CN112231749B - Distributed single-dimensional time sequence data real-time privacy protection publishing method with consistency - Google Patents
Distributed single-dimensional time sequence data real-time privacy protection publishing method with consistency Download PDFInfo
- Publication number
- CN112231749B CN112231749B CN202011097748.XA CN202011097748A CN112231749B CN 112231749 B CN112231749 B CN 112231749B CN 202011097748 A CN202011097748 A CN 202011097748A CN 112231749 B CN112231749 B CN 112231749B
- Authority
- CN
- China
- Prior art keywords
- data
- sampling
- node
- time
- information
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/60—Protecting data
- G06F21/62—Protecting access to data via a platform, e.g. using keys or access control rules
- G06F21/6218—Protecting access to data via a platform, e.g. using keys or access control rules to a system of files or objects, e.g. local or distributed file system or database
- G06F21/6245—Protecting personal data, e.g. for financial or medical purposes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
- G06Q50/06—Electricity, gas or water supply
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2113/00—Details relating to the application field
- G06F2113/04—Power grid distribution networks
Abstract
The invention discloses a distributed single-dimensional time sequence data real-time privacy protection publishing method with consistency, and belongs to the field of privacy protection. The method adjusts sampling frequency through a self-adaptive sampling strategy, perturbs sampling data through a Laplacian mechanism at sampling time, then performs distributed posterior estimation on published data of each node at the sampling time through a single-dimensional perception information correction strategy based on Kalman consistency filtering, and directly publishes prior predicted data at non-sampling time, so that a distributed single-dimensional time sequence perception data real-time publishing method which is independent of a central server and has differential privacy guarantee and consistency is finally realized, and the real-time requirement of dynamic data publishing and the consistency requirement of distributed data publishing are met. The method has a good application effect in an actual data scene, and can be used in distributed dynamic perception data distribution systems such as a distributed power utilization load monitoring system of an intelligent power grid and disease monitoring application.
Description
Technical Field
The invention belongs to the field of privacy protection, and particularly relates to a distributed single-dimensional time sequence data real-time privacy protection issuing method with consistency.
Background
With the rapid development of the internet of things technology and the continuous improvement of the hardware level, people's production and life are more and more digital, intelligent and networked, and various intelligent devices fill people's daily life to perceive data closely related to people themselves and the surrounding environment all the time. The system can collect, release and mine a large amount of user perception data, can provide all-round information and knowledge, and provides more specialized and personalized services for people. For example, the information such as browsing, shopping records and book image scores of the user is gathered and released, and a recommendation scheme which is more convenient for people to give preferential offers can be made; the method and the system can be used for gathering and releasing the environmental information or the position track collected by the smart phone, and can be used for digital planning of cities, intelligent traffic service and the like. Thus, the distribution of such perceptually aggregated data has tremendous application, whether to a business organization, research institution or government.
However, the convergence and distribution of sensory data provides services and convenience to people, as well as presents an unprecedented privacy concern. The data contains a large amount of personal sensitive information, such as disease information, consumption habits, location tracks and the like, which is likely to be leaked along with the distribution of the data. With the rapid development of technologies such as data mining and machine learning, there is a high possibility that data that appears to be secure indirectly reveals privacy. Research indicates that the identity of a mobile phone user can be identified in a correlation manner by combining flow information, time information and social network information of a mobile phone network; further, studies have demonstrated that a single user can be uniquely identified by performing correlation analysis using a plurality of pieces of attribute information other than IDs, and that even deletion of personal identification information (such as name, ID, and the like) is insufficient to secure private information. Therefore, the privacy protection problem in the data distribution application is very important and cannot be ignored.
Most of the existing privacy protection publishing technologies aim at a centralized perception data convergence system and solve the problem of publishing a static data set. And the rapidly developed application technology of the internet of things and the increasingly serious privacy disclosure problem put forward higher requirements on the privacy protection release technology. On the one hand, in recent years, data mining is more and more commonly applied to a time sequence data perception system, such as traffic flow monitoring, real-time financial analysis, epidemic situation monitoring, intelligent power grid power load monitoring and the like. The applications can provide real-time feedback results by mining and analyzing the dynamically updated and issued time sequence perception convergence data, so that the most accurate information is provided at present. On the other hand, with the continuous development of the internet of things technology, the system scale and the data scale of various sensing applications are continuously enlarged, and the centralized sensing data aggregation system cannot meet the era requirements of huge data volume and high information sharing. Therefore, more and more applications adopt a distributed system architecture for data aggregation and distribution, including many sensing applications based on dynamic time series data. However, in the application of distributed dynamic time-series aware data distribution, since time-series data contains rich time correlation, once being utilized by an attacker, more information can be mined out besides the data itself, resulting in more privacy being exposed. Meanwhile, distributed data publishing applications generally need to ensure the cooperation and consistency of data published by each node, so that operations such as communication, data sharing and the like need to be performed between different nodes, and these frequent and complicated operations may bring more attack opportunities to attackers, further aggravating the threat of privacy exposure. Thus, the privacy protection problem in distributed dynamic time-series aware data distribution applications is more severe and urgent.
Disclosure of Invention
Aiming at the problem of privacy protection and release of distributed time sequence perception data, the existing related researches cannot effectively solve the consistency requirement in the dynamic release process of the distributed data while improving the effectiveness of continuously released data. Therefore, the invention aims to provide a method for issuing the real-time privacy protection of the distributed single-dimensional time sequence data with consistency aiming at issuing the most commonly applied single-dimensional limited time sequence perception data. The method realizes the real-time property, high utility and consistency of the private data release in the decentralized distributed sensing system. The distributed dynamic sensing data distribution system can be used in a distributed power utilization load monitoring system of a smart power grid, a disease monitoring application and other distributed dynamic sensing data distribution systems.
In order to achieve the purpose, the invention is realized by adopting the following technical scheme:
a distributed single-dimensional time sequence data real-time privacy protection publishing method with consistency adjusts sampling frequency through a self-adaptive sampling strategy, perturbs sampling data through a Laplacian mechanism at the sampling moment, then performs distributed posterior estimation on the published data of each node at the sampling moment through a single-dimensional perception information correction strategy based on Kalman consistency filtering, and directly publishes prior predicted data at the non-sampling moment, and finally realizes a distributed single-dimensional time sequence perception data real-time publishing method which is independent of a central server and has differential privacy guarantee and consistency, and is used in a distributed dynamic perception data publishing system of a distributed power utilization load monitoring system and a disease monitoring application of an intelligent power grid, and specifically comprises the following steps:
1) Modeling data: taking a statistic convergence issuing scene as an example, defining a real-time privacy protection issuing problem of distributed single-dimensional time sequence data, and modeling a sensing node network interaction condition;
2) Initialization: presetting maximum sampling times M and a maximum privacy budget epsilon according to actual conditions and requirements, and allocating the privacy budget epsilon to each sampling point in advance according to an average allocation principle k The following steps are all performed locally at each node;
3) And (3) perception data sampling: each node determines whether the current moment k samples the input time sequence data or not through the latest sampling interval;
4) Perception data disturbance: the sampled node i truly senses data x of the node i through a Laplace algorithm LPA i (k) Disturbing to obtain a disturbance value z of the node satisfying the differential privacy guarantee i (k) Directly taking the released data at the previous moment as disturbance observation data at the current moment by the non-sampling node;
5) Perception information interaction: all nodes including non-sampling nodes locally calculate the prior estimated value of the node i through a Kalman consistency filtering-local information updating algorithm KCF-update
And broadcast information message i (k) Broadcasting to all neighbor nodes, and receiving broadcast information from the neighbor nodes at the same time, wherein the interactive information does not relate to real sampling data of each node;
6) And (3) correcting the perception information: the sampling node realizes the correction of local disturbance data by using the prior information and the sensing information of the neighbor node, performs posterior estimation and release on a sensing target at the current moment k through a Kalman consistency filtering-single-dimensional distributed estimation algorithm KCF-correction, and the non-sampling node directly releases the prior estimation at the moment and completes the iteration;
7) Perceptual update feedback: and the Sampling node feeds the prior posterior estimation error back to the PID controller, updates the Sampling interval and completes the iteration by an Adaptive Sampling algorithm, and if the Sampling frequency exceeds the maximum Sampling frequency M, the system stops Sampling.
The invention further improves the method that the specific operation of the step 1) is as follows: assuming that in a distributed sensing system with n nodes, each sensing node carries out real-time statistical aggregation on the same sensing target information to form a distributed single-dimensional sensing data sequence X of the ith node i ＝{x i (1),…,x i (k),…,x i (T) }, k =1, \8230, T, where x i (k) For sensing the sensing data of a node i at a moment k, namely the convergence information of the node i to all user individuals in a convergence range, T is the length of a time sequence, and the actual statistical sequence of sensing target information is assumed to be R = { R (1), \8230, R (k), \8230, R (T) }, wherein R (k) is the actual statistical value of the sensing target at a moment discrete time k, and the sensing data x of the node i at the moment k is the sensing data x of the sensing target at the moment discrete time k i (k) The theoretical relationship with the perception target actual data r (k) can be expressed as
x i (k)＝H i ·r(k)
Wherein H i For the perception coefficient of a node i, the privacy protection issuing method aims at issuing accurate perception target statistical data in real time by a distributed node in the dynamic convergence process and protecting the privacy of individual users, because the perception data of a single node is often not comprehensive and accurate and the issued data of the distributed node is usually inconsistent, in a distributed perception system without a central convergence server, in order to achieve the comprehensive, accurate and consistent dynamic issued information of each node, each node is communicated with other nodes and operated cooperatively in the privacy protection processing and issuing process, each node is connected with certain adjacent nodes according to communication conditions to carry out local interaction and does not interact with all other nodes in the perception system pairwise, G = (V, E) is adopted to model the network interaction condition of the perception nodes, wherein V = {1,2, \\ 8230, n } is an edge set, E is an edge set, every two nodes capable of communicating with each other are connected by an edge, namely adjacent nodes, in order to reduce the network end-to-end communication load, the perception system is assumed that each node is only communicated with the adjacent nodes in the distributed node at the perception time, and the privacy protection information is attacked by the adjacent nodes, and the adjacent nodes are not communicated by the local interaction of the distributed node, and the adjacent nodes are not communicated by the edge set, and the adjacent nodes in the distributed node before the perception node interaction process of the privacy protection information, and the node, the node is assumed to be communicated with the node i (k) Representing perception data x for node i i (k) Disturbing data after privacy protection; finally, the distributed distribution sequence of each node is denoted as O = { O i (1),…,o i (k),…,o i (T)},i＝1,…,n。
The further improvement of the invention is that the specific operation of the step 2) is as follows: presetting maximum sampling times M and a maximum privacy budget epsilon according to actual conditions and requirements, wherein the privacy budget
The invention further improves that the concrete operation of the step 3) is as follows: if it is currentlyTime k is equal to the latest sampling time sp of node i i And if the current sampling number n is smaller than the maximum sampling number M, sampling the input time sequence data at the current moment k, wherein the current moment is a sampling point, and the sampling number n = n +1, otherwise, not sampling the input time sequence data at the current moment k, and wherein the current moment is a non-sampling point.
The further improvement of the invention is that the specific operation of the step 4) is as follows: sensing data x of ith node needing sampling at k moment i (k) Addition based on sensitivity Δ f And privacy budget ε k Corrected Laplace noise to obtain disturbance data z i (k) I.e. z i (k)＝x i (k)+Lap(0,Δ f /ε k ) Wherein the sensitivity Δ f Usually, the calculation is performed according to a query function in practical application, and a node which does not need to be sampled directly takes the released data at the previous moment as the disturbance observation data at the current moment.
The invention further improves that the concrete operation of the step 5) is as follows: locally calculating the prior estimated value of the ith node through a process model at the moment k by using a Kalman consistency filtering-local information updating algorithm of all nodes
Wherein A is the transfer coefficient of the first and second groups,
calculating a posterior estimated value at the last moment of the ith node through a Kalman consistency estimator; then calculating the information vector u of the ith node i (k) And matrix U i ，
R j Is the observed noise variance, H, of node j i Is a node iThe information to be interacted of the node i contains a priori estimated value of k time
And information vector u i (k) And a matrix U i In a standardized form of
And will message i (k) Broadcast to neighboring nodes, i.e. j ∈ N i Simultaneously receiving broadcast information of all adjacent nodes, and finally fusing the node i and the adjacent node j belonging to the N i Information acquisition of (2) fused perception data
Fusion variance
The invention is further improved in that the specific operation of the step 6) is as follows: at the sampling node, the prior information is used
And neighbor node perception information, i.e. fusion perception data y i (k) And fusion variance S i (k) The node i can obtain the posterior estimation with accurate k time by correction
Firstly, a series of parameters of the current node i at the moment k are calculated: kalman gain M i Uniformity gain gamma and estimation error variance P i (ii) a Kalman gain M i (k)＝(P i -1 +S i (k)) -1 Gain of consistency
Where β > 0 is a relatively small constant, the error variance P is estimated i ＝AM i (k-1)A T + Q, where A is the transfer coefficient, Q is the variance of the noise in the user equation of state, and detectionThe change of the target is related and is obtained from historical data; computing a posteriori estimate using a series of parameters and fusion information
Namely:
on non-sampling nodes, prior estimates are issued directly
Namely that
Wherein A is the transfer coefficient of the first and second groups,
the posterior estimation value is calculated by the Kalman consistency estimator at the last moment of the ith node, and meanwhile, other nodes synchronously calculate local information and posterior estimation, so that all nodes can realize accurate distributed estimation.
The invention is further improved in that the specific operation of the step 7) is as follows: due to the lack of priori knowledge of time sequence data in the real-time release of dynamic data, the dynamic change of the data is detected to adjust the sampling frequency in real time, the change condition of the data is measured according to the deviation of the priori estimation and the posterior estimation by introducing a filtering strategy, and then the nth sampling time k is defined n (0＜k n Feedback error of < T)
Is composed of
Wherein
Is k n Is estimated a priori by a priori using a priori estimates of,
is k is n The parameter δ is set to prevent the situation where the divisor is 0, and is usually 1 in the statistical convergence scenario, assuming a posteriori estimate
Approaching the actual data of the perception target and estimating a priori
Is determined by a fixed process model, so that it is inferred if the feedback error is
Increasing means that the actual target data is undergoing rapid change, the error is fed back to the controller, the controller detects the error and correspondingly reduces the sampling interval, the dynamic adjustment of the sampling frequency can be realized, the most common feedback controller, namely PID controller, is adopted to measure the sampling performance, and the feedback error is based on
The output error Δ of the PID controller is:
wherein, C p 、C i 、C d Proportional, integral, differential control gain, T, respectively i For the integration time of the i-node, based on the analysis of the adaptive sampling strategy described above, k n Sampling interval of time of day
The update formula of (2) is:
wherein theta and xi are predefined parameters, theta determines the variation amplitude of the sampling interval, xi is the adjusting point of the sampling process, sp i Is the latest sampling moment of the node i, which needs to be based on the sampling interval
Are updated, i.e.
But if the sampling time n exceeds the maximum sampling time M, the system stops sampling, and the latest sampling time sp i No update is being performed.
The invention has at least the following beneficial technical effects:
according to the method for distributing the distributed single-dimensional time sequence data with consistency in real-time privacy protection and release, the sampling frequency is adjusted through a self-adaptive sampling strategy, and the sampled data is disturbed through a Laplace mechanism at the sampling moment. And then distributed posterior estimation is carried out on the published data of each node at the sampling moment through a single-dimensional perception information correction strategy based on Kalman consistency filtering, prior prediction data is published directly at the non-sampling moment, and finally, a distributed single-dimensional time sequence perception data real-time publishing method which does not depend on a central server and has differential privacy guarantee and consistency is realized, the real-time requirement of dynamic data publishing and the consistency requirement of distributed data publishing are met, and the real-time property, the high utility property and the consistency of data publishing are realized. The distributed dynamic sensing data distribution system can be used in a distributed power utilization load monitoring system of a smart power grid, a disease monitoring application and other distributed dynamic sensing data distribution systems.
Drawings
Fig. 1 is a distributed dynamic awareness data publishing process.
FIG. 2 is a flow chart of the method of the present invention.
Fig. 3 (a) is a graph of the Average Relative Error (Average Relative Error) between the published statistical sequence and the actual statistical sequence of the real data set uemploy data according to the FAST method and the present invention, as a function of the maximum privacy budget epsilon.
Fig. 3 (b) is a graph of the Average consistency Error (Average Consensus Error) of the distribution sequence of the real data set unnamploy data and the Average value of the distribution data of all nodes according to the maximum privacy budget epsilon of the present invention and the FAST method.
Fig. 4 (a) is a graph of the Average Relative Error (Average Relative Error) between the actual statistical sequence and the actual statistical sequence published by the tru data set Flu data according to the FAST method of the present invention as a function of the maximum privacy budget epsilon.
Fig. 4 (b) is a graph showing the variation of the Average consistency Error (Average Consensus Error) of the real data set Flu data publishing sequence and the Average value of all the node publishing data according to the maximum privacy budget epsilon of the present invention and the FAST method.
Detailed Description
The invention is described in further detail below with reference to the figures and the specific embodiments.
Referring to fig. 1, in the process of distributing distributed dynamic sensing data, each server node aggregates the sensing data of the user individuals in its local range and interactively shares with other nodes, so that the distributed nodes can simultaneously and consistently distribute global data, thereby providing further analysis and application. However, interactive sharing among distributed server nodes may pose a certain threat to the privacy of data, and published aggregated data may be subjected to link attacks, so that the privacy of a single user is revealed therein. For example, in a distributed power load monitoring system of a smart grid, each distributed site aggregates real-time power consumption data of users in different areas, and performs cooperative interaction with other sites to obtain comprehensive, efficient and accurate aggregated data, so as to provide services such as intelligent whole-grid scheduling, real-time electricity price policies, accurate prediction of whole-grid loads and the like. However, once the interaction process between the distributed sites is intercepted or an untrusted site exists, the aggregated data is exposed, and then an attacker can further deduce the privacy of the user such as daily behavior and life habit through the power utilization load curve of the user. Similarly, in disease monitoring applications, multiple hospitals perform real-time statistics on the number of patients with various diseases, and perform information interactive sharing with other hospitals to obtain comprehensive statistical information for scientific research or social services (such as intelligent monitoring and prevention of infectious diseases or epidemic infection in various regions). However, it is very important to provide the necessary privacy protection to the aggregated data before distribution because the unprotected aggregated data is likely to reveal the individual diseased condition.
Referring to fig. 2, the distributed single-dimensional time series data real-time privacy protection publishing method with consistency provided by the invention includes the following steps:
1) Modeling data: suppose that in a distributed sensing system with n nodes, each sensing node performs real-time statistics and aggregation on the same sensing target information (for example, statistics on the number of infectious disease patients per day, or statistics on the total electricity consumption per minute of all users in a certain area), so as to form a distributed single-dimensional sensing data sequence X of the ith node i ＝{x i (1),…,x i (k),…,x i (T) }, k =1, \ 8230;, T. Wherein x is i (k) And T is the length of the time sequence, namely sensing data of the sensing node i at the moment k, namely the information of the node i on the convergence of all the user individuals in the convergence range. Suppose that the actual statistical sequence of the perception target information is R = { R (1), \8230;, R (k), \8230;, R (T) }, where R (k) is the actual statistical value of the perception target at the discrete time k time instant. Perception data x of node i i (k) The theoretical relationship with the perception target actual data r (k) can be expressed as
x i (k)＝H i ·r(k)
Wherein H i Is the perceptual coefficient of the node i. The privacy protection issuing method aims at issuing the sensing target statistical data which is as accurate as possible in real time by the distributed nodes in the dynamic convergence process and protecting the privacy of individual users.
Because the perception data of a single node is often incomplete and accurate, the published data of distributed nodes is usually inconsistent. Therefore, in the distributed sensing system without the central convergence server, in order to realize comprehensive and accurate information dynamic release of each nodeTherefore, each node needs to perform mutual communication and cooperative operation with other nodes in the process of performing privacy protection processing and publishing. Each node only needs to establish connection with some adjacent nodes according to communication conditions to carry out local interaction, and does not need to carry out pairwise interaction with all other nodes in the sensing system. And modeling the interaction condition of the sensing node network by adopting an undirected graph G = (V, E), wherein V = {1,2, \8230;, n } is a node set, and E is an edge set. Every two nodes which can communicate with each other are connected by edges, namely, the nodes are neighbor nodes. In order to reduce network communication load, it is assumed that a sensing node in the distributed system performs information interaction only with a neighbor node at each time, and the model graph G is a connected graph. Since the sensing data of each node may be attacked in the interaction process, and the privacy security between the nodes does not have transparency, before performing information interaction with other nodes, each node needs to perform privacy protection processing on the sensing data at the local end. By z i (k) Representing perception data x for node i i (k) And (5) disturbing data after privacy protection. Finally, the distributed distribution sequence of each node is denoted as O = { O i (1),…,o i (k),…,o i (T)},i＝1,…,n。
2) Initialization: presetting the maximum sampling times M and the maximum privacy budget epsilon according to actual conditions and requirements, and allocating the well-distributed privacy budget in advance
3) Perceptual data sampling: if the current time k is equal to the latest sampling time sp of the node i i And the current sampling number n is less than the maximum sampling number M, then the input time series data is sampled at the current time k, the current time is a sampling point, and the sampling number n = n +1. Otherwise, the input time sequence data is not sampled at the current moment k, and the current moment is a non-sampling point.
4) Perception data disturbance: sensing data x of ith node needing sampling at k moment i (k) Addition is based on sensitivity Δ f And privacy budget ε k Calibrated Laplace noise, resulting in a disturbanceData z i (k) I.e. z i (k)＝x i (k)+Lap(0,Δ f /ε k ) Wherein the sensitivity Δ f It is usually calculated according to a query function in practical applications, for example, it is usually 1 in counting statistics. And directly taking the released data at the previous moment as the disturbance observation data at the current moment by the node without sampling.
5) Perception information interaction: the local of all the nodes calculates the prior estimated value of the ith node through a process model at the moment k by a Kalman consistency filtering-local information updating algorithm
Wherein A is the transfer coefficient of the first and second groups,
the posterior estimated value is calculated by a Kalman consistency estimator at the last moment of the ith node.
Then, the information vector u of the ith node needs to be calculated i (k) And matrix U i ，
R j Is the observed noise variance, H, of node j i Is the perceptual coefficient of the node i. The information to be interacted of the node i comprises a priori estimated value of k time
Information vector u i (k) And matrix U i In a standardized form of
And will message i (k) Broadcast to neighboring nodes, i.e. j ∈ N i And simultaneously receiving broadcast information of all adjacent nodes. Finally, the fusion node i and the adjacent node j thereof belong to N i Information acquisition of (2) fused perception data
Fusion variance
6) And (3) correcting the perception information: at the sampling node, the prior information is passed
And neighbor node perception information, i.e. fusion perception data y i (k) And fusion variance S i (k) The node i can obtain the posterior estimation with accurate k time after correction
Firstly, a series of parameters of the current node i at the moment k are calculated: kalman gain M i Uniformity gain gamma and estimation error variance P i . Kalman gain M i (k)＝(P i -1 +S i (k)) -1 Gain of consistency
Where β > 0 is a relatively small constant, the error variance P is estimated i ＝AM i (k-1)A T + Q, where A is the transfer coefficient and Q is the variance of the noise in the user equation of state, associated with the change in the detection target itself, may be obtained from historical data. Computing a posteriori estimates using a series of parameters and fusion information
Namely:
on non-sampling nodes, prior estimates are issued directly
Namely, it is
Wherein A is the transfer coefficient of the first layer,
the posterior estimated value is calculated by a Kalman consistency estimator at the last moment of the ith node. Meanwhile, other nodes are synchronously calculating local information and posterior estimation, so that all nodes can realize accurate distributed estimation.
7) Perceptual update feedback: due to the lack of prior knowledge of the timing data in the real-time distribution of dynamic data, dynamic changes in the data need to be detected to adjust the sampling frequency in real-time. The introduced filtering strategy measures the change condition of data according to the deviation of the prior estimation and the posterior estimation, and defines the nth sampling time k n (0＜k n Feedback error of < T)
Is composed of
Wherein
Is k n Is estimated a priori of the time-of-flight,
is k n The parameter δ is set to prevent the case where the divisor is 0, and in the case of statistical convergence, it is usually 1. Hypothesis posterior estimation
Approaching the actual data of the sensing target and estimating a priori
Is determined by a fixed process model, so it can be inferred if the feedback error is correct
Increasing means that the actual target data is passingUndergoing rapid changes. At this time, the error is fed back to the controller, and the controller detects the error and correspondingly reduces the sampling interval, so that the dynamic adjustment of the sampling frequency can be realized.
The most common feedback controller, the PID (proportional, integral, and derivative) controller, is used to measure the performance of the sample. Based on feedback error
Has an output error Delta of
Wherein, C p 、C i 、C d Proportional, integral, differential control gain, T, respectively i Is the integration time. Analysis based on the adaptive sampling strategy described above, k n Sampling interval of time of day
Is updated by the formula
And theta determines the change amplitude of a sampling interval, and ξ is an adjusting point of the sampling process. sp i Is the latest sampling moment of the node i, which needs to be based on the sampling interval
Is updated, i.e.
But if the sampling time n exceeds the maximum sampling time M, the system stops sampling, and the latest sampling time sp i No update is being performed.
Referring to fig. 3 (a) to (b) and fig. 4 (a) to (b), the superiority of the present invention in data protection is analyzed as follows:
FIGS. 3 (a) - (b) are graphs comparing the effect of the method of the present invention (PropDP) and the FAST method on the change of the Unamploy data in the real data set with the maximum privacy budget ε; wherein, fig. 3 (a) is a graph of the Average Relative Error (Average Relative Error) between the published statistical sequence and the actual statistical sequence of the FAST method according to the present invention as a function of the maximum privacy budget epsilon. Fig. 3 (b) is a graph of the Average Consensus Error (Average Consensus Error) between the release sequence of the present invention and FAST method and the Average of all node release data as a function of the maximum privacy budget epsilon.
In fig. 3 (a), as the privacy budget increases, i.e., the privacy level decreases, the average relative error of the two mechanisms over different data sets exhibits different degrees of decrease. The FAST mechanism also adopts a self-adaptive sampling strategy and a filtering idea, so that the real-time performance and high utility of privacy protection issued data are ensured. However, the FAST employs a general kalman filter, which is only suitable for independent real-time publishing of a single data source, and can not achieve consistent publishing of each node in the distributed sensing system. This chapter uses it in a distributed sensing system, with each node independently performing the method. The average relative error is generally used for measuring the following condition of the output sequence to the input sequence, and is a generalized index which does not need to consider the actual specific application field, and a smaller average relative error means that the published sequence is closer to the actual sequence, that is, the utility of the published data is better. Conversely, the larger the average relative error, the less effective the utility. As the privacy budget increases, the less error the laplace perturbation introduces. And the average relative error of the PropDP mechanism is still much lower than that of FAST, so that the PropDP mechanism has higher utility than that of FAST mechanism under the condition of ensuring the same privacy level. And the utility of the promdp mechanism on the unneploy data set is still greatly improved when the privacy budget is large. The reason is that the Unamploy data set and the numerical value are generally small, and when privacy budgets are close to 1, errors caused by Laplace disturbance are large, so that the PropDP method has more advantages and the effectiveness is improved more remarkably. As such, the average relative error of the PropDP mechanism is much lower across all data sets than that of the FAST mechanism, as the perturbation error is large, when the privacy budget is close to 0.1.
In fig. 3 (b), as the privacy budget increases, i.e., the privacy level decreases, the average consistency error of both mechanisms over different data sets exhibits different degrees of decrease. The Average consistency Error (Average Consensus Error) is used to measure the consistency of the distributed node privacy protection release data. The average consistency error is defined as follows
I.e. distributed release sequence o i (k) And the average value o average (k) The deviation therebetween. o average (k) Representing the mean value of data released by all nodes, i.e.
The smaller the average consistency error is, the better the consistency between the issuing sequences of the sensing nodes is, and the worse the opposite is.
Because the inconsistency of data published by each node is mainly caused by Laplace perturbation, the larger the privacy budget is, the smaller the perturbation noise is, and the better the consistency among the nodes is. The average coherence error of the PropDP mechanism is much lower than that of the FAST mechanism, both at lower and higher privacy budgets. It can be seen that the promdp mechanism publishes data much more consistently than the FAST mechanism, while ensuring the same privacy level.
FIGS. 4 (a) - (b) are graphs comparing the effect of the method of the present invention (PropDP) and the FAST method on the variation of the data of the real data set FLU data with the maximum privacy budget ε; fig. 4 (a) is a graph of the Average Relative Error (Average Relative Error) between the published statistical sequence and the actual statistical sequence of the FAST method according to the present invention as a function of the maximum privacy budget epsilon. Fig. 4 (b) is a graph of the Average consistency Error (Average Consensus Error) of the release sequence of the FAST method and the Average of all node release data according to the present invention as a function of the maximum privacy budget epsilon.
In fig. 4 (a), as the privacy budget increases, i.e. the privacy level decreases, the average relative error of the PropDP and FAST on different data sets shows different decreases, and the average relative error curve of the PropDP mechanism is closer to that of the FAST mechanism, but the average relative error of the PropDP mechanism is still smaller than that of the FAST mechanism in general. It can be seen that the PropDP mechanism has a higher utility than the FAST mechanism, ensuring the same privacy level.
In fig. 4 (b), as the privacy budget increases, i.e., the privacy level decreases, the average consistency error of both mechanisms over different data sets exhibits different degrees of decrease. The average coherence error of the PropDP mechanism is much lower than that of the FAST mechanism, both at lower and higher privacy budgets. It can be seen that the PropDP mechanism has greater consistency in publishing data than the FAST mechanism, while ensuring the same privacy level.
Claims (2)
1. A distributed single-dimensional time sequence data real-time privacy protection publishing method with consistency is characterized in that sampling frequency is adjusted through a self-adaptive sampling strategy, sampled data are disturbed through a Laplace mechanism at the sampling time, then distributed posterior estimation is carried out on published data of nodes at the sampling time through a single-dimensional perception information correction strategy based on Kalman consistency filtering, prior prediction data are published directly at the non-sampling time, and finally the distributed single-dimensional time sequence perception data real-time publishing method which is independent of a central server and has differential privacy guarantee and consistency is used in a distributed power consumption load monitoring system of an intelligent power grid or a distributed dynamic perception data publishing system applied to disease monitoring specifically comprises the following steps:
1) Modeling data: in a statistical convergence issuing scene, defining a real-time privacy protection issuing problem of distributed single-dimensional time sequence data, and modeling a sensing node network interaction condition; the specific operation is as follows: in a distributed sensing system with n nodes, each sensing node carries out real-time statistical aggregation on the same sensing target information to form single-dimensional sensing of the distributed ith nodeData sequence X i ＝{x i (1),…,x i (k),…,x i (T) }, k =1, \8230, T, where x i (k) Sensing data of a sensing node i at a moment k, namely the convergence information of the node i to all user individuals in a convergence range, T is the length of a time sequence, the actual statistical sequence of sensing target information is R = { R (1), \8230 =, R (k), \8230, R (T) }, wherein R (k) is the actual statistical value of the sensing target at a moment discrete time k, and the sensing data x of the node i at the moment k is the sensing data x of the sensing target at the moment discrete time k i (k) The theoretical relationship with the perception target actual data r (k) can be expressed as
x i (k)＝H i ·r(k)
Wherein H i In order to realize the comprehensive, accurate and consistent dynamic information release of each node in a distributed sensing system without a central convergence server, each node performs mutual communication and cooperative operation with other nodes in the privacy protection processing and releasing process, and each node establishes connection with certain adjacent nodes according to communication conditions to perform local interaction, the method is characterized in that pairwise interaction is not carried out with all other nodes in a sensing system, an undirected graph G = (V, E) is adopted to model the interaction situation of the sensing node network, wherein V = {1,2, \8230, n } is a node set, E is an edge set, every two nodes capable of communicating with each other are connected by edges, namely neighbor nodes, in order to reduce network communication load, the sensing nodes in a distributed system only carry out information interaction with the neighbor nodes at each moment, and the model graph G is a connected graph i (k) Representing perception data x for node i i (k) Number of disturbances after privacy protectionAccordingly; finally, the distributed distribution sequence of each node is denoted as O = { O i (1),…,o i (k),…,o i (T)},i＝1,…,n；
2) Initialization: presetting the maximum sampling times M and the maximum privacy budget epsilon of all nodes according to actual conditions and requirements, and allocating the privacy budget for each sampling time point k in advance according to an average allocation principle
3) And (3) perception data sampling: each node determines whether the current moment k samples the input time sequence data or not through the latest sampling interval; the specific operation is as follows: if the current instant k is equal to the latest sampling instant sp of node i i If the current sampling number n is smaller than the maximum sampling number M, sampling the input time sequence data at the current moment k, wherein the current moment is a sampling point, and the sampling number n = n +1, otherwise, not sampling the input time sequence data at the current moment k, and wherein the current moment is a non-sampling point;
4) And (3) perception data disturbance: the sampled node i truly senses data x of the node i through a Laplace algorithm LPA i (k) Disturbing to obtain a disturbance value z of the node satisfying the differential privacy guarantee i (k) Directly taking the released data at the previous moment as disturbance observation data at the current moment by the non-sampling node; the specific operation is as follows: sensing data x of ith node needing sampling at k moment i (k) Addition based on sensitivity Δ f And privacy budget ε k Calibrated Laplace noise to obtain disturbance data z i (k) I.e. z i (k)＝x i (k)+Lap(0,Δ f /ε k ) Wherein the sensitivity Δ f Usually, the calculation is carried out according to a query function in practical application, and a node which does not need to be sampled directly takes the released data at the previous moment as disturbance observation data at the current moment;
5) Perception information interaction: all nodes including non-sampling nodes locally calculate the prior estimated value of the node i through a Kalman consistency filtering-local information updating algorithm KCF-update
And broadcast information message i (k) Broadcasting to all neighbor nodes, and receiving broadcast information from the neighbor nodes at the same time, wherein the interactive information does not relate to real sampling data of each node; the specific operation is as follows: locally calculating the prior estimated value of the ith node through a process model at the moment k by using a Kalman consistency filtering-local information updating algorithm of all nodes
Wherein A is the transfer coefficient of the first layer,
calculating a posterior estimated value at the last moment of the ith node through a Kalman consistency estimator; then, the information vector u of the ith node is calculated i (k) And matrix U i ，
R j Is the observed noise variance, H, of node j i The information to be interacted of the node i contains a priori estimated value of k time as a perception coefficient of the node i
And an information vector u i (k) And a matrix U i In a standardized form of
And will message i (k) Broadcast to neighboring nodes, i.e. j ∈ N i Simultaneously receiving broadcast information of all adjacent nodes, and finally fusing the node i and the adjacent node j belonging to the node N i Information acquisition of (2) fused perception data
Fusion variance
6) And (3) correcting the perception information: the sampling node realizes the correction of local disturbance data by using prior information and sensing information of neighbor nodes, the sensing target at the current moment k is subjected to posterior estimation and issued by a Kalman consistency filtering-single-dimensional distributed estimation algorithm KCF-correction, and the non-sampling node directly issues the prior estimation at the moment and completes the iteration; the specific operation is as follows: at the sampling node, the prior information is passed
And neighbor node perception information, i.e. fusion perception data y i (k) And fusion variance S i (k) The node i can obtain the posterior estimation with accurate k time by correction
Firstly, a series of parameters of the current node i at the moment k are calculated: kalman gain M i Uniformity gain gamma and estimation error variance P i (ii) a Kalman gain
Gain in consistency
Where β > 0 is a relatively small constant, the error variance P is estimated i ＝AM i (k-1)A T + Q, where A is the transfer coefficient and Q is the variance of the noise in the user equation of state, associated with the change in the detection target itself, obtained from historical data; computing a posteriori estimate using a series of parameters and fusion information
Namely:
on non-sampling nodes, prior estimates are issued directly
Namely, it is
Wherein A is the transfer coefficient of the first layer,
the posterior estimation value is calculated by the Kalman consistency estimator at the last moment of the ith node, and meanwhile, other nodes synchronously calculate local information and posterior estimation, so that all nodes can realize accurate distributed estimation;
7) Perceptual update feedback: and the Sampling node feeds the prior posterior estimation error back to the PID controller, updates the Sampling interval and completes the iteration through Adaptive Sampling algorithm Adaptive Sampling, and if the Sampling times exceed the maximum Sampling times M, the system stops Sampling.
2. The method for real-time privacy-preserving publication of coherent distributed single-dimensional time-series data according to claim 1, wherein the specific operations in step 7) are: due to the lack of priori knowledge of time sequence data in the real-time release of dynamic data, the dynamic change of the data is detected to adjust the sampling frequency in real time, the change condition of the data is measured according to the deviation of the priori estimation and the posterior estimation by introducing a filtering strategy, and then the nth sampling time k is defined n Feedback error of
Is composed of
Wherein
Is k n A priori estimate of 0<k n <T，
Is k n The parameter delta is set to prevent the situation that the divisor is 0, 1 is usually taken in the statistic convergence scene, and the posterior estimation
Approaching the actual data of the perception target and estimating a priori
Is determined by a fixed process model, so that it is inferred if the feedback error is correct
Increasing means that the actual target data is undergoing rapid change, the error is fed back to the controller, the controller detects the error and correspondingly reduces the sampling interval, the dynamic adjustment of the sampling frequency can be realized, the most common feedback controller, namely PID controller, is adopted to measure the sampling performance, and the feedback error is based on
The output error Δ of the PID controller is:
wherein, C p 、C i 、C d Proportional, integral, differential control gain, T, respectively i For the integration time of the i-node, based on the analysis of the adaptive sampling strategy described above, k n Sampling interval of time of day
The update formula of (2) is:
wherein theta and xi are predefined parameters, theta determines the variation amplitude of the sampling interval, xi is the adjusting point of the sampling process, sp i Is the latest sampling moment of the node i, which needs to be based on the sampling interval
Are updated, i.e.
But if the sampling time n exceeds the maximum sampling time M, the system stops sampling, and the latest sampling time sp i No further updates are made.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011097748.XA CN112231749B (en) | 2020-10-14 | 2020-10-14 | Distributed single-dimensional time sequence data real-time privacy protection publishing method with consistency |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011097748.XA CN112231749B (en) | 2020-10-14 | 2020-10-14 | Distributed single-dimensional time sequence data real-time privacy protection publishing method with consistency |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112231749A CN112231749A (en) | 2021-01-15 |
| CN112231749B true CN112231749B (en) | 2022-12-09 |
Family
ID=74112734
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011097748.XA Active CN112231749B (en) | 2020-10-14 | 2020-10-14 | Distributed single-dimensional time sequence data real-time privacy protection publishing method with consistency |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112231749B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113473361B (en) * | 2021-06-28 | 2022-04-15 | 武汉大学 | Check-in data dynamic collection and release method and system based on B-DP differential privacy mechanism |
| CN115661617B (en) * | 2022-12-28 | 2023-03-14 | 成都中轨轨道设备有限公司 | Dynamic self-adaptive distributed cooperative working method for remote sensing big data |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105631360A (en) * | 2016-01-06 | 2016-06-01 | 西安交通大学 | Private data aggregating method based on multidimensional decomposition in sensor network |
| CN109450889A (en) * | 2018-11-02 | 2019-03-08 | 西安交通大学 | The secret protection dissemination method of data flow is converged in a kind of Internet of Things |
| CN110472437A (en) * | 2019-07-29 | 2019-11-19 | 上海电力大学 | A kind of period susceptibility difference method for secret protection of user oriented electricity consumption data |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US7853545B2 (en) * | 2007-02-26 | 2010-12-14 | International Business Machines Corporation | Preserving privacy of one-dimensional data streams using dynamic correlations |
-
2020
- 2020-10-14 CN CN202011097748.XA patent/CN112231749B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105631360A (en) * | 2016-01-06 | 2016-06-01 | 西安交通大学 | Private data aggregating method based on multidimensional decomposition in sensor network |
| CN109450889A (en) * | 2018-11-02 | 2019-03-08 | 西安交通大学 | The secret protection dissemination method of data flow is converged in a kind of Internet of Things |
| CN110472437A (en) * | 2019-07-29 | 2019-11-19 | 上海电力大学 | A kind of period susceptibility difference method for secret protection of user oriented electricity consumption data |
Non-Patent Citations (5)
| Title |
|---|
| Differentially Private Two-Dimension Sparse Data Publication under Consistency;Tingting Chen等;《2013 International Conference on Cloud Computing and Big Data》;20140526;全文 * |
| temporal Monitoring of IoT;Teng Wang等;《2019 IEEE 38th International Performance Computing and Communications Conference (IPCCC)》;20200116;全文 * |
| 可穿戴设备的数值型流数据差分隐私均值发布;涂子璇等;《计算机应用》;20200430(第06期);全文 * |
| 面向动态数据发布的差分隐私保护研究综述;屈晶晶等;《北京信息科技大学学报(自然科学版)》;20191231(第06期);全文 * |
| 面向时序数据发布的隐私保护方法研究;于东等;《通信学报》;20151130;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112231749A (en) | 2021-01-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112231749B (en) | Distributed single-dimensional time sequence data real-time privacy protection publishing method with consistency | |
| Li et al. | Approximate holistic aggregation in wireless sensor networks | |
| Yang et al. | Survey on improving data utility in differentially private sequential data publishing | |
| LIU et al. | Data Publication Based On Differential Privacy In V2G Network | |
| Al-Hussaeni et al. | Privacy-preserving trajectory stream publishing | |
| Chen et al. | An approach of flow compensation incentive based on Q-learning strategy for IoT user privacy protection | |
| Pang et al. | Towards personalized privacy-preserving truth discovery over crowdsourced data streams | |
| Ren et al. | LDP-IDS: Local differential privacy for infinite data streams | |
| Otgonbayar et al. | K-VARP: K-anonymity for varied data streams via partitioning | |
| Wang et al. | Improved Kalman filter based differentially private streaming data release in cognitive computing | |
| CN112214788B (en) | Ubiquitous power Internet of things dynamic data publishing method based on differential privacy | |
| CN108234493A (en) | The space-time crowdsourcing statistical data dissemination method of secret protection under insincere server | |
| Yan et al. | Parallel alternating direction method of multipliers | |
| Wang et al. | Elastic-net regularized latent factor analysis-based models for recommender systems | |
| Fei et al. | Fog computing perception mechanism based on throughput rate constraint in intelligent Internet of Things | |
| CN111416735A (en) | Federal learning-based safety QoS prediction method under mobile edge environment | |
| CN111353173B (en) | Sensitive tag track data publishing method using graph difference privacy model | |
| Yan et al. | Dynamic release of big location data based on adaptive sampling and differential privacy | |
| Cui et al. | A novel spatial cloaking scheme using hierarchical hilbert curve for location-based services | |
| Fan et al. | Adaptively sharing time-series with differential privacy | |
| Abdelgawad | Distributed Kalman filter with fast consensus for wireless sensor networks | |
| Chen et al. | The bifurcation analysis and optimal feedback mechanism for an SIS epidemic model on networks | |
| Xiong et al. | Predict: Privacy and security enhancing dynamic information collection and monitoring | |
| Liu et al. | A dynamic privacy protection mechanism for spatiotemporal crowdsourcing | |
| CN114662152B (en) | Real-time data-oriented localization differential privacy data stream publishing method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |