynamic behavior characteristics of time-series data, a proportional, integral, and derivative (PID) control framework is introduced to model the change pattern of each data point. A dynamic importance metric is constructed by comprehensively analyzing instantaneous fluctuations, cumulative trends, and accelerations. First, the data is represented as a slope, and the error between the slope at the current time step and the slope at the previous time step is used as the feedback error. ; in, It is the first The slope at each timestamp is represented as follows: ; Furthermore, the PID control characteristics are constructed, and their expressions are as follows: ; in, Let these represent the proportional, integral, and differential coefficients (empirical values ​​are 0.5, 0.3, and 0.2, respectively). The first term... The second term represents the dynamic proportional error of the current data. It is the integral error, where the parameter The time window length for integration is defined to set the historical time range covered by the accumulated error of the integration term; the third term This represents the differential error in predicting the current change. Finally, the PID characteristics are converted into standardized metrics, normalized, and calculated as follows: ; In one or more embodiments, the step of calculating the adaptive perturbation radius for each data point in both the time and numerical dimensions based on the extracted normalized feature indices is specifically expressed as follows: ; ; in, For the first The adaptive perturbation radius of each observation in the time dimension For the first The adaptive perturbation radius of each observation in the numerical dimension. and They are respectively The corresponding acceptable minimum and maximum perturbations, For the first Normalized characteristic index of observations, and They are respectively The corresponding acceptable minimum and maximum perturbations.

[0036] Specifically, the adaptive perturbation radius is calculated for each data point in both the time and numerical dimensions, including the following processing steps: First, set the time disturbance radius. sum of numerical perturbation radius Upper and lower limits: ; ; in, Indicates the first The perturbation radius of an observation in the time dimension, that is, in the index space, limits the maximum discrete offset step size of the time index relative to the original position; Indicates the first The perturbation radius of an observation in the numerical dimension, i.e. in the index space, limits the maximum discrete offset step size of the numerical index relative to the original position. and This determines the size of the local neighborhood that the index mechanism will enumerate and randomly sample in the index space.

[0037] Next, based on the normalized dynamic features obtained earlier, linear interpolation is used to map them to... Upper and lower boundaries and The time radius for dynamically updating data points corresponds to the minimum and maximum acceptable perturbations, respectively. and numerical radius The above formula represents this.

[0038] In one or more embodiments, the step of constructing a joint candidate output set containing time and numerical index pairs in the discrete index space based on the adaptive perturbation radius of the time and numerical dimensions specifically includes the following processing steps: After determining the adaptive perturbation radius for each observation, a candidate output set conforming to the sampling is defined and constructed in the index space based on this pair of radii. The specific steps are as follows: Given a set of time indices and numeric indices: ; First, calculate the unpruned interval: [ ; Then, restrict them to the valid index range respectively: ; Therefore, time candidate set candidate set of sum values This can be formalized as: ; ; Finally, a bounded and enumerable joint candidate set is obtained. The subsequent index mechanism will perform weight calculation and sampling on the candidate set according to a predefined normalized utility function.

[0039] In one or more embodiments, the probability sampling through a preset screening strategy specifically includes: Based on the obtained joint candidate output set, a utility score is calculated for each candidate output using a pre-constructed utility function; Based on the obtained utility scores, probability sampling is performed using an index mechanism combined with a tournament selection strategy.

[0040] In specific implementation, the probability sampling through a preset screening strategy involves the following processing steps: First, calculate the utility score. Based on the constructed candidate set, for any candidate pair... It is necessary to process each candidate output within this local neighborhood. A score is assigned so that the index mechanism can allocate sampling probabilities based on the score. Specifically, the three distances—"time offset," "numerical offset," and "time-value correlation offset"—are weighted and normalized, and a utility function is constructed by taking a negative sign. The definition of the function is as follows: ; in, This is the time index after the perturbation. For the current number Time index of observations, For the first The adaptive perturbation radius of each observation in the time dimension This represents the normalized distance between the perturbed time index and the current observation time index, relative to the allowable radius. This is the index of the perturbed values. For the current number Numerical index of observations, For the first The adaptive perturbation radius of each observation in the numerical dimension. This represents the normalized distance between the perturbed numerical index and the currently observed numerical index, relative to the allowable radius. The normalized result is used to measure the degree of deviation of the point after the perturbation from the original linear trend; and To adjust the parameters, where Used to weigh the priorities of time and numerical values. Larger values ​​take priority over time, and vice versa. Factors used to control and ensure correlation. The larger the value, the more likely it is to preserve the original sequence shape; At this time, we only focus on the respective offsets of time and value.

[0041] Based on the utility function, for each candidate output After calculating the real utility value, this utility value is then mapped to a positive weight, which is called the "normalized weight," and is calculated as follows: ; in, It is a privacy budget, used to quantify the level of privacy protection. The smaller the size, the stronger the privacy protection. The larger the value, the weaker the privacy protection. This represents the sensitivity of the utility function, which measures the maximum impact that adding or deleting a single record might have on any output score. The formula indicates that the higher the utility function score, the higher its weight, and the higher the probability of it being selected during sampling.

[0042] like Figure 3 As shown, the selection process for multiple tournaments is as follows: First, initialize an empty list of winners. And set the total number of rounds of the tournament to be The number of candidates participating in each round is For each round of the tournament From the joint candidate set of this observation point Randomly sample without replacement and select... Groups of different candidate index pairs Then, the utility score and its corresponding non-normalized weight for each pair of participating candidates are calculated. In this round of the tournament, the pair with the highest weight is selected. As the winner, add him to the list of winners. In the middle. (After completion) Round the tournament and build a list of winners The system will recalculate the nonnormalized weight of each winning element in the list. : ; Since the size of the candidate space is directly determined by the adaptive perturbation radius, as the radius increases, the feasible region expands rapidly, leading to a sharp increase in the number of candidate points and a dramatic expansion of the search space. This, in turn, significantly increases the number of sampling and evaluation operations for candidate points in each iteration. Considering the computing power limitations and real-time publishing requirements of IoT terminals, the system introduces a quantized judgment threshold based on the size of the candidate space. (generally defined in) When the size of the candidate set exceeds this threshold, it is considered a larger candidate set; otherwise, it is considered a smaller candidate set. Based on this division, we provide two efficient sampling schemes for embodiments of this solution: If the candidate set is small, explicit normalized sampling can be used. After obtaining the non-normalized weights of the winners list, the total weights can be calculated. :

[0043] Normalized to the choice probability: ; Finally, sampling can be performed according to the probability.

[0044] If the candidate set is large and suitable for online candidate set generation, a prefix sum sampling method can be used to calculate the total weight after obtaining the non-normalized weights of the winners list. :

[0045] Generate a threshold This involves generating a uniformly random number. This random value is used as a "fishhook" to fish on a "weighted roulette wheel"—the interval with the larger the weight, the higher the probability of being hooked.

[0046] ; Then, an accumulated weight is defined. We add the weights to each candidate in turn. When the accumulated weights exceed the threshold for the first time, we stop and return that index. This index represents the candidate pair we need. .

[0047] In one or more embodiments, the step of restoring the perturbed index pairs to continuous timestamps and values ​​through inverse mapping, and then sorting them in chronological order to generate the final published data, specifically involves the following processing steps: First, in terms of numerical dimensions, based on the numerical range and discretization level recorded during the discretization stage, the perturbed numerical indexes are restored to their actual values. Regarding timestamp reconstruction, based on the baseline time and time granularity used in discretization, the perturbed time indexes are converted into specific time points. It should be noted that the perturbation operation may disrupt the original temporal sequence; therefore, after completing the bidirectional reconstruction of numerical and temporal data, the system reorders all data points according to the generated timestamps to restore the correct temporal logic, ultimately forming published data with temporal consistency.

[0048] Example 2

[0049] In one or more embodiments, corresponding to the above method embodiments, such as Figure 2 As shown, this embodiment provides a local differential privacy system for publishing IoT time-series data, including: The data preprocessing unit is used to discretize the timestamps and values ​​in the original time series data, mapping continuous time and values ​​to a discrete index space. The dynamic feature extraction unit is used to model the change pattern of each data point based on the proportional-integral-derivative control framework and extract normalized feature indicators that represent the dynamic behavior of the data. The adaptive perturbation radius calculation unit is used to calculate the adaptive perturbation radius for each data point in both the time and numerical dimensions based on the extracted normalized feature indices. A joint candidate set construction unit is used to construct a joint candidate output set containing time and numerical index pairs in a discrete index space based on the adaptive perturbation radius of the time and numerical dimensions. A probability sampling unit is used to perform probability sampling in the joint candidate output set through a preset screening strategy and output perturbed index pairs. The data reconstruction unit is used to restore the perturbed index pairs to continuous timestamps and values ​​through inverse mapping, and then sort them in chronological order to generate the final published data.

[0050] It is understood that the system described in this embodiment corresponds one-to-one with the method described in Embodiment 1, and its technical details have been described in detail in Embodiment 1, so they will not be repeated here.

[0051] In further embodiments, the following is also provided: An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the methods described in the above embodiments. For brevity, further details are omitted here.

[0052] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0053] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0054] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the methods described in the above embodiments.

[0055] The methods described in the above embodiments can be directly executed by a hardware processor, or executed by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0056] Those skilled in the art will recognize that the units, i.e., algorithm steps, of the various examples described in connection with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0057] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Various modifications and variations can be made to this disclosure by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A local differential privacy method for publishing IoT time-series data, characterized in that, include: The timestamps and values ​​in the original time series data are discretized respectively, and continuous time and values ​​are mapped to a discrete index space. The change pattern of each data point is modeled based on the proportional-integral-derivative control framework, and normalized feature indicators representing the dynamic behavior of the data are extracted. Based on the extracted normalized feature indices, the adaptive perturbation radius is calculated for each data point in both the time and numerical dimensions. Based on the adaptive perturbation radius of the time and numerical dimensions, a joint candidate output set containing time and numerical index pairs is constructed in the discrete index space; In the joint candidate output set, probability sampling is performed through a preset filtering strategy to output perturbed index pairs; The perturbed index pairs are restored to continuous timestamps and values ​​through inverse mapping, and then sorted in chronological order to generate the final published data.

2. The local differential privacy method for publishing IoT time-series data as described in claim 1, characterized in that, The extraction of the normalized feature indicators specifically includes: Calculate the slope error between the current time and the previous time. Construct a weighted combination of proportional error term, integral error term, and differential error term; The combined results are normalized and used as a dynamic importance measure.

3. The local differential privacy method for publishing IoT time-series data as described in claim 1, characterized in that, The proportional-integral-derivative control framework models the change pattern of each data point and extracts normalized feature indicators representing the dynamic behavior of the data, specifically as follows: ； ； in, These are the proportional, integral, and differential coefficients, respectively. For feedback error, This is the length of the time window for integration. This indicates the dynamic proportional error of the current data; Indicates the integral error; This represents the differential error in predicting the current change. For the first Normalized characteristic index of observations, These are preset normalization reference parameters used to normalize the original data. Eigenvalues ​​mapped to standard intervals For time indexing, For the first The specific timestamp of each observation.

4. The local differential privacy method for publishing IoT time-series data as described in claim 1, characterized in that, The adaptive perturbation radius is calculated for each data point in both the time and numerical dimensions based on the extracted normalized feature indices, as specifically shown below: ； ； in, For the first The adaptive perturbation radius of each observation in the time dimension For the first The adaptive perturbation radius of each observation in the numerical dimension. and They are respectively The corresponding acceptable minimum and maximum perturbations, For the first Normalized characteristic index of observations, and They are respectively The corresponding acceptable minimum and maximum perturbations.

5. The local differential privacy method for publishing IoT time-series data as described in claim 1, characterized in that, The probability sampling through a preset screening strategy specifically includes: Based on the obtained joint candidate output set, a utility score is calculated for each candidate output using a pre-constructed utility function; Based on the obtained utility scores, probability sampling is performed using an index mechanism combined with a tournament selection strategy.

6. The local differential privacy method for publishing IoT time-series data as described in claim 5, characterized in that, The utility function is specifically represented as follows: ； in, This is the time index after the perturbation. For the current number Time index of observations, For the first The adaptive perturbation radius of each observation in the time dimension This represents the normalized distance between the perturbed time index and the current observation time index, relative to the allowable radius. This is the index of the perturbed values. For the current number Numerical index of observations, For the first The adaptive perturbation radius of each observation in the numerical dimension. This represents the normalized distance between the perturbed numerical index and the currently observed numerical index, relative to the allowable radius. The normalized result is used to measure the degree of deviation of the point after the perturbation from the original linear trend; and To adjust the parameters.

7. The local differential privacy method for publishing IoT time-series data as described in claim 1, characterized in that, The discretization process for the timestamps and values ​​in the original time-series data specifically includes: Numerical discretization: Determine the numerical range of the original data, map the values ​​to the interval [0,1], divide them into M equal segments, and then convert them into integer indices in the range [0, M-1]. Timestamp discretization: Determine the baseline reference time, convert continuous timestamps into integer indices based on the preset time discretization granularity, and perform boundary pruning on the indices to limit them to the range of valid indices.

8. A local differential privacy system for publishing IoT time-series data, characterized in that: include: The data preprocessing unit is used to discretize the timestamps and values ​​in the original time series data, mapping continuous time and values ​​to a discrete index space. The dynamic feature extraction unit is used to model the change pattern of each data point based on the proportional-integral-derivative control framework and extract normalized feature indicators that represent the dynamic behavior of the data. The adaptive perturbation radius calculation unit is used to calculate the adaptive perturbation radius for each data point in both the time and numerical dimensions based on the extracted normalized feature indices. A joint candidate set construction unit is used to construct a joint candidate output set containing time and numerical index pairs in a discrete index space based on the adaptive perturbation radius of the time and numerical dimensions. A probability sampling unit is used to perform probability sampling in the joint candidate output set through a preset screening strategy and output perturbed index pairs. The data reconstruction unit is used to restore the perturbed index pairs to continuous timestamps and values ​​through inverse mapping, and then sort them in chronological order to generate the final published data.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements a local differential privacy method for publishing IoT time-series data as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements a local differential privacy method for publishing IoT time-series data as described in any one of claims 1-7.

Images