Differential privacy-based double-round privacy protection method

Abstract

The invention belongs to the technical field of privacy protection, and particularly relates to a double-round privacy protection method based on differential privacy. The specific process of the method is as follows: user grouping: in a first round of interaction, users are grouped by judging whether the number of candidate values of the users is within an interval specified by a set variable parameter and judging whether user sampling candidate values are meaningless candidate values, part of the users do not need to participate in a first stage, and part of the users do not need to participate in a second stage; reducing meaningless privacy budget segmentation through user grouping operation; the first-stage interaction is used for screening out high-frequency items from the candidate value set to serve as a frequent item candidate set of the second stage; in the second stage of interaction, after the user side filters the answers which are not in the high-frequency candidate value set, all the answers interact with the data collector side by using the provided lightweight single-value frequency counting method, and the candidate value frequency is estimated to obtain a final user high-frequency answer result.

Description

Technical Field

[0001] The present invention belongs to the technical field of privacy protection, and in particular relates to a two-round privacy protection method based on differential privacy. Background Art

[0002] The issue of privacy protection was first raised in the 1970s. With the advent of the big data era, a large amount of personal data is collected, including personal privacy information. How to protect this information has become a hot topic. Differential privacy was first defined in 2006. Its core is to add random noise to the data to cover up the impact of individual data on the overall query results, thereby preventing attackers from using the output results to infer personal information. It provides mathematically provable privacy protection, strictly defines the attack model, and provides strong protection in theory. At the same time, differential privacy protection is applicable to a wide range of scenarios and is one of the main methods of privacy protection today.

[0003] In scenarios with multiple data sources and multiple local data, high-dimensional problems are high-dimensional candidate value sets and multiple local answer scenarios. In order to ensure data availability and the accuracy of model estimation, the encoding results and sending result dimensions of user answers are often expanded with the candidate value dimensions and interference schemes and other algorithm design methods. High-dimensional user answer candidate sets and high-dimensional user answer sets will inevitably increase the dimensionality of the data to be desensitized, resulting in reduced data availability, difficulty in model fitting on the data collector side, and increased algorithm complexity. At this time, it is necessary to consider not only the design of privacy protection solutions, but also the efficiency of algorithms in engineering scenarios. By designing appropriate encoding, desensitization and other solutions, the dimensionality of user local data can be reduced, and the lightweight nature of the algorithm and data availability can be maintained.

[0004] Patent CN202211156864.3 proposes a big data frequency estimation method based on local differential privacy. The invention first samples the client by the server. The client sends the data processed by hash function and differential privacy protection method to the server. The server uses these data to train a frequency prediction model. The server sends the trained model and the corresponding high-frequency data threshold to the client. The client determines 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. The invention first groups users by the server, and makes the grouping results and the number of data items contained in each group, privacy budget and other information public. Users fill or truncate the answers based on the information received, and then flatten and interfere with the data and send it back to the server. The server constructs an initial heavy hitters set, calculates the frequency estimate value in the answer data of each group of users according to the order of user grouping, iterates this value with the initial set, and constructs 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 required set.

[0006] Patent CN202310212497.2 proposes a top-k network flow estimation method based on local differential privacy. First, the user is randomly assigned or joins a group, the corresponding network flow size and network flow cardinality are counted, and the results are sent to the server. After the server records these data, it calculates the corresponding optimal privacy protection parameters for users in different groups and sends them back to the corresponding users. The user on the client stores the network flow, perturbs the data using the optimal parameters provided by the server, and sends the perturbed data back to the server. The server aggregates these perturbed data and finally iteratively estimates the top-k network flow.

[0007] At present, there are three main problems with privacy protection technology in multi-value scenarios based on two rounds of interaction. The first is the problem of over-segmentation of the privacy budget. The second is the problem of reduced statistical accuracy in the second round caused by the selection bias of the candidate value set in the first round. The third is the problem of poor interference algorithm effect in the privacy protection process. For example, the operation of the first round of interaction of the LDPMiner algorithm is that the user samples and interferes with the answer. The server collects the sampled data to infer the statistical candidate value set containing the topk, and sends the set to the user end to reduce the communication cost of the second round of interaction. In the LDPminer algorithm, all users participate in the first round of interaction and divide the privacy budget for this purpose, but not all users' answers are necessary. There are some users whose unnecessary answers occupy the privacy budget. In addition, the selection bias of their topk statistical candidate value set will affect the second round of interaction. The interference methods used in the LDPminer algorithm are SH and RAPPOR algorithms. The statistical effect of the SH algorithm is strongly dependent on the generation of the encoding random matrix, so it shows the problem of unstable statistical effect in real experiments. The RAPPOR algorithm chooses to use Bloom filters to encode data, but Bloom filters are a probabilistic data structure with a certain error rate. The encoding idea relies on binary vectors and a series of random mapping functions. Its main purpose is to determine whether the queried element exists rather than encoding, which makes the interference algorithm ineffective. Summary of the invention

[0008] In view of this, the present invention proposes a two-round privacy protection method based on differential privacy, which can reduce meaningless privacy budget segmentation and enable data collectors to accurately estimate the user answers of data producers.

[0009] The technical solution of the present invention is as follows:

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

[0011] User grouping: The data generator randomly samples the local data, and the sampled user answers are processed according to the selected single-value frequency statistical method to achieve user grouping, selecting users who participated in the first and second stage privacy interactions and users who only participated in the second stage privacy interactions;

[0012] The first stage of privacy interaction: the data generator selects candidates according to the sampling position in the user answer set i Extract a single user's answer value and combine it with the total privacy budget ε to generate the result message report i And send it to the data collector; the data collector reports the result messages returned by all users iand the total privacy budget, generate frequency estimates of each candidate value, filter the candidate value set according to the candidate set adjustable parameter λ2 and the candidate value frequency estimate, and select high-frequency items from the candidate value set candidates as the frequent item candidate set in 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 data collectors; the data collector will collect user message reports for grouping and statistics, and finally determine the frequent item sets of user answers. topk .

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

[0015] (1) Set the threshold λ1 for a 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 selected user user i The answer set candidates i Truncate the set length (1+λ1)l-1), and fill the user answer set of insufficient length with virtual items; randomly sample sample_bit for the intercepted user answer data. If the sampled bit falls on the virtual item filling bit, the user will give up participating in the first stage of interaction and update the participation flag flag; finally output the retained sample bit sample_bit and participation flag flag.

[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 phase. 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 set that does not meet the above length;

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

[0021] Frequent item set generation: Based on the privacy budget and the user's answer set, the user message report is generated and sent 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 computing, 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 candidate values ​​with the highest frequency 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, in the second stage of privacy interaction of the present invention, the data collector sets the privacy budget according to whether the user participates in the first stage of privacy interaction. If the user participates in the first stage of execution, the total privacy budget is set to ε / 2; if the user does not participate in the first stage of execution, the total privacy budget is set to ε, and the user is divided into different groups according to the 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 fill-in items. The data collector will finally determine the final top k frequent item sets 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, and 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 d×m random projection matrix M, and ensures that the matrix is ​​a column orthogonal matrix with positive and negative elements. 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 reply 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 is The probability of setting the answer to the opposite case, ε is the privacy budget, and the value of the i-th dimension of d_vector is used to generate the message result report i .

[0030] Furthermore, when the OSH algorithm is used to generate interference, the data collector performs the following processing:

[0031] First, the data collector initializes a statistical vector reports with all dimensions of data being 0, and reports the answer messages of each user collected. i and random code bit j, add report to the random code bit j dimension of the statistical vector i , generate the final statistical vector report;

[0032] 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 matrix columns, and the sampling bits corresponding to each encoding result are extracted to form the decoding vector Me of the candidate value. candidate ;

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

[0034] Furthermore, the NCHOSH algorithm described in the present invention is:

[0035] 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;

[0036] Secondly, a random seed res of length l is generated, the user data is encoded using a hash algorithm, and the encoded result is reduced in dimension using a remainder operation;

[0037] Again, the encoding result is disturbed, the probability of which remains unchanged with 1-p+q, and a number is randomly selected from 0 to d-1 with probability pq;

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

[0039] Furthermore, when the NCHOSH algorithm is used to generate interference, the decoding process of the data collector is as follows:

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

[0041] Secondly, after the dimension reduction operation is performed on each candidate value encoding result encode_candidate, it is compared with the user message report. If they are the same, the dimension of the frequency statistics vector freq corresponding to the candidate value is increased by one;

[0042] Again, after performing the above operation on all user answers, a biased statistical vector freq is obtained;

[0043] Finally, the frequency calibration formula is used to calibrate the data of the biased statistical vector freq to obtain the unbiased frequency estimation results corresponding to each candidate value.

[0044] Beneficial effects:

[0045] First, user grouping technology introduces an adjustable parameter λ1. In the first round of interaction, users are grouped by judging whether the number of candidate values ​​of users is within the interval specified by the parameter and whether the user's sampled candidate values ​​are meaningless candidate values. The user grouping operation reduces meaningless privacy budget segmentation.

[0046] Second, a lightweight data desensitization technology that complies with ε-differential privacy for numerical data, and an efficient privacy-preserving histogram estimation algorithm based on encoding optimization (Local, Private, Efficient Protocols SuccinctHistograms Based on Encoding Optimization, OSH), which supports The transmission cost is reduced to maintain the lightweight of the overall desensitization technology. The ordinary coding matrix is ​​converted into a column orthogonal matrix by improving the generation of the coding matrix. At the same time, random numbers are introduced to optimize the coding method to complete the coding operation. After subsequent sampling and interference operations, the data is sent to the data collector for decoding and statistical analysis.

[0047] Third, for text-type data, a lightweight data desensitization technology that complies with ε-differential privacy is used, and a lightweight privacy protection technology based on non-cryptographic hash coding (Local, Private, Efficient Protocols SuccinctHistograms Based on non-cryptographic Hash Algorithm, NCHOSH) is used. This technology supports the transmission cost of λ+lbit to maintain the lightweight of the overall desensitization technology. By introducing non-cryptographic hash functions to achieve fast encoding of user data, combined with data mapping operations to narrow the data range, and after subsequent sampling and interference operations, the data is sent to the data collector for decoding and statistical analysis.

[0048] Fourth, statistical item adjustment technology, which introduces an adjustable parameter λ2. At the end of the first round of interaction, the data collector needs to return the topk candidate value set to optimize the communication cost of the second round of interaction. By introducing this parameter, the size of the returned topk candidate value set is adjusted.

[0049] Fifth, the user grouping technology of the present invention eliminates meaningless operations in the first round of interaction and enhances the utilization of the privacy budget; the OSH algorithm and the NCHOSH algorithm implement data desensitization operations for different types of data types, protect user privacy from the source, and reduce communication costs; 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, which enhances the adaptability and accuracy of the technology. The combination of the four technologies realizes a two-round interaction privacy protection solution for high-dimensional data multi-value scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0051] Figure 1 is a flow chart of the method of the present invention;

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

[0053] Figure 3 Overall design for the NCHOSH algorithm. DETAILED DESCRIPTION

[0054] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0055] It should be noted that the following embodiments and features in the embodiments may be combined with each other in the absence of conflict; and, based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making any creative work are within the scope of protection of the present disclosure.

[0056] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein may be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present disclosure, it should be understood by those skilled in the art that an aspect described herein may be implemented independently of any other aspect, and two or more of these aspects may be combined in various ways. For example, any number of aspects described herein may be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein may be used to implement this device and / or practice this method.

[0057] Symbol Description: User i The answer set candidates i , candidate value set candidates, report, the result message sent by the data producer in the first stage i , the frequent itemsets of the user answer set sent by the data collector in the first stage The result message report sent by the data producer in the second stage, and the frequency statistics of frequent itemsets generated by the data collector in the second stage topk .

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

[0059] User grouping: The data generator randomly samples the local data, and the sampled user answers are processed according to the selected single-value frequency statistical method to achieve user grouping, selecting users who participated in the first and second stage privacy interactions and users who only participated in the second stage privacy interactions;

[0060] The first stage of privacy interaction: the data generator selects candidates according to the sampling position in the user answer set i Extract a single user's answer value and combine it with the total privacy budget ε to generate the result message report i And send it to the data collector; the data collector reports the result messages returned by all users iand the total privacy budget, generate frequency estimates of each candidate value, filter the candidate value set according to the candidate set adjustable parameter λ2 and the candidate value frequency estimate, and select high-frequency items from the candidate value set candidates as the frequent item candidate set in the second stage Send to the data generator end;

[0061] 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 report sets according to user grouping and send them to the data collector; the data collector will collect user report sets for grouping and statistics, and finally determine the frequent item sets of user answers estimations topk .

[0062] The embodiment of the present application converts a user's single multiple-answer question into a single-value question for processing, and proposes user grouping and adjustable parameters for the size of the statistical item set to balance the segmentation of the privacy budget and the consumption of communication costs, thereby achieving better frequency statistics effects under low communication costs.

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

[0064] User grouping: The data generator randomly samples the local data, and the sampled user answers are processed according to the set single-value frequency statistical method to achieve user grouping, selecting users who participate in the first and second stage interactions and users who only participate in the second stage interaction; specifically:

[0065] The two stages of the common two-round interactive privacy protection scheme need to split the total privacy budget. Since the user answers need to be padded or truncated to the same length, the possibility of unnecessary privacy budget splitting increases. Based on this problem, the present invention proposes a user grouping rule to divide users into two camps corresponding to different interaction rounds to avoid unnecessary privacy budget splitting. The specific user grouping rules are:

[0066] (1) Set the threshold λ1 for a 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.

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

[0068] (2) The selected user user i The answer set candidates i Truncate the set length (1+λ1)l-1), and fill the user answer set of insufficient length with virtual items; randomly sample sample_bit for the intercepted user answer data. If the sampled bit falls on the virtual item filling bit, the user will give up participating in the first stage of interaction and update the participation flag flag; finally output the retained sample bit sample_bit and participation flag flag.

[0069] In this step, the user user is 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 stage:

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

[0078] The users who participate in the first round of interaction are determined by the above user grouping stage. The total privacy budget for the first stage will be divided locally, and the perturbation scheme will select the OSH algorithm or NCHOSH algorithm proposed in this invention according to the selection of numerical type and the variability of the candidate value set. The algorithm description of the random perturbation of the data generator in the first stage 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 message report i

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

[0083] 2:

[0084] 3: return report i

[0085] The first stage generates a set of candidate values ​​for frequent items: the data collector generates a set of candidate values ​​for frequent items based on the result report. i and the total privacy budget, generate frequency estimates of each candidate value, filter the candidate value set according to the candidate set adjustable parameter λ2 and the candidate value frequency estimate, and select high-frequency items from the candidate value set candidates as the frequent item candidate set in 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 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 collect the topk inference results 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 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 Items

[0090] 1:estimations=FO(reports,ε)

[0091] 2:

[0092] 3:

[0093] 4:return

[0094] The following is a detailed description of the execution process of the second phase:

[0095] User answer filtering, filling and desensitization: The data generator generates the candidate set of frequent items based on the data collector's feedback in the first stage. Set candidates for user answers i Filter out the non-high-frequency items, 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 set that does not meet the above length; calculate the privacy budget value available for each candidate value based on whether the user participates in the first-stage interaction. When the user participates in the first-stage interactive operation, his privacy budget is When the user does not participate in the first phase of interactive computing, his privacy budget is Because the single-value frequency statistics method is used, the privacy budget must be divided equally for each value, so there is a division operation to divide the total privacy budget by λ2k. Using the privacy budget, the user answer set is disturbed, and the interference results are stored in the reports array, and the reports array and the corresponding flag are sent to the data collector.

[0096] In this embodiment, during the second round of interaction, users need to determine the privacy budget for the second stage based on their own grouping obtained in the first round, and perturb the local answer according to the privacy budget of the corresponding group. The data perturbation algorithm is described in Algorithm 4.

[0097] Algorithm 4: Second-stage user desensitization generation

[0098] Input: total privacy budget ε, user user i Answer set candidate i , candidate value set of frequent itemsets The first round of participation flag

[0099] Output: user user i Result message report

[0100]

[0101] Determination of frequent item sets of user answers: The data collector groups and counts the collected data and finally determines the frequent item sets of user answers.

[0102] In the second phase, the data collector will determine whether the user participated in the first round of interaction based on the user's interaction record, and then complete the setting of the corresponding privacy budget. If the user participated in the execution of the first phase, the total privacy budget will be set to ε / 2. If the user did not participate in the execution of the first phase, the total privacy budget of the user will be set to ε. The users will be divided into different groups according to their privacy budgets. Within each group, the data collector will set the corresponding privacy budget according to the privacy budget determined in the first phase. Frequent item set candidate set A new round of frequency statistics is performed with the message data collected in the second phase. The statistical process will filter out virtual filler items. The data collector will finally determine the final topk frequent item set estimations based on the frequency statistics of the two groups of users. topk ,The algorithm for determining the topk frequent itemsets in the second stage is described in Algorithm 5.

[0103] Algorithm 5: Determination of frequent itemsets in the second stage

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

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

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

[0107] 2:

[0108] 3:

[0109] 4: return estimations topk

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

[0111] OSH and NCHOSH are both single-value frequency statistics methods. The following algorithm describes the algorithm input on the data producer side as user answer sets candidates. i A candidate in the actual use is the user answer set candidiates i Perform algorithmic operations on each answer value, and then form an array of algorithm results for each answer value and return it to the data collector end. The messages in the user message set reports received by the data collector end in the algorithm description refer to the results returned after each user performs algorithmic operations on a single answer value. In actual use, it is equivalent to a large array spliced ​​together from the message arrays of each user's answers. There is a sentence in the specification that the embodiment of the present application converts the user's single multiple answer questions into single-value questions for processing, which is reflected in this. A single-value question means that the user will only choose one candidate value to answer at a time, which is a single single answer. The OSH and NCHOSH algorithm descriptions are written in this context. A single multiple answer means that the user will return multiple candidate values ​​at a time, which is the case of the double-round protection method. Converting a user's single multiple answer question into a single-value question for processing is similar to treating a single multiple answer as multiple single answers.

[0112] 1. Local, Private, Efficient Protocols Succinct Histograms Based on Encoding Optimization (OSH)

[0113] The overall structure of the OSH algorithm is as follows: Figure 2 As shown, the data collector first generates a d×m random projection matrix M, and ensures that the matrix is ​​a column orthogonal matrix and the elements in the matrix are positive and negative. The matrix is ​​sent to the data producer, and the matrix columns are randomly extracted to help encode the user data, thereby reducing the dimension from the high-dimensional m-dimensional space to the d-dimensional space. After sampling and interference operations, the data is transmitted back to the data collector. The specific algorithm implementation is shown in Algorithm 6.

[0114] First, the data collector generates a d×m random projection matrix M and ensures that the matrix is ​​a column orthogonal matrix with positive and negative elements. Send the matrix to the data producer;

[0115] Secondly, the data producer generates a random code bit j from the interval (0, m-1), and uses the projection matrix M and the random code bit j to answer the user's answer set candidiates i Encode to generate a mapping vector d_vector representing the 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 a 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 situation 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 of 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 added noise is 0, that is, the sum of the weighted contributions of the two noises is: It can be seen that the statistical situation is an unbiased estimate of the actual situation, that is, the data collector does not need to make corrections to the statistical results.

[0128] The data collector needs to generate a d×m random projection matrix M to complete operations such as data collection and data decoding. The specific process is as follows:

[0129] First, the data collector initializes a statistical vector reports with all dimensions of data being 0, and reports the answer messages of each user collected. i and random code bit j, add report to the random code bit j dimension of the statistical vector i , generate the final statistical vector reports;

[0130] 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 matrix columns, and the sampling bits corresponding to each encoding result are extracted to form the decoding vector Me of the candidate value. candidate ;

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

[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-cryptographic hash coding (Local, Private, Efficient Protocols Succinct Histograms Based on non-cryptographic Hash Algorithm, NCHOSH)

[0138] The overall design of the NCHOSH algorithm is as follows: Figure 3 As shown in the figure, the data producer side includes a data encoding module, a data mapping module and a perturbation module. The data encoding relies on a non-encrypted hash function to quickly encode the user data. The encoded data will undergo a one-step mapping process to project the data in a large range into a smaller range. While reducing the communication cost, it provides a layer of protection for user privacy through data collision. The encoded and mapped data will be desensitized locally by the user using a privacy data desensitization scheme based on a generalized random response (GRR). The desensitized data will be sent back to the data collector side. The specific algorithm flow 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] Secondly, a random seed res with a length of 1 is generated, the user data is encoded using a hash algorithm, and the encoded result is reduced in dimension using a remainder operation;

[0141] Again, the encoding result is disturbed, the probability of which remains unchanged with 1-p+q, and a number is randomly selected from 0 to d-1 with probability pq;

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

[0143] Algorithm 8NCHOSH algorithm User-side 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: Definition

[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] (interfere with the encoding result, keep it unchanged with a probability of 1-p+q, and randomly select a number from 0 to d-1 with a probability of pq)

[0150] 6: return (report, random_seed)

[0151] The data desensitization scheme of NCHOSH is designed based on the GRR multivariate random response mechanism: the user codes and maps the answer to a random solution that is not a real solution in the λ-dimensional mapping space with a probability of q, and the real answer will be maintained with a probability of p. The noise addition function is defined as A, and d is the length of the mapping interval, that is, 2 λ , for any user input v: Depend on It can be obtained that the interference function A complies with ε-differential privacy.

[0152] In order to achieve privacy protection and reduce communication costs, it is necessary to map the encoded user data to an interval of length d, and calculate the relationship between the variance of the frequency estimate and d. When d = e ε +1 has the smallest variance. In real application scenarios, the choice of d cannot be simply set to e. ε +1, in order to ensure that the collision results are not biased due to periodicity and other issues, usually the closest e ε The best result is obtained when the value is a prime number of +1.

[0153] After the anonymized data is sent to the data collector, the data collector 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, use the random seed res to encode each candidate value in the overall candidate value set candidates, and the encoding result of each candidate value is encode_candidate;

[0155] Secondly, after the dimension reduction operation is performed on each candidate value encoding result encode_candidate, it is compared with the user message report. If they are the same, the dimension of the frequency statistics vector freq corresponding to the candidate value is increased by one;

[0156] Again, after performing the above operation on all user answers, a 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 frequency estimation results corresponding to each candidate value.

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

[0159] Output: frequency statistics result freq

[0160] 1: Definition

[0161] 2: for report in reports do

[0162] Each user data is processed as follows

[0163] 3: report,random_seed←dataset

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

[0165] 4: for candidates in candidates do

[0166] All candidate values ​​in the candidate value set are processed 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] The encoded result is reduced in dimension and compared with the result message sent by the user. If they are the same, the frequency of the candidate value is increased by one.

[0171] 7: end for

[0172] 8: end for

[0173] 9:

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

[0175] 10: return freq

[0176] Assume that there are n data producers, and the data producers generate data candidate∈candidates locally, t candidateis the number of users whose real answers are candidates, c candidate The number of candidates whose answers are counted can be obtained according to the interference rule:

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

[0178] Therefore, the calibrated frequency estimate should be:

[0179] The present invention proposes an efficient, lightweight, and accurate privacy protection scheme for high-dimensional data that is adapted to respective scenarios, achieving good data statistical effects. The key points of the present invention are: 1. Proposing two lightweight data desensitization algorithms that comply with ε-differential privacy; 2. Proposing a two-round interactive privacy protection scheme for multi-value scenarios; 3. Proposing user grouping rules to avoid meaningless division of privacy budget; 4. Proposing a statistical item adjustment scheme.

[0180] The present invention realizes the adjustable functions of user grouping and statistical item set by introducing adjustable parameters to design a sampling optimization scheme, making the overall privacy budget allocation more reasonable. At the same time, the improved coding perturbation algorithm proposed by the present invention is introduced in the two-round interaction, which improves the accuracy of frequent item mining and obtains a more stable model performance while reducing the communication cost, achieving a balance between model performance and communication cost, and providing a new idea for frequency publishing in the field of multi-valued frequency statistics.

[0181] This paper proposes a local differential privacy histogram estimation algorithm (OSH) based on coding optimization. Different from the previous fixed matrix coding mapping scheme, this method randomly maps the data to a vertex of the hypercube, and replaces the random sampling method with a method of generating fixed sampling points based on local data, thereby reducing the communication cost while improving the accuracy of the numerical data statistical results. The algorithm uses a dimensionality reduction matrix to realize the mapping of high-dimensional numerical space to low-dimensional space, greatly reducing the communication cost of the algorithm, and optimizes the traditional random mapping matrix into a column orthogonal coding matrix, achieving dimensionality reduction while avoiding the problem of unstable model effect introduced by randomness.

[0182] The present invention proposes a local differential privacy histogram estimation algorithm (NCHOSH) based on non-encrypted hash coding optimization. Different from the previous coding algorithms that avoid collisions, the algorithm combines a fast hash algorithm and a local variable random seed to achieve the encoding of user data. The existence of the local variable random seed enables the algorithm to optimize the model performance by adjusting the dimension of the random seed when facing a high-dimensional candidate value set, and obtain better statistical effects at a lower communication cost. By optimizing the mapping rules between small cells, data dimensionality reduction and uniform collision between candidate values ​​are achieved, and better model effects are achieved at a low communication cost in the case of non-encrypted hash functions and dimensionality reduction. The algorithm can avoid the assumption that most algorithms need to know a fixed candidate value set in advance. Even if the candidate value set changes, the algorithm is still valid and the transmitted data will not fail, thereby improving the robustness of the algorithm, with low computational complexity and adjustable communication cost, ensuring the lightweight of the overall algorithm.

[0183] The present invention supports the processing of both numerical data and textual data, and supports a more comprehensive range of usage scenarios. It can be applied to a variety of data usage scenarios such as user behavior data statistics, such as GPS data, exercise duration and other numerical data statistics, and user identity data, such as address, gender, work unit and other textual information statistics.

[0184] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A two-round privacy protection method based on differential privacy, characterized in that: Includes the following: User grouping: The data generator randomly samples the local data, and the sampled user answers are processed according to the selected single-value frequency statistical method to achieve user grouping, selecting users who participated in the first and second stage privacy interactions and users who only participated in the second stage privacy interactions; The first stage of privacy interaction: the data generator selects candidates according to the sampling position in the user answer set i Extract a single user's answer value and combine it with the total privacy budget ε to generate the result message report i and sent to the data collector; The data collector reports the result messages returned by all users. i and the total privacy budget, generate frequency estimates of each candidate value, filter the candidate value set according to the candidate set adjustable parameter λ2 and the candidate value frequency estimate, and select high-frequency items from the candidate value set candidates as the frequent item candidate set in the second stage Send to the data generator end; 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 Perform filtering and filling, and according to the user grouping, interfere with the generation of user message reports and send them to the data collector; The data collector will group and count the user reports collected, and finally determine the frequent item sets of the user's answers. topk .

2. The two-round privacy protection method based on differential privacy according to claim 1 is characterized in that: The specific process of user grouping is as follows: (1) Set the threshold λ1 for a 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; (2) The selected user user i The answer set candidates i Truncate the set length (1+λ1)l-1), and fill the user answer set of insufficient length with virtual items; randomly sample sample_bit for the intercepted user answer data. If the sampled bit falls on the virtual item filling bit, the user will give up participating in the first stage of interaction and update the participation flag flag; finally output the retained sample bit sample_bit and participation flag flag.

3. The two-round privacy protection method based on differential privacy according to claim 1 is characterized in that: During the second stage of privacy interaction, the following operations are performed: Filtering: The data generator filters the candidate value set of frequent items based on the set of candidate values ​​returned by the data collector in the first stage. Filter the user answer set to remove non-high-frequency items; Filling: Set the length of the user's final answer set to the length of the high-frequency item candidate value set, and fill the user answer set that does not meet the above length; Privacy budget: The data producer calculates the privacy budget value available for each candidate value based on whether the user participates in the first phase of interaction; Frequent item set generation: Based on the privacy budget and the user's answer set, the user message report is generated and sent to the data collector.

4. The two-round privacy protection method based on differential privacy according to claim 3 is characterized in that: When a user participates in the first phase of interactive computing, his privacy budget is When the user does not participate in the first phase of interactive computing, his privacy budget is Where ε is the total privacy budget and k is the number of final results of the set two-round interaction.

5. The two-round privacy protection method based on differential privacy according to claim 3 is characterized in that: In the second stage of privacy interaction, the data collector sets the privacy budget according to whether the user participates in the first stage of privacy interaction. If the user participates in the first stage of execution, the total privacy budget is set to ε / 2. If the user does not participate in the first stage of execution, the total privacy budget is set to ε, and the user is divided into different groups according to the privacy budget. In each group, the data collector calculates the candidate set of frequent items. A new round of frequency statistics is performed with the user report set reports. The statistical process will filter out the virtual fill-in items. The data collector will finally determine the final top k frequent item sets estimations based on the frequency statistics of the two groups of users. topk .

6. The two-round privacy protection method based on differential privacy according to claim 1, characterized in that: When the user's answer set is of numeric type, the OSH algorithm is used to interfere with the generation of the result message; when the user's answer set is of text type, the NCHOSH algorithm is used to interfere with the generation of the result message.

7. The two-round privacy protection method based on differential privacy according to claim 6, characterized in that: The OSH algorithm is: First, the data collector generates a d×m random projection matrix M and ensures that the matrix is ​​a column orthogonal matrix with positive and negative elements. Send the matrix to the data producer; Secondly, the data producer generates a random code bit j from the interval (0, m-1), and uses the projection matrix M and the random code bit j to answer the user set candidates i Encode to generate a mapping vector d_vector representing the user input and a transmission bit i; 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 .

8. The two-round privacy protection method based on differential privacy according to claim 7 is characterized in that: When the OSH algorithm is used to generate interference, the data collector processes it as follows: First, the data collector initializes a statistical vector reports with all dimensions of data being 0, and reports the answer messages of each user collected. i and random code bit j, add report to the random code bit j dimension of the statistical vector i , 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 matrix columns, 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 and the decoding vector Me corresponding to each candidate value are combined. candidate Combined with the inverse mapping, the projection of the decoding vector onto the statistical vector is the frequency of the candidate value required.

9. The two-round privacy protection method based on differential privacy according to claim 6, characterized in that: The NCHOSH algorithm is: 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; Secondly, a random seed res of length l is generated, the user data is encoded using a hash algorithm, and the encoded result is reduced in dimension using a remainder operation; Again, the encoding result is disturbed, the probability of which remains unchanged with 1-p+q, and a number is randomly selected from 0 to d-1 with probability pq; Finally, the encoded result user message report and the random seed are sent to the data receiver.

10. The two-round privacy protection method based on differential privacy according to claim 9, characterized in that: When the NCHOSH algorithm is used to generate interference, the decoding process at the data collector end is as follows: First, for each user message report and random seed res, use the random seed res to encode each candidate value in the overall candidate value set candidates, 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, it is compared with the user message report. If they are the same, the dimension of the frequency statistics vector freq corresponding to the candidate value is increased by one; Again, after performing the above operation 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 to obtain the unbiased frequency estimation results corresponding to each candidate value.

Images