A double round privacy protection method based on differential privacy

Abstract

The application belongs to the technical field of privacy protection, and particularly relates to a double-wheel privacy protection method based on differential privacy. The specific process of the method is as follows: user grouping: in the first round of interaction, the user is grouped by judging whether the number of candidate values of the user is within the interval specified by the variable parameter, and judging whether the user sampling candidate value is a meaningless candidate value, part of the users do not need to participate in the first stage, and the meaningless privacy budget segmentation is reduced through the user grouping operation; first stage interaction: used for screening high-frequency items from the candidate value set as the frequent item candidate set of the second stage; second stage interaction: after filtering the answers not in the high-frequency candidate value set, the user end interacts with the data collector end by using the lightweight single-value frequency statistical method, estimates the candidate value frequency, and obtains the final user high-frequency answer result.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of privacy protection, and particularly relates to a double-round privacy protection method based on differential privacy. BACKGROUND

[0002] The problem of privacy protection was first proposed in the 1970s. With the advent of the big data era, personal data is collected in large quantities, which contains personal privacy information. How to protect this information has become a hot issue. Differential privacy was first defined in 2006. Its core is to add random noise to the data to mask the influence of individual data on the overall query result, so as to resist attackers from inferring personal information by using the output result. It provides mathematically provable privacy protection, strictly defines the attack model, and provides strong protection in theory. At the same time, the scenario applicable to differential privacy protection is very extensive, and it is one of the main methods of privacy protection at present.

[0003] In the scenario of facing multiple data sources and multiple local data, the high-dimensional problem is the high-dimensional candidate value set and the multiple local answer scenario. In order to ensure the usability of data and the accuracy of model estimation effect, the encoding result of user answer and the dimension of sending result will be expanded with the candidate value dimension and interference scheme and other algorithm design methods. High-dimensional user answer candidate set, high-dimensional user answer set, the dimension expansion of the data to be desensitized will inevitably lead to the reduction of data usability, the difficulty of model fitting at the data collector end and the improvement of algorithm complexity. At this time, not only the design of privacy protection scheme needs to be considered, but also the efficiency of algorithm in engineering scenario needs to be considered. Through the design of appropriate coding, desensitization and other schemes, the dimensionality reduction of user local data is realized, the lightness of algorithm and the usability of data are maintained.

[0004] Patent CN202211156864.3 proposes a big data frequency estimation method based on local differential privacy. First, the server samples the client, the client sends the data processed by the hash function and the differential privacy protection method to the server, the server trains a frequency prediction model using these data, and the server sends the trained model and the corresponding high-frequency data threshold to the client. The client judges whether it belongs to high-frequency data and responds.

[0005] Patent CN202211196477.2 proposes a big data domain set data collection method based on local differential privacy. First, the server groups the users and publishes the grouping results, the number of data contained in each group, and other information such as privacy budget. Users fill in or truncate the answers according to the received information, then perform flattening and interference operations on the data and send it back to the server. An initial heavy hitters set is constructed in the server, and the frequency estimation value in each group of user answer data is calculated according to the user grouping order. The value is iterated with the initial set to construct the next round of heavy hitters set. Until each group of user data is processed, the heavy hitters set obtained in the last round is the final set.

[0006] Patent CN202310212497.2 proposes a top-k network flow estimation method based on local differential privacy. First, users are randomly assigned or added to a certain group, and the corresponding network flow size and network flow base are counted and sent to the server. The server records these data, calculates the corresponding optimal privacy protection parameters for different groups of users, and sends them back to the corresponding users. The client's users store the network flow, use the optimal parameters provided by the server to perturb the data, and send the perturbed data back to the server. The server aggregates the perturbed data and finally iteratively estimates the top-k network flow.

[0007] The problems of the current privacy protection technology based on two-round interaction of multi-value scene mainly include three aspects: the first is the excessive segmentation of privacy budget, the second is the selection deviation of the first-round candidate value set, which reduces the statistical accuracy of the second round, and the third is the poor interference algorithm effect in the privacy protection process. In the LDPMiner algorithm, the operation of the first-round interaction is to sample and interfere with the user's answer, the server collects the sampling data to speculate the statistical candidate value set containing topk, and sends the set to the user end to reduce the communication cost of the second-round interaction. In the LDPminer algorithm, all users participate in the first-round interaction and divide the privacy budget for this purpose, but not all user answers are necessary, and the unnecessary answers of some users occupy the privacy budget. In addition, the selection deviation of the topk statistical candidate value set will affect the second-round interaction. The interference method used in the LDPminer algorithm is the SH and RAPPOR algorithms. The SH algorithm depends on the generation of the encoding random matrix to achieve strong statistical effect, and thus shows unstable statistical effect in the real experiment. The RAPPOR algorithm uses a Bloom filter to encode the data, but the Bloom filter is a probabilistic data structure and has a certain false positive rate. The encoding idea depends on the binary vector and a series of random mapping functions, and is mainly used to judge whether the queried element exists or not, rather than encoding, so that the interference algorithm effect is poor. SUMMARY

[0008] Therefore, the present application provides a two-round privacy protection method based on differential privacy, which can reduce meaningless privacy budget segmentation and accurately estimate the user's answer of the data producer by the data collector.

[0009] The technical scheme of the present application is as follows:

[0010] A two-round privacy protection method based on differential privacy includes the following contents:

[0011] User grouping: the local data is randomly sampled by the data producer, and the sampled user answer is processed according to the selected single-value frequency statistical method to realize user grouping, and the users participating in the first and second stages of privacy interaction and the users participating in only the second stage of privacy interaction are selected;

[0012] First-stage privacy interaction: the data producer extracts a single user answer value from the user answer set candidates i according to the sampling position, combines the total privacy budget ε, generates an interference result message report i , and sends it to the data collector; the data collector generates an interference result message report iAnd the total privacy budget, generate frequency estimations of each candidate value, filter the candidate value set according to the candidate set adjustable parameter λ2 and the candidate value frequency estimation, and select high-frequency items from the candidate value set candidates as the frequent item candidate set of the second stage Send to the data generator end;

[0013] The second stage of privacy interaction: the data generator is based on the frequent item candidate set A collection of candidates for the user's answers i Filter and fill, and generate user message reports according to user grouping and send them to the data collector; the data collector will collect user message reports, group and count them, and finally determine the frequent item sets answered by users. topk .

[0014] Furthermore, the specific process of user grouping described in the present invention is as follows:

[0015] (1) Set the threshold λ1 for the user to answer a round of interaction, and calculate the selection interval [λ1l, (1+λ1)l] based on the threshold λ1 and the sampling length l to filter the user's answer set candidates i The number of users in the interval user i Participate in the first phase of interaction and set the corresponding participation flag;

[0016] (2) The filtered user user i The answer set candidates i The set length (1+λ1)l-1) is truncated to fill the insufficient user answer set with virtual items; the truncated user answer data is randomly sampled sample_bit. If the sampled bit falls on the virtual item filling bit, the user is abandoned from participating in the first stage interaction and the participation flag flag is updated; finally, the retained sample bit sample_bit and participation flag flag are output.

[0017] Furthermore, during the second stage of privacy interaction of the present invention, the following operations are performed:

[0018] Filtering: The data generator filters the candidate value set of frequent itemsets based on the data collected in the first stage. Filter the user answer set to remove non-high-frequency items;

[0019] Filling: Set the length of the user's final answer set to the length of the high-frequency candidate value set, and fill the user answer sets that do not meet the above length;

[0020] Privacy budget: The data producer calculates the privacy budget available for each candidate value based on whether the user participates in the first-stage interaction;

[0021] Frequent item set generation: Based on the privacy budget and the user's answer set, interference generates a user message report and sends it to the data collector.

[0022] Furthermore, in the present invention, when a user participates in the interactive operation in the first phase, his privacy budget is When the user does not participate in the first phase of interactive operation, his privacy budget is Where ε is the total privacy budget, k is the number of final results of the two-round interaction, which is the k most frequent candidate values ​​among all candidate values, and λ2k is the number of candidate values ​​that the producer of the second-round interaction data wants to send to the collector.

[0023] Furthermore, during the second stage of privacy interaction of the present invention, the data collector sets a privacy budget based on whether the user participated in the first stage of privacy interaction. If the user participated in the first stage, the total privacy budget is set to ε / 2. If the user did not participate in the first stage, the total privacy budget is set to ε. Users are then divided into different groups based on their privacy budget.

[0024] In each group, the data collector calculates the candidate set of frequent itemsets A new round of frequency statistics is performed with the user report set reports. The statistical process will filter out the virtual filler items. The data collector will finally determine the final top k frequent item set estimations based on the frequency statistics of the two groups of users. topk .

[0025] Furthermore, in the present invention, when the numerical type and the candidate value set are: the numerical type uses the OSH algorithm to interfere with generating the result message; when the numerical type and the candidate value set are: the text type uses the NCHOSH algorithm to interfere with generating the result message.

[0026] Furthermore, the OSH algorithm of the present invention is:

[0027] First, the data collector generates a random projection matrix M of d×m, and ensures that the matrix is ​​a column-orthogonal matrix and the elements in the matrix are positive and negative. Send the matrix to the data producer;

[0028] Secondly, the data producer generates a random coding bit j from the interval (0, m-1), and uses the projection matrix M and the random coding bit j to encode the user response to generate a mapping vector d_vector representing the user input and a transmission bit i;

[0029] Again, with the probability of maintaining the true answer, and the probability of setting the answer to the opposite case, and epsilon is the privacy budget, and the value of the i-th dimension of the d_vector is used to generate the report i .

[0030] Further, when the OSH algorithm is used for interference generation, the process of processing at the data collector end is:

[0031] First, the data collector end initializes a statistical vector reports with all dimensions being 0, and according to each user's answer report i and a random encoding bit j, adds report i to the j-th dimension of the statistical vector, and generates a final statistical vector report;

[0032] Second, the data collector end encodes each candidate value in the candidate value set candidates according to the random projection matrix M, and each candidate value needs to be encoded and calculated with all encoding matrixes, and the sampling bits corresponding to each encoding result are extracted to form the decoding vector Me candidate of the candidate value;

[0033] Finally, the statistical vector reports and the decoding vector Me candidate corresponding to each candidate value are combined for inverse mapping, and the projection of the decoding vector to the statistical vector is the frequency of the candidate value corresponding to the candidate value.

[0034] Further, the NCHOSH algorithm of the present application is:

[0035] First, determine the privacy budget epsilon, set the mapping interval length d and the random seed length l, and generate parameters p and q according to the privacy budget epsilon and the mapping interval length d;

[0036] Second, generate a random seed res with a length of l, encode the user data using a hash algorithm, and reduce the dimension of the encoded result using a modulo operation;

[0037] Third, interfere with the encoded result, maintain it with a probability of 1-p+q, and randomly select a number from 0 to d-1 with a probability of p-q;

[0038] Finally, send the encoded result user message report and the random seed to the data receiver.

[0039] Further, when the NCHOSH algorithm is used for interference generation, the decoding process at the data collector end is:

[0040] First, for each user message report and random seed res, each candidate value in the overall candidate value set candidates is encoded using the random seed res, and the encoding result of each candidate value is encode candidate;

[0041] Second, after dimension reduction operation on each candidate value encoding result encode candidate, compare with the user message report, if same, add one to the dimension corresponding to the candidate value in the frequency statistics vector freq;

[0042] Third, after the above operation on all user answers, an biased statistics vector freq is obtained;

[0043] Finally, the biased statistics vector freq is calibrated using the frequency calibration formula to obtain the unbiased estimation result of the frequency corresponding to each candidate value.

[0044] Advantages:

[0045] First, the user grouping technology introduces an adjustable parameter λ1. In the first round of interaction, the user is grouped by judging whether the number of candidate values of the user is within the interval specified by the parameter, and whether the sampled candidate value of the user is a meaningless candidate value. The meaningless privacy budget is divided by the user grouping operation.

[0046] Second, the light-weight data de-sensitization technology for numerical type data in accordance with ε-differential privacy, based on the efficient privacy protection histogram estimation algorithm (Local, Private, Efficient Protocols Succinct Histograms Based on Encoding Optimization, OSH), which supports the transmission of The cost of maintaining the lightness of the overall de-sensitization technology. By improving the generation of the encoding matrix to convert the ordinary encoding matrix into a column orthogonal matrix, and introducing a random number to optimize the encoding method, the encoding operation is completed. After the subsequent sampling and interference operation, the data is sent to the data collector for decoding and statistical analysis.

[0047] Third, the lightweight data desensitization technology for text type data in accordance with the e-differential privacy is a lightweight privacy protection technology (Local, Private, Efficient Protocols Succinct Histograms Based on non-cryptographic Hash Algorithm, NCHOSH) based on a non-encrypted hash coding. The technology supports a transmission cost of lambda+l bits to maintain the lightweight of the overall desensitization technology. The non-encrypted hash function is introduced to realize fast coding of user data, the data range is reduced by combining the data mapping operation, and the data is sent to the data collector for decoding and statistical analysis after the subsequent sampling and interference operations.

[0048] Fourth, the statistical item adjustment technology. The technology introduces an adjustable parameter lambda2. At the end of the first round of interaction, the data collector side needs to return a topk candidate value set to optimize the communication cost of the second round of interaction. The parameter is introduced to adjust the size of the returned topk candidate value set.

[0049] Fifth, the user grouping technology of the present application. The meaningless operation in the first round of interaction is eliminated, and the utilization of the privacy budget is enhanced. The OSH algorithm and the NCHOSH algorithm are used for different types of data to realize data desensitization operation, protect user privacy from the source, and reduce the communication cost. The statistical item adjustment technology makes the size of the topk candidate value set returned at the end of the first round of interaction adjustable, enhances the adaptability and accuracy of the technology, and realizes the double-round interaction privacy protection scheme for high-dimensional data multi-value scene by combining the four technologies. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0051] Figure 1 The flowchart of the method of the present application;

[0052] Figure 2 The overall structure of the OSH algorithm;

[0053] Figure 3 The overall design of the NCHOSH algorithm. DETAILED DESCRIPTION

[0054] The embodiments of the present application will be described in detail below with reference to the drawings.

[0055] It should be noted that the following embodiments and features in the embodiments can be combined with each other in the case of no conflict; and all other embodiments obtained by those skilled in the art based on the embodiments in the present disclosure without creative labor are within the scope of protection of the present disclosure.

[0056] It should be noted that various aspects of the embodiments described below are within the scope of the appended claims. As will be apparent, the aspects described herein can be implemented in various ways, and that any particular structure is merely an example. Based on this disclosure and the knowledge of one of ordinary skill in the art, it will be apparent how to make and use a one aspect described herein, independently of any other aspect, and that any two or more aspects can be combined in various ways. For example, an apparatus can be implemented and / or a method can be practiced using any number of the aspects set forth herein. In addition, such an apparatus can be implemented and / or such a method can be practiced using other structure and / or functionality in addition to or other than one or more of the aspects set forth herein.

[0057] Symbol explanation: user user i Set of answers candidates i Set of candidate values candidates, result message report sent by the first-stage data producer i Frequent item set of the set of user answers sent by the first-stage data collector Result message report sent by the second-stage data producer, frequent item set frequency statistics result estimations generated by the second-stage data collector topk .

[0058] As shown in Figure 1 , the embodiment of the present application is a two-round privacy protection method based on differential privacy, which includes the interaction of two-stage data producers and data collectors:

[0059] User grouping: the data producer randomly samples the local data, and the sampled user answers are processed according to the selected single-value frequency statistics method to realize user grouping, and select the users participating in the first and second-stage privacy interaction and the users participating in only the second-stage privacy interaction;

[0060] First-stage privacy interaction: the data producer extracts a single user answer value from the set of user answers candidates i according to the sampling position, combines the total privacy budget ε, generates a result message report report i and sends it to the data collector; the data collector generates a result message report report iand the total privacy budget, generate frequency estimations of each candidate value, filter the candidate value set according to the candidate set adjustable parameter λ2 and the candidate value frequency estimations, and select high-frequency items from the candidate value set candidates as the second-stage frequent item candidate set to the data producer end;

[0061] Second-stage privacy interaction: the data producer generates a frequent item candidate set to the user according to the frequent item candidate set i filter and fill, and generate user message sets reports according to user grouping and send them to the data collector; the data collector end groups and counts the collected user message sets reports to finally determine the frequent item set estimations of the user answers topk .

[0062] Embodiments of the present application convert single-time multi-answer questions of users into single-value questions for processing, and propose adjustable parameters of user grouping and statistical item set size to balance the division of privacy budget and the consumption of communication cost, so that better frequency statistical effect is achieved under low communication cost.

[0063] The specific process of user grouping is described in detail as follows:

[0064] User grouping: the data producer randomly samples local data, processes the sampled user answers according to the set single-value frequency statistical method, realizes user grouping, and selects users participating in the first and second stages of interaction and users participating in only the second stage of interaction; specifically:

[0065] Both stages of the common two-round interaction privacy protection scheme need to divide the total privacy budget, and since the user answers need to be filled or truncated to the same length, the possibility of unnecessary privacy budget division is increased. Based on this problem, the present application proposes a user grouping rule, which divides users into two camps corresponding to different interaction rounds to avoid unnecessary privacy budget division. The specific user grouping rule is as follows:

[0066] (1) Set the threshold λ1 of user answers for one round of interaction, and calculate the selection interval [λ1l, (1+λ1)l] based on the threshold λ1 and the sampling length l to select users user i whose answer set candidates i participate in the first stage of interaction, and set the corresponding participation flag flag.

[0067] This step randomly selects a user user iThe answer is combined with a round of interaction processing, and the local user user i Grouping is performed to reduce the bias in the selection of statistical candidate value sets caused by excessive data preprocessing, i.e., virtual item filling and truncation, in the first round;

[0068] (2) The filtered user user i The answer set candidates i The set length (1+λ1)l-1) is truncated to fill the insufficient user answer set with virtual items; the truncated user answer data is randomly sampled sample_bit. If the sampled bit falls on the virtual item filling bit, the user is abandoned from participating in the first stage interaction and the participation flag flag is updated; finally, the retained sample bit sample_bit and participation flag flag are output.

[0069] This step will use the user user according to the above process i Users are divided into those who participate in the first-stage interaction and those who do not. For users who do not participate in the first-stage interaction, the privacy budget is no longer divided.

[0070] User grouping algorithm 1 description.

[0071] Algorithm 1: User grouping function

[0072] Input: User user i The answer set candidates i , sampling length is l, threshold λ1

[0073] Output: participation flag flag, sampling bit sample_bit

[0074]

[0075]

[0076] The following is a detailed description of the execution process of the first phase:

[0077] The first stage results message generation: the data generator answers candidates based on the user corresponding to the sampling position i and the total privacy budget ε, interference generation result message report i and sent to the data collector.

[0078] The users participating in the first round of interaction, determined by the user grouping stage described above, will be used to locally divide the total privacy budget for the first phase. The perturbation scheme will select the OSH algorithm or NCHOSH algorithm proposed in this invention based on the selection of the numerical type and the variability of the candidate value set. The algorithm description of the random perturbation of the data generator in the first phase is shown in Algorithm 2.

[0079] Algorithm 2: Result message report i Generative Algorithms

[0080] Input: total privacy budget ε, user user i The answer set candidates i , sampling bit sample_bit

[0081] Output: user user i Result report i

[0082] 1: data = candidates i [sample_bit]

[0083] 2:

[0084] 3: return report i

[0085] The first stage generates the candidate value set of frequent items: the data collector reports the result message i And the total privacy budget, generate frequency estimations of each candidate value, filter the candidate value set according to the candidate set adjustable parameter λ2 and the candidate value frequency estimation, and select high-frequency items from the candidate value set candidates as the frequent item candidate set of the second stage Sent to the data generator.

[0086] In order to reduce the problem of reduced statistical accuracy in the second round caused by the selection bias of the candidate value set in the first round, an adjustable parameter λ2 of the candidate value set returned in the first round is introduced. This parameter determines the size of the topk set returned in the first round. By setting this parameter, the accuracy of the final topk inference result can be improved while reducing the communication cost in the second round. Regardless of whether the user participates in the first round of interaction, the data collector will use the data obtained based on the participating users. The candidate value set is sent to all users. The algorithm description of the one-stage data collector statistical candidate value set selection scheme with the introduction of the adjustable parameter λ2 is shown in Algorithm 3.

[0087] Algorithm 3: One-stage data collector statistical candidate value set generation module

[0088] Input: user report set reports, privacy budget ε, candidate value set candidates, parameter λ2

[0089] Output: frequent item

[0090] 1: estimations = FO(reports, ε)

[0091] 2:

[0092] 3:

[0093] 4: return

[0094] The execution process of the second stage is described in detail as follows:

[0095] User answer filtering, filling and desensitization: the data producer filters the frequent item candidate set returned by the data collector in the first stage according to the user answer set Filter the non-high-frequency items in the user answer set candidates i , and set the length of the final user answer set to the length of the high-frequency candidate value set. The user answer set that does not meet the length is filled. According to whether the user participates in the first stage interaction, the available privacy budget value of each candidate value of the user is calculated. When the user participates in the first stage interaction, the privacy budget is When the user does not participate in the first stage interaction, the privacy budget is Because the single value frequency statistics method is used, the privacy budget is divided equally to each value, so there is a division operation, and the total privacy budget is divided by λ2k. Using the privacy budget, the user answer set is disturbed, and the disturbance result is stored in the reports array. The reports array and the corresponding flag bit are sent to the data collector.

[0096] In this embodiment, the user needs to determine the privacy budget of the second stage according to the self-grouping situation obtained in the first round in the second round of interaction. The local answer is disturbed according to the corresponding privacy budget of the grouping, and the data disturbance algorithm is described as shown in Algorithm 4.

[0097] Algorithm 4: Second stage user desensitization generation

[0098] Input: total privacy budget ε, user user i answer set candidate i , frequent item set candidate value set first round participation flag flag

[0099] Output: user i Result message report

[0100]

[0101] User answer frequent item set determination: the data collector groups and counts the collected data to determine the frequent item set answered by the user.

[0102] The second stage data collector determines whether the user participates in the first round of interaction according to the user's interaction record, and then sets the corresponding privacy budget. If the user participates in the first stage, the total privacy budget is set to ε / 2, and if the user does not participate in the first stage, the total privacy budget of the user is set to ε; and the user is divided into different groups according to the privacy budget. In each group, the data collector determines the Frequent item set candidate set A new round of frequency statistics is performed on the message data collected in the second stage, and the statistical process filters the virtual filling items. The data collector finally integrates and determines the final topk frequent item set estimations according to the frequency statistics of the two groups of users topk The second stage topk frequent item set determination algorithm is described as algorithm 5.

[0103] Algorithm 5: Second stage data frequent item set determination

[0104] Input: user message set reports, total privacy budget ε, frequent item set candidate set

[0105] Output: frequent item set frequency statistics result estimations topk

[0106] 1: if user participate in step 1 do

[0107] 2:

[0108] 3:

[0109] 4: return estimations topk

[0110] Next, two innovative privacy interference technologies (i.e. desensitization technology) used in the architecture are introduced:

[0111] OSH and NCHOSH are single-value frequency statistics methods, and the following algorithm description is that the algorithm input at the data producer end is one candidate in the user answer set candidates i In actual use, the algorithm is operated on each answer value in the user answer set candidates i , and then an array composed of the algorithm results of each answer value is returned to the data collector end. The message in the message set reports received by the data collector end in the algorithm description refers to the result returned after the algorithm operation on each user answer value, which is equivalent to a large array formed by splicing the message arrays of each user answer in actual use. The description in the specification that the embodiments of the present application convert single-time multiple-answer questions of users into single-value questions for processing is embodied in the above. A single-value question is that a user will only select one candidate value for answering at a time, which is single-time single answering. The algorithm descriptions of OSH and NCHOSH are written under this condition. Single-time multiple answering is that a user will return multiple candidate values at a time, which is the case of double-round protection method. Converting single-time multiple-answer questions of users into single-value questions for processing is similar to treating single-time multiple answering as multiple-time single answering.

[0112] 1. Efficient privacy protection histogram estimation algorithm based on encoding optimization (Local, Private, Efficient Protocols Succinct Histograms Based on Encoding Optimization, OSH)

[0113] The overall structure of the OSH algorithm is shown in Figure 2 . First, a d x m random projection matrix M is generated by the data collector end, and the matrix is ensured to be a column-orthogonal matrix, and the elements in the matrix are positive and negative . The matrix is sent to the data producer end, and the user data is encoded by randomly extracting matrix columns to help users to reduce the dimension from m-dimensional space to d-dimensional space. After sampling and interference operations, the data is transmitted back to the data collector end. The specific algorithm implementation is shown in Algorithm 6.

[0114] First, a d x m random projection matrix M is generated by the data collector end, and the matrix is ensured to be a column-orthogonal matrix, and the elements in the matrix are positive and negative . The matrix is sent to the data producer end;

[0115] Second, the data producer generates a random encoding bit j from the interval (0, m-1), and encodes the user answer set candidates i using the projection matrix M and the random encoding bit j to generate a mapping vector d_vector representing user input and a transmission bit i;

[0116] Again, with The probability of maintaining the true answer is The probability of setting the answer to the opposite case, ε is the total privacy budget, and the value of the i-th dimension of the mapping vector d_vector is used to generate the message result report i .

[0117] Algorithm 6: OSH algorithm user-side execution process

[0118] Input: User original data candidate i , the server sends the projection matrix M, the privacy budget ε, and the candidate set size m

[0119] Output: User coding result report i , random bit j

[0120] 1: Generate random code bit j = getSampleBit(0,m-1)

[0121] 2: Generate the mapping vector d_vector representing the user input and the transmission bit i, d_vector,i=Encode(M,j,candidate i )

[0122] 3:

[0123] 4:

[0124] d_vector i Represents the value of the i-th dimension of the mapping vector d_vector.

[0125] 5:return(report i ,j)

[0126] The interference algorithm uses a certain probability The probability of maintaining the true answer is The probability of setting the answer to the opposite case and sending it to the server, where It can be seen that this step ensures that the design of the perturbation scheme satisfies ε-differential privacy.

[0127] Report for any bit i , this bit is The probability is set to c ε dx j (For ease of writing, d_vector i Use x j Instead), The probability is set to -c ε dxj The expected value of the noise added to it is 0, that is, the sum of the weighted contributions of the two noises is: It can be obtained that the statistics is an unbiased estimate of the true situation, that is, the data collector does not need to correct the statistical results for the bits.

[0128] The data collector side needs to generate a d x m random projection matrix M to complete data collection, data decoding and other operations, and the specific process is:

[0129] First, the data collector side initializes a statistical vector reports with each dimension data being 0, and according to each user's answer message report i and the random encoding bit j, adds report i to the statistical vector in the random encoding bit j dimension to generate the final statistical vector reports;

[0130] Second, the data collector side encodes each candidate value in the candidate value set candidates according to the random projection matrix M. Each candidate value needs to be encoded and calculated with all the encoding matrix columns, and the sampling bit corresponding to each encoding result is extracted to form the decoding vector Me candidate of the candidate value.

[0131] Finally, the statistical vector reports and the decoding vector Me candidate corresponding to each candidate value are combined for inverse mapping. The projection of the decoding vector to the statistical vector is the frequency of the candidate value.

[0132] Algorithm 7 Data collector side algorithm

[0133] Input: user data {userInfo i ∈candidates: 1≤i≤n}, privacy budget ε, confidence parameter β, 0≤β≤1, candidate set size m

[0134] Output: the frequency freq(candidate) corresponding to each candidate

[0135]

[0136]

[0137] 2. Lightweight privacy protection technology based on non-encryption hash coding (Local, Private, Efficient Protocols Succinct Histograms Based on non-cryptographic Hash Algorithm, NCHOSH)

[0138] The overall design of the NCHOSH algorithm is shown in Figure 3 As shown in the figure, the data producer end includes a data encoding module, a data mapping module, and a perturbation module. The data encoding relies on a non-encrypted hash function to realize fast encoding of user data. The encoded data is subjected to a mapping process, which projects data in a large range to a smaller range, thereby reducing communication costs and providing a layer of protection for user privacy through data collision. The encoded and mapped data is subjected to a privacy data desensitization scheme based on generalized random response (GRR) at the user's local end. The desensitized data is sent back to the data collector end. The specific algorithm process is shown in Algorithm 8.

[0139] First, determine the privacy budget ε, set the mapping interval length d and the random seed length l; generate parameters p and q according to the privacy budget ε and the mapping interval length d;

[0140] Second, generate a random seed res with a length of 1, encode the user data using a hash algorithm, and reduce the dimension of the encoded result using a modulo operation;

[0141] Third, disturb the encoded result. With a probability of 1-p+q, it remains unchanged, and with a probability of p-q, it randomly selects a number from 0 to d-1;

[0142] Finally, send the encoded result user message report and the random seed to the data receiver.

[0143] Algorithm 8 NCHOSH algorithm user end randomizer module input: user data candidate, privacy budget parameter ε, mapping interval length d, random seed length l output: user answer report, random seed random_seed

[0144] 1: Define

[0145] 2: random_seed = generateRandom(l)

[0146] 3: res = Encode(candidate) ← XXH3 / XXH128

[0147] 4: report = mapping(res) ← res mod d / Hadamrd Response

[0148] 5:

[0149] (Interference on the encoding result, keep unchanged with probability 1-p+q, and randomly select a number from 0 to d-1 with probability p-q)

[0150] 6: return(report, random_seed)

[0151] The data de-sensitization scheme of NCHOSH is based on the GRR multi-dimensional random response mechanism: the user encoding and the mapping answer will select a random answer of non-real answer in the λ-dimensional mapping space with the probability of q, and maintain the real answer with the probability of p. Define the noise addition function as A, and d is the length of the mapping interval, that is, 2 λ For any input v of the user, there is: By The interference function A meets the ε-differential privacy.

[0152] To achieve privacy protection and reduce communication cost, it is necessary to map the encoded user data to an interval with a length of d. By calculating the relationship between the variance of the frequency estimate and d, it is obtained that when d = e ε +1, the variance is minimum. In real application scenarios, the selection scheme of d cannot be simply set as e ε +1. To ensure that the collision result does not deviate due to periodicity and other problems, the optimal result is usually obtained when a prime number closest to e ε +1 is selected.

[0153] After the de-sensitized data is sent to the data collector end, the data collector end needs to perform statistics, correction and frequency estimation. The specific algorithm is described in Algorithm 9.

[0154] First, for each user message report and random seed res, encode each candidate value in the overall candidate value set candidates using the random seed res. The encoding result of each candidate value is encode_candidate.

[0155] Second, after dimension reduction operation on each candidate value encoding result encode_candidate, compare it with the user message report. If they are the same, add one to the dimension corresponding to the candidate value in the frequency statistical vector freq.

[0156] Third, after the above operation is performed on all user answers, an biased statistical vector freq is obtained.

[0157] Finally, the frequency calibration formula is used to calibrate the data of the biased statistical vector freq to obtain the unbiased estimation result of the frequency corresponding to each candidate value.

[0158] Algorithm 9 NCHOSH data collector side F0 module input: dataset of user data, set of candidate values candidates, privacy budget ε

[0159] output: frequency statistics result freq

[0160] 1: define

[0161] 2: for report in reports do

[0162] Process each user data as follows

[0163] 3: report, random_seed ← dataset

[0164] Take out the encoding result sent by the user and the random seed

[0165] 4: for candidate in candidates do

[0166] Process all candidate values in the candidate value set as follows

[0167] 5: encode_candidate = Encode(candidate, random_seed)

[0168] Encode the candidate value using the random seed used by the user

[0169] 6: freq candidate + = 1 ← if mapping(encode_candidate) == report

[0170] Compare the dimensionality reduction operation on the encoding result with the result message sent by the user, if they are the same, the frequency of the candidate value is incremented by one

[0171] 7: end for

[0172] 8: end for

[0173] 9:

[0174] The frequency estimate is a biased estimate, and the estimate result is calibrated

[0175] 10: return freq

[0176] Assume that there are n data producers, and the data producer generates data candidate ∈ candidates, t candidateThe number of real answers in the user is candidate, c candidate The number of answers for candidate is counted, and according to the interference rule:

[0177] t candidate ·p+(n-t candidate )·q=c candidate

[0178] Therefore, the frequency estimation after calibration should be:

[0179] The present application proposes a high-efficiency, light, and accurate privacy protection scheme suitable for each scene for high-dimensional data, and realizes good data statistical effect. The focus of the present application is: 1, two light data desensitization algorithms conforming to epsilon-differential privacy are proposed, 2, a double-round interaction privacy protection scheme for multi-value scene is proposed, 3, a user grouping rule is proposed to avoid meaningless division of privacy budget, and 4, a statistical item adjustment scheme is proposed.

[0180] The present application realizes the adjustable function of user grouping and statistical item set by introducing adjustable parameter design sampling optimization scheme, makes the overall privacy budget allocation more reasonable, at the same time introduces the improved coding disturbance algorithm proposed in the present application in the double-round interaction, reduces the communication cost, improves the accuracy of frequent item mining and obtains more stable model performance, realizes the balance of model performance and communication cost, and provides a new idea of frequency release in the field of multi-value frequency statistics.

[0181] The present application proposes a local differential privacy histogram estimation algorithm (OSH) based on coding optimization. Unlike the previous fixed matrix coding mapping scheme, the method adopts random mapping of data to a vertex of a hypercube, and replaces the random sampling method with a fixed sampling point based on local data generation, realizes the reduction of communication cost while improving the accuracy of numerical data statistical results. The algorithm uses a dimension reduction matrix to realize the mapping of high-dimensional numerical space to low-dimensional space, greatly reduces the communication cost of the algorithm, and optimizes the traditional random mapping matrix to a column orthogonal coding matrix, realizes dimension reduction while avoiding the problem of unstable model effect caused by randomness.

[0182] The application provides a local differential privacy histogram estimation algorithm (NCHOSH) based on non-encryption hash coding optimization, which is different from the avoidance of collision in the previous coding algorithm, the algorithm combines a fast hash algorithm and a local variable random seed mode to realize coding of user data, the existence of the local variable random seed enables the algorithm to optimize model performance by adjusting the dimension of the random seed when facing a high-dimensional candidate value set, so that better statistical results are obtained with smaller communication cost, data dimension reduction and uniform collision between candidate values are realized by optimizing the cell mapping rule, better model results are obtained with low communication cost in the case of non-encryption hash function and dimension reduction, the algorithm can avoid the assumption that most algorithms need to know the fixed candidate value set in advance, even if the candidate value set changes, the algorithm is still effective and the transmitted data will not have the problem of invalidation, the robustness of the algorithm is improved, the calculation complexity is low and the communication cost is adjustable, and the lightness of the overall algorithm is ensured.

[0183] The application simultaneously supports processing of numerical data and text data, supports a more comprehensive use scene, and can be applied to various data use scenes, such as user behavior data statistics, such as statistical numerical data of GPS data and motion duration, and statistical text information of user identity data, such as address, gender, work unit and the like.

[0184] The above is only a specific embodiment of the application, but the protection scope of the application is not limited to this, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the application, which should be covered in the protection scope of the application. Therefore, the protection scope of the application should be subject to the protection scope of the claims.

Claims

1. A two-round privacy preserving method based on differential privacy, characterized in that, The application relates to a privacy protection method and device for a data producer and a data collector. The application comprises the following contents: First stage privacy interaction: the data producer extracts a single user answer value from the set of user answers candidates i in accordance with the sampling bit, combines the total privacy budget ε, and generates a result message report i and sends it to the data collector; The data collector generates frequency estimations of each candidate value based on the report messages returned by all users i and the total privacy budget, generates frequency estimations of each candidate value, filters the candidate value set based on the candidate set adjustable parameter λ2 and the candidate value frequency estimations, and selects high-frequency items from the candidate value set candidates as the second-stage frequent item candidate set to the data producer end; Second stage privacy interaction: data producer filters and fills the answer set candidates based on the frequent item candidate set to the user i and sends the report to the data collector. The data collector end will collect the user message report to group and statistics, and finally determine the frequent item set estimations of user answers topk .

2. The two-round privacy preserving method based on differential privacy according to claim 1, characterized in that, User grouping: the data producer randomly samples local data, and the sampled user answers are processed according to a selected single-value frequency statistical method to realize user grouping and select users participating in the first and second stage privacy interactions and users participating in only the second stage privacy interaction; (1) Set a threshold λ1 for the user to answer a round of interaction, and calculate the selection interval [λ1l, (1+λ1)l] based on the threshold λ1 and the sampling length l, and screen the users whose number of answer set candidates i is in the interval i participate in the first stage of interaction, and set the corresponding participation flag flag; (2) the user user i answer set candidates i The length of the answer set of the user is less than the length of the answer set of the user. The virtual item is filled. The length of the user answer data is randomly sampled sample_bit. If the sample bit falls on the virtual item filling bit, the user is abandoned in the first stage interaction, and the participation flag flag is updated. The final output of the sample bit and the participation flag flag is retained. 3.The two-round privacy preserving method based on differential privacy according to claim 1, wherein, The specific process of the user grouping is as follows: Filtering: The data producer filters the set of frequent item candidates returned by the first phase data collector Filtering the user answer set to remove non-high frequency items from it; In the second stage privacy interaction, the following operations are performed: Padding: the length of the user final answer set is set as the length of the high-frequency item candidate value set, and the user answer set that does not meet the length is padded; Privacy budget: the data producer calculates the available privacy budget value of each candidate value according to whether the user participates in the first stage interaction; 4. The method of claim 3, wherein, When the user participates in the first stage of interactive computation, the privacy budget is When the user does not participate in the first stage of interactive computation, the privacy budget is where ε is the total privacy budget, and k is the number of final results of the double-round interaction set.

5. The method of claim 3, wherein, Frequent item set generation: according to the privacy budget and the answer set of the user, a user report is disturbed and sent to the data collector. Within each group, the data collector performs a new round of frequency statistics based on the set of frequent item candidates and the set of user reports, and the statistics process filters out the dummy items. The data collector finally integrates the two sets of frequency statistics to determine the final set of top-k frequent item estimations topk .

6. The two-round privacy preserving method based on differential privacy according to claim 1, characterized in that, In the second stage privacy interaction, the data collector side sets the privacy budget according to whether the user participates in the first stage privacy interaction, sets the total privacy budget as epsilon / 2 if the user participates in the first stage, sets the total privacy budget as epsilon if the user does not participate in the first stage, and sets different groups for the user according to the privacy budget; 7. The two-round privacy preserving method based on differential privacy according to claim 6, characterized in that, When the answer set of the user is of a numerical type, an OSH algorithm is used to disturb the generation of a result report, and when the answer set of the user is of a text type, an NCHOSH algorithm is used to disturb the generation of a result report. First, a d x m random projection matrix M is generated by the data collector, and it is ensured that the matrix is a column-orthogonal matrix, and the elements in the matrix are positive and negative The matrix is issued to the data producer; Secondly, the data producer generates a random encoding bit j from the interval (0, m-1), and encodes the user answer set candidates i using the projection matrix M and the random encoding bit j, to generate a mapping vector d_vector representing the user input and a transmission bit i. Again, with probability the true answer is maintained, and with probability the answer is set to the opposite, ε is the total privacy budget, and the value of the i-th dimension of the mapping vector d_vector is used to generate the message result report i .

8. The two-round privacy preserving method based on differential privacy according to claim 7, characterized in that, The OSH algorithm is as follows: First, the data collector end initializes a statistical vector reports with all dimensions being 0, and according to the answer message report of each user collected i and the random encoding bit j, adds report in the random encoding bit j dimension of the statistical vector i to generate the final statistical vector reports; Secondly, the data collector encodes each candidate value in the candidate value set candidates according to the random projection matrix M, each candidate value needs to be encoded and calculated with all the encoding matrixes, and the sampling bits corresponding to each encoding result are extracted to form the decoding vector Me of the candidate value candidate ; Finally, the statistical vector reports is compared with the decoding vector Me corresponding to each candidate value candidate The projection of the decoding vector to the statistical vector, i.e. the frequency of the candidate value corresponding to the sought candidate value, is determined in connection with the inverse mapping.

9. The method of claim 6, wherein, When the OSH algorithm is used to disturb the generation, the process of the data collector for processing is as follows: The NCHOSH algorithm is as follows: Firstly, the privacy budget epsilon is determined, the mapping interval length d and the random seed length l are set, and parameters p and q are generated according to the privacy budget epsilon and the mapping interval length d; Secondly, a random seed res with a length of l is generated, user data is encoded by using a hash algorithm, and the encoded result is reduced in dimension by using a remainder operation; Thirdly, the encoded result is disturbed, and a number is randomly selected from 0 to d-1 with a probability of p-q and remains unchanged with a probability of 1-p+q; 10. The two-round privacy preserving method based on differential privacy according to claim 9, characterized in that, Finally, the encoded result user report and the random seed are sent to the data receiver. When the NCHOSH algorithm is used to disturb the generation, the process of the data collector side for decoding is as follows: Firstly, for each user report and random seed res, each candidate value in the whole candidate value set candidates is encoded by using the random seed res, and the encoding result of each candidate value is encode_candidate; Secondly, after the dimension reduction operation is performed on each candidate value encoding result encode_candidate, the user report is compared, if the same, the dimension corresponding to the candidate value in the frequency statistical vector freq is added by one; Thirdly, after the above operation is performed on all user answers, a biased statistical vector freq is obtained; Finally, the frequency calibration formula is used to calibrate the data of the biased statistical vector freq, and the frequency unbiased estimation result corresponding to each candidate value is obtained.

Images