CN107871087A - The personalized difference method for secret protection that high dimensional data is issued under distributed environment - Google Patents
The personalized difference method for secret protection that high dimensional data is issued under distributed environment Download PDFInfo
- Publication number
- CN107871087A CN107871087A CN201711092850.9A CN201711092850A CN107871087A CN 107871087 A CN107871087 A CN 107871087A CN 201711092850 A CN201711092850 A CN 201711092850A CN 107871087 A CN107871087 A CN 107871087A
- Authority
- CN
- China
- Prior art keywords
- attribute
- data
- bayesian network
- sensitive
- mutual 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.)
- Granted
Links
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
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/21—Design, administration or maintenance of databases
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Databases & Information Systems (AREA)
- Bioethics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Computer Security & Cryptography (AREA)
- Software Systems (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
- Storage Device Security (AREA)
Abstract
The present invention disclose a kind of personalized difference method for secret protection of the issue of high dimensional data under distributed environment, and by the correlation between mutual information quantified property, the mutual information of respective attributes pair is calculated using mutual information formula.Approximate k degree Bayesian network is built according to mutual information, Bayesian network can be good at embodying the dependence between attribute and attribute.According to the personalized distribution privacy budget of the quantity of Sensitive Attributes and the non-quick attribute for meeting condition.According to the privacy budget distributed, carry out plus make an uproar to data each participant processing, is carried out plus is made an uproar using random response mechanism.Adding the data after making an uproar to be sent to manager, integrated data collection, then externally issue are synthesized by manager's aggregation.The present invention ensure that privacy requirement when issuing data, and the data of processing are largely reduced, and so as to which the change to data reduces, improve the utility of data, is advantageous to data analysis person and carries out correlation analysis.
Description
Technical field
The present invention relates to network data security technical field, and in particular to high dimensional data is issued under a kind of distributed environment
Personalized difference method for secret protection.
Background technology
In recent years, as the fast development of internet and information technology, secret protection data publication are of considerable interest,
People for protect individual privacy, realize data sharing idea understand it is more and more deep.Many specific data owners
The initial data (such as user data of the medical data of certain hospital, certain social network sites etc.) for needing to be possessed oneself is issued out
Go, so that other mechanisms are researched and analysed or for other purposes.It may be included in the initial data to be issued substantial amounts of
Sensitive individual privacy information (such as wages, disease condition, individual deposit etc.), directly issue may cause privacy information
Leakage, therefore data owner needs initial data carrying out secret protection and then issued again.These secret protection technologies are substantially
3 classes can be divided into:1) distribution technology based on restrictive condition:According to the characteristic of initial data, selectively issue is local special
Property data.Guard method or model have:K- anonymities method for secret protection, l- diversity indexes, t- approximate models, m- consistency
Deng.2) distribution technology based on data distortion:Make private data distortion but keep some characteristics of legacy data.Guard method
Or model has:Randomized technique, block techniques, switching technology, condensation technique etc..3) distribution technology based on data encryption:It is
The method of sensitive data is hidden in data mining process using encryption technology.Data encryption technology now development very into
Ripe, guard method or model have:Multi-party computations (SMC), des encryption technology, rsa encryption technique etc..
Difference privacy (differential privacy) is that Dwork carries in the privacy leakage problem for staqtistical data base
A kind of new privacy definition gone out.It is the distribution technology based on data distortion, and the technology using addition noise makes data distortion
But keep some data or data attribute constant simultaneously.Difference secret protection technology can ensure, when the addition in data set or
Person deletes a data record and does not interfere with inquiry output result, therefore even if in the worst case, mesh known to attacker
All sensitive datas outside mark record, still can ensure that target record will not be compromised.The protection technique is hidden unlike tradition
Private protection technique is the same, and it provides very strong theoretical guarantee on the privacy concern of data publication, i.e., not to the back of the body of attacker
Scape knowledge does any ad hoc hypothesis.The technology is not only that the issue of private data provides higher safety guarantee, and in reality
It is widely used in trampling.But existing difference secret protection technology can not handle high dimensional data under distributed environment well
RELEASE PROBLEM, particularly when including substantial amounts of attribute field in the data of issue, existing technology is injected greatly into data
The noise information of amount, it is possible to cause the data of issue to lose due effectiveness.
The content of the invention
To be solved by this invention is the deficiency present in existing privacy dissemination method, there is provided high under a kind of distributed environment
The personalized difference method for secret protection of dimension data issue.
To solve the above problems, the present invention is achieved by the following technical solutions:
The personalized difference method for secret protection that high dimensional data is issued under distributed environment, including step are as follows:
The each participant of step 1. by local data carry out plus make an uproar processing after be sent to manager;
Step 2. manager collects the data that each participant is sent, and the attribute of all participants is integrated, and goes
Except the attribute repeated, property set X is formed;
Step 3. participant cooperates with manager carries out sensitivity mark to the attribute in property set X, by property set X
Attribute be divided into Sensitive Attributes and the class of non-sensitive attribute two;
The attribute pair that each attribute is formed with other all properties in addition to itself in step 4. computation attribute collection X
Mutual information;
Step 5. initializes Bayesian networkAnd set
An attribute x is arbitrarily chosen in step 6. dependence collection X0And its parent attribute is set to beAttribute to (x0,)
It is added in Bayesian network N, attribute x0It is added in set V;
Step 7. is chosen in the remaining attribute in property set X and chooses an attribute as current attribute, and judges whether same
When meet following two conditions:
1. current attribute belongs to property set X, but is not belonging to set V;
2. the number of all subset elements is min (k, | V |) in set V, wherein | V | the number of attribute in set V is represented,
K represents that the maximum of any attribute in the Bayesian network of setting enters angle value, min represent k and | V | in smaller;
If it is, go to step 8;Otherwise, current Bayesian network N is returned, and goes to step 10;
Step 8. selects the maximum attribute pair of mutual information from the attribute centering of all about current attribute, and the maximum
Another attribute of attribute centering is added in Bayesian network N;
Step 9. judge current Bayesian network N any attribute node enter angle value whether be respectively less than be equal to k:If so,
Current Bayesian network N is then returned, and goes to step 10;Otherwise, return to step 7, until all properties addition property set X
Into Bayesian network N;
The attribute pair of step 10. all Sensitive Attributes and non-sensitive attribute composition from the current Bayesian network N of return
In, the attribute pair that mutual information is more than given threshold θ is selected, and the attribute selected by statistics is to the quantity of the Sensitive Attributes included
N and non-sensitive attribute quantity m;
Step 11. calculates each respectively according to the quantity n of the Sensitive Attributes counted and the quantity m of non-sensitive attribute
The privacy budget ε ' of participant:
Wherein, ε represents given total privacy budget, and α represents the weights of given Sensitive Attributes, and n ' represents the participant
The number of middle Sensitive Attributes, m ' represent the number of non-sensitive attribute in the participant;
For step 12. when user is inquired about the local data of some participant, manager is first with the participant
Privacy budget ε ' to Query Result carry out difference privacy add make an uproar after be then forwarded to user.
In above-mentioned steps 4, attribute is to (xi, ∏j) mutual phase information MI (xi, ∏j) calculation formula is:
Wherein, Pr [xi, Πj] represent attribute to (xi, Πj) joint probability distribution, Pr [xi] represent attribute xiEdge
Probability distribution, Pr [Πj] represent attribute ∏jMarginal probability distribution;
In above-mentioned steps 4, it is also necessary to according to mutual information size to all properties to being ranked up.
In above-mentioned steps 4, by all properties to carrying out descending sort according to mutual information, i.e. the big attribute of mutual information is to coming
Before, after the small attribute of mutual information is to coming.
In above-mentioned steps 5, constructed k- degree Bayesian network is directed acyclic graph.
Consider that existing difference privacy methods use same privacy budget to all participants, the standard of being uniformly processed can not
What is avoided causes workload to maximize.Also privacy budget allocation when do not account for personalization, the budget of mean allocation privacy certainly will
Cause some local data sets protections weaker, and the problem of some local data sets protections are too strong.The present invention passes through
Correlation between mutual information quantified property, the mutual information of respective attributes pair is calculated using mutual information formula.According to mutual information structure
Approximate k- degree Bayesian network is built, Bayesian network can be good at embodying the dependence between attribute and attribute.According to
The personalized distribution privacy budget of quantity of Sensitive Attributes and the non-quick attribute for the condition that meets.Each participant is according to the individual character distributed
Change privacy budget, carry out plus make an uproar to data processing, is carried out plus is made an uproar using random response mechanism.When user is to the sheet of some participant
When ground data are inquired about, manager is carried out after difference privacy adds and make an uproar first with the privacy budget ε ' of the participant to Query Result
It is then forwarded to user.The present invention ensure that privacy requirement when issuing data, and the data of processing are largely reduced, so as to data
Change reduce, improve the utilities of data, be advantageous to data analysis person and carry out correlation analysis.
Brief description of the drawings
Fig. 1 is the flow chart of the preferred embodiment of the present invention.
Fig. 2 is the 1- degree Bayesian networks N of 8 attribute nodes in the preferred embodiment of the present invention.
Embodiment
For the object, technical solutions and advantages of the present invention are more clearly understood, below in conjunction with instantiation, and with reference to attached
Figure, the present invention is described in more detail.
The present invention considers that data attribute has juxtaposition situation in distributed environment, according to the different sensitivitys of attribute
Relevance between and is distributed to realize that privacy budget is personalized, to solve the problems of prior art.Base of the present invention
It is credible in half it is assumed that each local data base is carried out plus made an uproar so that it meets ε according to the privacy budget of distributioni-
differential privacy(εiFor the privacy budget of distribution), after be sent to data manager and polymerize, data aggregate
Meet total ε-differential privacy afterwards (whereinK participant), while preferably keep original
The utility of data, be advantageous to the purpose of data analysis person carries out data analysis.
The system model of the present invention includes three roles:What k participant (i.e. local data owner), one and half trusted
Data manager and user's (data analysis).Each participant PkIt is owned by a local data base Dk.Manager is half credible
, assist k participant to issue integrated data collection D, the purpose of data of issue are directed to different data analyses for user.
Assuming that total property set:X={ x1,x2,x3,x4,x5,x6,x7,x8, wherein xi(i=1,2 .., 8) represents different
Attribute.When some attribute x in the local data of participant being presentiWhen, represented with " 1 ", when not deposited in the local data of participant
In some attribute xiWhen, represented with " 0 ", Pk(k=1,2,3,4) represents participant, then 4 participants:P1:1,1,0,1,0,0,
0,0 }, P2:{ 1,1,0,0,0,1,0,0 }, P3:{ 1,0,1,0,1,0,1,1 }, P4:{ 0,0,1,0,0,1,1,0 }.
The personalized difference method for secret protection that high dimensional data is issued under a kind of distributed environment, as shown in figure 1, combining tool
The present invention will be described in further detail for the example of body.It is as follows that it includes step:
Step 1:Each participant is sent to manager after local data to be added to processing of making an uproar.
Because manager is half believable, it is impossible to real result of calculation directly is sent into it, then will be to local
Data carry out adding processing of making an uproar to be then sent to manager.Here plus make an uproar processing be with one kind generation random number (scope is typically 0
To between 1) mode the marginal probability distribution of attribute is handled.Pr [x are used in examplei] represent plus make an uproar the attribute after handling
xiMarginal probability distribution.
Step 2:Manager collects the data that each participant is sent, and the attribute progress to all participants of collection is whole
Close, the attribute of repetition is removed during integration, to form property set X.
Step 3:Participant cooperates with manager and carries out sensitivity mark to the property set X for integrating generation, category
Property is labeled as Sensitive Attributes and non-sensitive attribute.
Property set X={ x in example1,x2,x3,x4,x5,x6,x7,x8, attribute is divided into Sensitive Attributes and non-sensitive category
Property, attribute { x in example4,x7Sensitive Attributes are defined as, remaining attribute is non-sensitive attribute.So any attribute is to (xi,
Πj) (x hereiRepresent any attribute in property set X, ΠjRepresent xiAny parent attribute, except attribute xiIt is any outside itself
Attribute is all probably its parent attribute) composition can be divided into three kinds of situations, i.e., Sensitive Attributes and Sensitive Attributes, Sensitive Attributes with it is non-quick
The attribute pair that sense attribute, non-sensitive attribute are formed with non-sensitive attribute.
Step 4:The joint probability point of each attribute and other all properties in addition to itself in computation attribute collection X
Cloth, it is expressed as Pr [xi, Πj], and then the association relationship MI of all properties pair after integration is calculated, and by all properties to being stored in
Sorted in array candidate and in a manner of descending.
The mutual information of attribute pair can be obtained by mutual information calculation formula, wherein an attribute is to (xi, Πi) mutual information
MI(xi, Πj) calculation formula be:
Wherein, xiFor any attribute in property set X, ΠjFor attribute xiParent attribute collection Π in any parent attribute, Pr
[xi, Πj] for attribute to (xi, Πj) joint probability distribution, Pr [xi] it is to add the attribute x after processing of making an uproariMarginal probability point
Cloth, Pr [Πj] it is to add the parent attribute Π after processing of making an uproariMarginal probability distribution.Wherein xi∈ X, Πj∈ X, i ≠ j.
Step 5:Initialize Bayesian networkAnd set
The empty list of Bayesian network N to one AP pairs of initialization, wherein AP is to referring to any attribute to (xi, ∏j),
xi∈ X, ∏j∈ ∏, i ≠ j, ∏ are attribute xiThe set of all parent attributes, initialization| ∏ |≤k's, i.e. some attribute
Parent attribute quantity is less than or equal to k, it is ensured that Bayesian network N is a k- degree Bayesian network.To ensure Bayesian network N simultaneously
For directed acyclic graph (DAG).
Set V deposits the set of the attribute of its known parent attribute, such as known attribute x when being structure N1Parent attribute for category
Property x2, then just attribute x1It is stored in V;
Step 6:An attribute x is randomly selected in dependence collection X (assuming that including d attribute altogether)0And its parent attribute is set
ForAttribute to (x0,) be added in Bayesian network N, x0It is added in set V.
For each selected properties xi∈ X V and eachX V represent attribute xiBelong to property set X but do not belong to
In set V, if for example, xaIt is added in V, then subsequently choose xiWhen will exclude xa。Represent all in V
(subset refers to that all properties are combined formed set in V to subset here, for example, it is assumed that V={ x1,x2, institute
It is { x to have subset1}、{x2}、{x1,x2) in attribute number be min (k, | V |).
Step 7:For the remaining attribute in property set X, d-1 iteration is carried out using greedy algorithm, until property set X
All properties be added in Bayesian network N;
In each iteration, attribute x is judgediWhether following two conditions are met simultaneously:
(1) attribute xiBelong to property set X and be not belonging to set V;
(2) number of all subset elements is min (k, | V |) in set V, wherein | V | represent the individual of attribute in set V
Number;The maximum of any attribute enters angle value in the Bayesian network that k represents to set;
If it is satisfied, then go in next step;Otherwise, current Bayesian network is returned, and goes to step 10;
In fact, if being unsatisfactory for above-mentioned 2 conditions, it is on stricti jurise that the Bayesian network of return, which can not can be regarded as,
Bayesian network, but in order to allow algorithm to be smoothed out, still exported as Bayesian network, in the case of this can pair
Last Query Result, which adds to make an uproar, produces larger error.
Step 8:According to array candidate in step 2 as attribute to Candidate Set.Regard the attribute in V as parent attribute,
Chosen from X V has the attribute of larger association relationship next in Bayesian network N as being added to these parent attributes
Attribute.By d-1 iteration, each iteration, to adding in N, finally returns to shellfish as far as possible the attribute for possessing larger mutual information
This network of leaf.Such as establish 1- degree Bayesian networks, then selection is exactly and attribute xaThere is the attribute of maximum mutual information value,
The attribute being added to as next time in Bayesian network N.It is identical if there is the mutual information of multiple attributes pair, preferentially select during selection
The attribute pair related to Sensitive Attributes is taken, is added in Bayesian network N.
Step 9:Whether the Bayesian network N that judgment step 8 is built meets k- degree, and meeting the standard of k- degree is:Current shellfish
Any attribute node of this network N of leaf enters whether angle value is respectively less than equal to k, if k- degree Bayesian networks are then returned to, in reality
The Bayesian network returned in example is as shown in Figure 2.Otherwise iterations adds 1, return to step 7.
The correlation of mutual information quantified property in information theory, using the GreedyBayes algorithms in PrivBayes
Build the Bayesian network of approximate k- degree (it is k that the maximum of any attribute node, which enters angle value, in the Bayesian network built).
Node represents attribute in Bayesian network, and directed edge or directed walk represent certain dependence between 2 attributes.Such as
In Fig. 2, attribute x1With attribute x2There is a directed edge, represent attribute x2Depend directly on attribute x1.Attribute x1With attribute x4Have one
Bar directed walk, represent attribute x4Indirectly depend on attribute x1.Bayesian network can be good at embodying between attribute and attribute
Dependence.
Step 10:Count in the Bayesian network returned and meet the Sensitive Attributes of condition and the quantity of non-sensitive attribute.
First, all attributes pair being made up of Sensitive Attributes and non-sensitive attribute are chosen from the Bayesian network of return.
In example, return is 1- degree Bayesian networks, and Sensitive Attributes are { x4,x7The attribute chosen is to for { x1,x4}、{x2,x4}、
{x3,x4}、{x5,x4}、{x6,x4}、{x8,x4}、{x1,x7}、{x2,x7}、{x3,x7}、{x5,x7}、{x6,x7}、{x8,x7}。
Then, the mutual information MI of these Sensitive Attributes and the attribute pair of non-sensitive attribute is compared with set threshold θ
Compared with selecting attributes pair of the mutual information MI more than those Sensitive Attributes and the non-sensitive attribute of threshold θ.Setting for threshold θ,
Depend on the circumstances, found by test of many times optimal.Such as in instances, threshold θ=0.2 is set, it is assumed that there are MI (x2,x4)>θ,
MI(x8,x7)>θ, therefore the non-sensitive attribute for needing exist for carrying out secret protection has 2, respectively x2、x8。
Finally, the selected quantity n of attribute centering Sensitive Attributes and the quantity m of non-sensitive attribute taken out of statistics.With it is set
Determine the size of threshold θ, if MI is more than θ, count the quantity of associated sensitive and non-sensitive attribute, it is otherwise insensitive to these
Handled with Sensitive Attributes.N=2, m=2 in example.
Step 11:Personalization distribution privacy budget.
Because existing difference privacy methods use same privacy budget to all participants, the standard of being uniformly processed can not
What is avoided causes workload to maximize.Also privacy budget allocation when do not account for personalization, the budget of mean allocation privacy certainly will
Cause some local data sets protections weaker, and some local data sets protections are too strong.So we use individual character
The privacy budget allocation scheme of change, the quantity n of the Sensitive Attributes counted according to step 10 and the quantity m of non-sensitive attribute,
The privacy budget ε ' of each participant is calculated respectively:
Wherein, ε represents given total privacy budget, and α represents the weights of given Sensitive Attributes, and n ' represents the participant
The number of middle Sensitive Attributes, m ' represent the number of non-sensitive attribute in the participant;
Pay attention to, the ratio i.e. α of the privacy budget distributed for Sensitive Attributes and non-sensitive attribute, to depend on the circumstances,
Found by test of many times optimal, but distribute total privacy budget and should be less than being equal to ε, so as to meet ε-difference privacy.
In instances, total privacy budget ε=0.1, α=0.6 are set.Now, ε1=α ε=0.6 ε, ε2=(1- α) ε=
0.4 ε, then the privacy budget for distributing to Sensitive Attributes are respectively ε (x4)=ε (x7ε/2 of)=0.6, distribute to the hidden of non-sensitive attribute
Private budget is respectively ε (x2)=ε (x8ε/2 of)=0.4.
Step 12:The process of making an uproar is added to the difference privacy of Query Result.When user is inquired about data, manager according to
The privacy budget ε ' distributed carries out adding processing of making an uproar to data.Added using random response (RR) mechanism for meeting ε '-difference privacy
Plus noise.
It is main to be applied to the inquiry of logarithm value type and tie because universal in existing difference privacy plus mode of making an uproar be Laplace mechanism
The protection of fruit, add used here as random response (RR) mechanism and make an uproar, be better than Laplace mechanism in some scenes.Random response machine
System is made up of following algorithm:S is that a random algorithm is used to generate a problem q.Res be using q and correct option tA as input,
Answer nA with noise is the random algorithm of output, is expressed as Res (q, tA), generally omits q, is expressed as Res (tA).Eval
(nA1..., nAn) it is with the answer with noise, nA1..., nAnAs input, export as valuation.So given ε >=0, with
Machine response mechanism M=(S, Res, Eval) meets ε-difference privacy, if for any two correct option tA0∈R、tA1∈ R,
For arbitrary element s ∈ R, Pr [Res (tA0)=s]≤eε×Pr[Res(tA1)=s].Wherein (S, Res, Eval) is represented respectively
Three random algorithms, ε are secret protection budget.In some scenarios by adding caused mean square error of making an uproar to be less than at random
Mean square error caused by Laplace mechanism, someone demonstrate this point.
Step 13:Terminate.
The present invention calculates the mutual of respective attributes pair by the correlation between mutual information quantified property using mutual information formula
Information.Approximate k- degree Bayesian network is built according to mutual information, Bayesian network can be good at embodying attribute and attribute it
Between dependence.According to the personalized distribution privacy budget of the quantity of Sensitive Attributes and the non-quick attribute for meeting condition.It is each to participate in
According to the privacy budget distributed, carry out plus make an uproar to data root processing, is carried out plus is made an uproar using random response mechanism.After adding and making an uproar
Data are sent to manager, and integrated data collection, then externally issue are synthesized by manager's aggregation.The present invention protects when issuing data
Privacy requirement is demonstrate,proved, the data of processing are largely reduced, and so as to which the change to data reduces, improve the utility of data, favorably
Correlation analysis is carried out in data analysis person.
It should be noted that although embodiment of the present invention is illustrative above, but it is to the present invention that this, which is not,
Limitation, therefore the invention is not limited in above-mentioned embodiment.Without departing from the principles of the present invention, it is every
The other embodiment that those skilled in the art obtain under the enlightenment of the present invention, it is accordingly to be regarded as within the protection of the present invention.
Claims (5)
1. the personalized difference method for secret protection that high dimensional data is issued under distributed environment, it is characterized in that, including step is as follows:
The each participant of step 1. by local data carry out plus make an uproar processing after be sent to manager;
Step 2. manager collects the data that each participant is sent, and the attribute of all participants is integrated, and removes weight
Multiple attribute, form property set X;
Step 3. participant cooperates with manager carries out sensitivity mark to the attribute in property set X, by the category in property set X
Property is divided into Sensitive Attributes and the class of non-sensitive attribute two;
The attribute pair that each attribute is formed with other all properties in addition to itself in step 4. computation attribute collection X it is mutual
Information;
Step 5. initializes Bayesian networkAnd set
An attribute x is arbitrarily chosen in step 6. dependence collection X0And its parent attribute is set to beAttribute pairAdd
It is added in Bayesian network N, attribute x0It is added in set V;
Step 7. is chosen in the remaining attribute in property set X and chooses an attribute as current attribute, and judges whether simultaneously full
The following two conditions of foot:
1. current attribute belongs to property set X, but is not belonging to set V;
2. the number of all subset elements is min (k, | V |) in set V, wherein | V | represent the number of attribute in set V, k tables
Show that the maximum of any attribute in the Bayesian network of setting enters angle value;
If it is, go to step 8;Otherwise, current Bayesian network N is returned, and goes to step 10;
Step 8. selects the maximum attribute pair of mutual information from the attribute centering of all about current attribute, and the maximum attribute
Another attribute of centering is added in Bayesian network N;
Step 9. judge current Bayesian network N any attribute node enter angle value whether be respectively less than be equal to k:If so, then return
Current Bayesian network N is returned, and goes to step 10;Otherwise, return to step 7, until property set X all properties are added to shellfish
In this network N of leaf;
The attribute centering of step 10. all Sensitive Attributes and non-sensitive attribute composition from the current Bayesian network N of return, choosing
Go out the attribute pair that mutual information is more than given threshold θ, and count selected attribute to the quantity n of Sensitive Attributes that is included and non-
The quantity m of Sensitive Attributes;
Step 11. calculates each participate in respectively according to the quantity n of the Sensitive Attributes counted and the quantity m of non-sensitive attribute
The privacy budget ε ' of side:
Wherein, ε represents given total privacy budget, and α represents the weights of given Sensitive Attributes, and n ' represents quick in the participant
Feel the number of attribute, m ' represents the number of non-sensitive attribute in the participant;
Step 12. when user is inquired about the local data of some participant, manager first with the participant privacy
Budget ε ' to Query Result carry out difference privacy add make an uproar after be then forwarded to user.
2. the personalized difference method for secret protection that high dimensional data is issued under distributed environment according to claim 1, it is special
Sign is, in step 4, attribute is to (xi, Πj) mutual phase information MI (xi, Πj) calculation formula is:
Wherein, Pr [xi, Πj] represent attribute to (xi, Πj) joint probability distribution, Pr [xi] represent attribute xiMarginal probability
Distribution, Pr [Πj] represent attribute ΠjMarginal probability distribution.
3. the personalized difference method for secret protection that high dimensional data is issued under distributed environment according to claim 1 or claim 2, its
It is characterized in, in step 4, it is also necessary to according to mutual information size to all properties to being ranked up.
4. the personalized difference method for secret protection that high dimensional data is issued under distributed environment according to claim 3, it is special
Sign is, in step 4, by all properties to according to mutual information carry out descending sort, i.e., the big attribute of mutual information to coming before, mutual trust
After ceasing small attribute to coming.
5. the personalized difference method for secret protection that high dimensional data is issued under distributed environment according to claim 1, it is special
Sign is, in step 5, constructed k- degree Bayesian network is directed acyclic graph.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711092850.9A CN107871087B (en) | 2017-11-08 | 2017-11-08 | Personalized differential privacy protection method for high-dimensional data release in distributed environment |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201711092850.9A CN107871087B (en) | 2017-11-08 | 2017-11-08 | Personalized differential privacy protection method for high-dimensional data release in distributed environment |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107871087A true CN107871087A (en) | 2018-04-03 |
CN107871087B CN107871087B (en) | 2020-10-30 |
Family
ID=61752616
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201711092850.9A Active CN107871087B (en) | 2017-11-08 | 2017-11-08 | Personalized differential privacy protection method for high-dimensional data release in distributed environment |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107871087B (en) |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108763954A (en) * | 2018-05-17 | 2018-11-06 | 西安电子科技大学 | Linear regression model (LRM) multidimensional difference of Gaussian method for secret protection, information safety system |
CN108776763A (en) * | 2018-06-08 | 2018-11-09 | 哈尔滨工程大学 | One kind being based on the relevant difference method for secret protection of attribute |
CN108959956A (en) * | 2018-06-07 | 2018-12-07 | 广西师范大学 | Difference private data dissemination method based on Bayesian network |
CN108959958A (en) * | 2018-06-14 | 2018-12-07 | 中国人民解放军战略支援部队航天工程大学 | A kind of method for secret protection and system being associated with big data |
CN109241770A (en) * | 2018-08-10 | 2019-01-18 | 深圳前海微众银行股份有限公司 | Information value calculating method, equipment and readable storage medium storing program for executing based on homomorphic cryptography |
CN109299436A (en) * | 2018-09-17 | 2019-02-01 | 北京邮电大学 | A kind of ordering of optimization preference method of data capture meeting local difference privacy |
CN110334539A (en) * | 2019-06-12 | 2019-10-15 | 北京邮电大学 | A kind of personalized method for secret protection and device based on random response |
CN110334546A (en) * | 2019-07-08 | 2019-10-15 | 辽宁工业大学 | Difference privacy high dimensional data based on principal component analysis optimization issues guard method |
CN111242194A (en) * | 2020-01-06 | 2020-06-05 | 广西师范大学 | Differential privacy protection method for affinity propagation clustering |
CN111259442A (en) * | 2020-01-15 | 2020-06-09 | 广西师范大学 | Differential privacy protection method for decision tree under MapReduce framework |
WO2020177484A1 (en) * | 2019-03-01 | 2020-09-10 | 华南理工大学 | Localized difference privacy urban sanitation data report and privacy calculation method |
CN112131604A (en) * | 2020-09-24 | 2020-12-25 | 合肥城市云数据中心股份有限公司 | High-dimensional privacy data publishing method based on Bayesian network attribute cluster analysis technology |
CN112395630A (en) * | 2020-11-26 | 2021-02-23 | 平安普惠企业管理有限公司 | Data encryption method and device based on information security, terminal equipment and medium |
CN112395638A (en) * | 2019-08-16 | 2021-02-23 | 国际商业机器公司 | Collaborative AI with respect to privacy-assured transactional data |
CN112528316A (en) * | 2020-09-18 | 2021-03-19 | 江苏方天电力技术有限公司 | Privacy protection lineage workflow publishing method based on Bayesian network |
CN112822004A (en) * | 2021-01-14 | 2021-05-18 | 山西财经大学 | Belief network-based targeted privacy protection data publishing method |
CN113379062A (en) * | 2020-03-10 | 2021-09-10 | 百度在线网络技术(北京)有限公司 | Method and apparatus for training a model |
WO2022199473A1 (en) * | 2021-03-25 | 2022-09-29 | 支付宝(杭州)信息技术有限公司 | Service analysis method and apparatus based on differential privacy |
CN115329898A (en) * | 2022-10-10 | 2022-11-11 | 国网浙江省电力有限公司杭州供电公司 | Distributed machine learning method and system based on differential privacy policy |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104050267A (en) * | 2014-06-23 | 2014-09-17 | 中国科学院软件研究所 | Individuality recommendation method and system protecting user privacy on basis of association rules |
CN104573560A (en) * | 2015-01-27 | 2015-04-29 | 上海交通大学 | Differential private data publishing method based on wavelet transformation |
CN105608388A (en) * | 2015-09-24 | 2016-05-25 | 武汉大学 | Differential privacy data publishing method and system based on dependency removal |
CN105608389A (en) * | 2015-10-22 | 2016-05-25 | 广西师范大学 | Differential privacy protection method of medical data dissemination |
US20170024575A1 (en) * | 2015-07-22 | 2017-01-26 | International Business Machines Corporation | Obfuscation and protection of data rights |
CN106778314A (en) * | 2017-03-01 | 2017-05-31 | 全球能源互联网研究院 | A kind of distributed difference method for secret protection based on k means |
-
2017
- 2017-11-08 CN CN201711092850.9A patent/CN107871087B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104050267A (en) * | 2014-06-23 | 2014-09-17 | 中国科学院软件研究所 | Individuality recommendation method and system protecting user privacy on basis of association rules |
CN104573560A (en) * | 2015-01-27 | 2015-04-29 | 上海交通大学 | Differential private data publishing method based on wavelet transformation |
US20170024575A1 (en) * | 2015-07-22 | 2017-01-26 | International Business Machines Corporation | Obfuscation and protection of data rights |
CN105608388A (en) * | 2015-09-24 | 2016-05-25 | 武汉大学 | Differential privacy data publishing method and system based on dependency removal |
CN105608389A (en) * | 2015-10-22 | 2016-05-25 | 广西师范大学 | Differential privacy protection method of medical data dissemination |
CN106778314A (en) * | 2017-03-01 | 2017-05-31 | 全球能源互联网研究院 | A kind of distributed difference method for secret protection based on k means |
Non-Patent Citations (3)
Title |
---|
LIN ZHANG等: "Efficient privacy-preserving classification construction model with differential privacy technology", 《JOURNAL OF SYSTEMS ENGINEERING AND ELECTRONICS》 * |
孙奎等: "一种增强的差分隐私数据发布算法", 《计算机工程》 * |
王玲玲等: "基于位置服务的隐私保护机制度量研究综述", 《计算机应用研究》 * |
Cited By (28)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108763954A (en) * | 2018-05-17 | 2018-11-06 | 西安电子科技大学 | Linear regression model (LRM) multidimensional difference of Gaussian method for secret protection, information safety system |
CN108763954B (en) * | 2018-05-17 | 2022-03-01 | 西安电子科技大学 | Linear regression model multidimensional Gaussian difference privacy protection method and information security system |
CN108959956A (en) * | 2018-06-07 | 2018-12-07 | 广西师范大学 | Difference private data dissemination method based on Bayesian network |
CN108959956B (en) * | 2018-06-07 | 2021-06-22 | 广西师范大学 | Differential privacy data publishing method based on Bayesian network |
CN108776763A (en) * | 2018-06-08 | 2018-11-09 | 哈尔滨工程大学 | One kind being based on the relevant difference method for secret protection of attribute |
CN108959958A (en) * | 2018-06-14 | 2018-12-07 | 中国人民解放军战略支援部队航天工程大学 | A kind of method for secret protection and system being associated with big data |
CN109241770A (en) * | 2018-08-10 | 2019-01-18 | 深圳前海微众银行股份有限公司 | Information value calculating method, equipment and readable storage medium storing program for executing based on homomorphic cryptography |
CN109299436A (en) * | 2018-09-17 | 2019-02-01 | 北京邮电大学 | A kind of ordering of optimization preference method of data capture meeting local difference privacy |
CN109299436B (en) * | 2018-09-17 | 2021-10-15 | 北京邮电大学 | Preference sorting data collection method meeting local differential privacy |
WO2020177484A1 (en) * | 2019-03-01 | 2020-09-10 | 华南理工大学 | Localized difference privacy urban sanitation data report and privacy calculation method |
CN110334539A (en) * | 2019-06-12 | 2019-10-15 | 北京邮电大学 | A kind of personalized method for secret protection and device based on random response |
CN110334539B (en) * | 2019-06-12 | 2021-06-22 | 北京邮电大学 | Personalized privacy protection method and device based on random response |
CN110334546A (en) * | 2019-07-08 | 2019-10-15 | 辽宁工业大学 | Difference privacy high dimensional data based on principal component analysis optimization issues guard method |
CN112395638B (en) * | 2019-08-16 | 2024-04-26 | 国际商业机器公司 | Collaborative AI with respect to transaction data with privacy guarantee |
CN112395638A (en) * | 2019-08-16 | 2021-02-23 | 国际商业机器公司 | Collaborative AI with respect to privacy-assured transactional data |
CN111242194B (en) * | 2020-01-06 | 2022-03-08 | 广西师范大学 | Differential privacy protection method for affinity propagation clustering |
CN111242194A (en) * | 2020-01-06 | 2020-06-05 | 广西师范大学 | Differential privacy protection method for affinity propagation clustering |
CN111259442A (en) * | 2020-01-15 | 2020-06-09 | 广西师范大学 | Differential privacy protection method for decision tree under MapReduce framework |
CN113379062A (en) * | 2020-03-10 | 2021-09-10 | 百度在线网络技术(北京)有限公司 | Method and apparatus for training a model |
CN113379062B (en) * | 2020-03-10 | 2023-07-14 | 百度在线网络技术(北京)有限公司 | Method and device for training model |
CN112528316A (en) * | 2020-09-18 | 2021-03-19 | 江苏方天电力技术有限公司 | Privacy protection lineage workflow publishing method based on Bayesian network |
CN112528316B (en) * | 2020-09-18 | 2022-07-15 | 江苏方天电力技术有限公司 | Privacy protection lineage workflow publishing method based on Bayesian network |
CN112131604A (en) * | 2020-09-24 | 2020-12-25 | 合肥城市云数据中心股份有限公司 | High-dimensional privacy data publishing method based on Bayesian network attribute cluster analysis technology |
CN112131604B (en) * | 2020-09-24 | 2023-12-15 | 合肥城市云数据中心股份有限公司 | High-dimensional privacy data release method based on Bayesian network attribute cluster analysis |
CN112395630A (en) * | 2020-11-26 | 2021-02-23 | 平安普惠企业管理有限公司 | Data encryption method and device based on information security, terminal equipment and medium |
CN112822004A (en) * | 2021-01-14 | 2021-05-18 | 山西财经大学 | Belief network-based targeted privacy protection data publishing method |
WO2022199473A1 (en) * | 2021-03-25 | 2022-09-29 | 支付宝(杭州)信息技术有限公司 | Service analysis method and apparatus based on differential privacy |
CN115329898A (en) * | 2022-10-10 | 2022-11-11 | 国网浙江省电力有限公司杭州供电公司 | Distributed machine learning method and system based on differential privacy policy |
Also Published As
Publication number | Publication date |
---|---|
CN107871087B (en) | 2020-10-30 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107871087A (en) | The personalized difference method for secret protection that high dimensional data is issued under distributed environment | |
US20170250796A1 (en) | Trans Vernam Cryptography: Round One | |
Zhang et al. | Identifying influential nodes in complex networks with community structure | |
Chen et al. | Differentially private transit data publication: a case study on the montreal transportation system | |
Gupta et al. | Centrality measures for networks with community structure | |
Task et al. | A guide to differential privacy theory in social network analysis | |
Navarro-Arribas et al. | Information fusion in data privacy: A survey | |
CN101916256A (en) | Community discovery method for synthesizing actor interests and network topology | |
Zeng et al. | Stackelberg game under asymmetric information in critical infrastructure system: From a complex network perspective | |
Ahmed et al. | A random matrix approach to differential privacy and structure preserved social network graph publishing | |
Rabelo et al. | Multigraph approach to quantum non-locality | |
CN107729767A (en) | Community network data-privacy guard method based on figure primitive | |
Kumar et al. | A novel architecture to identify locations for Real Estate Investment | |
CN110413652A (en) | A kind of big data privacy search method based on edge calculations | |
Wei et al. | Differential privacy-based trajectory community recommendation in social network | |
Le et al. | Full autonomy: A novel individualized anonymity model for privacy preserving | |
Lv et al. | Edge-fog-cloud secure storage with deep-learning-assisted digital twins | |
Miller | Equivalence of several generalized percolation models on networks | |
Kaleli et al. | SOM-based recommendations with privacy on multi-party vertically distributed data | |
Wang et al. | Differential privacy for weighted network based on probability model | |
CN115438227A (en) | Network data publishing method based on difference privacy and compactness centrality | |
Dhanalakshmi et al. | Privacy preserving data mining techniques-survey | |
da Silva et al. | Inference in distributed data clustering | |
Mazalov et al. | Game-theoretic centrality measures for weighted graphs | |
CN103200034B (en) | Network user structure disturbance method based on spectral constraint and sensitive area partition |
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 |