Robustness evaluation method and system for data stream local differential privacy algorithm

Abstract

The invention discloses a robustness evaluation method and system for a data stream local differential privacy algorithm. According to the method, the robustness of the algorithm is judged according to the proximity distance between algorithm release frequency distribution on each timestamp and algorithm robustness evaluation target distribution. According to the method, a branch judgment mechanism is adopted, and the strategy into which the data stream local differential privacy algorithm enters can be adaptively judged on each timestamp, so that the robustness evaluation of the algorithm can obtain a better evaluation effect. Based on the judgment of the branch judgment mechanism, the privacy algorithm can enter a strategy method expected by an evaluator through a dissimilarity calculation module influencing the data stream local differential privacy algorithm, and the method is divided into input evaluation and output evaluation. If the data stream local differential privacy algorithm enters the release strategy, the method provides a fine-grained algorithm robustness evaluation method for the frequency statistics local differential privacy algorithm in the release strategy, and the method is also divided into an input evaluation method and an output evaluation method.

Description

Technical Field

[0001] The present invention belongs to the field of data security, and in particular relates to a robustness evaluation method and system for a data stream local differential privacy algorithm. Background Art

[0002] As one of the specific technologies in the field of privacy computing, the data stream local differential privacy algorithm can safely and efficiently complete data statistics tasks such as frequency statistics without collecting original user data. When performing data statistics, the algorithm first collects user data at the current timestamp collected by the frequency statistics local differential privacy algorithm and calculates the dissimilarity with the published data at the previous timestamp. At the same time, the potential publishing error of the current timestamp is calculated as The algorithm will then determine and The relative size of Greater than , indicating that publishing data at this timestamp will bring smaller errors, and the algorithm will enter the publishing strategy and use the frequency statistics local differential privacy algorithm to publish data; otherwise, the algorithm will enter the approximate strategy and use the publishing frequency of the previous moment as the publishing frequency of the current timestamp. Some studies have discussed the robustness of local differential privacy algorithms and found that the collector of local differential privacy algorithms is very sensitive to the distribution of data sent by users. Some studies have discussed the robustness of frequency statistics local differential privacy algorithms. Some studies have also proposed local differential privacy algorithms for mean and variance statistics and discussed the robustness of the algorithm. Since the goal of its robustness evaluation algorithm can be adjusted according to the expectations of the evaluator, it is also called a fine-grained robustness evaluation algorithm. However, the existing robustness evaluation of local differential privacy algorithms is only applicable to evaluating static local differential privacy algorithms, not to evaluating data stream local differential privacy algorithms. Summary of the invention

[0003] The purpose of the present invention is to provide a robustness evaluation method and system for a data stream local differential privacy algorithm. The evaluation party of the method sends prepared data to a data stream local differential privacy algorithm server to change the publishing frequency of the data stream local differential privacy algorithm. Make it close to the evaluation target frequency set by the evaluator , that is, when the length is On the data stream of , solve the following optimization problem:

[0004] That is, the evaluator hopes to minimize the frequency of the algorithm's release at each timestamp. and the target frequency of evaluation set by the evaluator at each timestamp Approach distance At the same time, the evaluator evaluates the robustness of the local differential privacy algorithm of the data stream based on the proximity distance; The average close distance between the release frequency of the data stream local differential privacy algorithm and the evaluation target frequency on the data stream The larger it is, the more robust the algorithm is. This evaluation method can adaptively select the strategy for entering the local differential privacy algorithm of the data stream to achieve the above robustness evaluation goal.

[0005] In order to achieve the above object, the present invention adopts the following technical solution: A robustness evaluation method for a data stream local differential privacy algorithm, comprising: Step 1: Assessment preparation and selection of branching strategy; At each timestamp, the evaluator first sets the evaluation target frequency for that timestamp. And from the total Select from users users as evaluation users, The users are the remaining users; at the same time, the evaluator calculates the privacy budget used by the server to publish the policy at this timestamp from the evaluated users. And the number of users sampled by the server at that timestamp for publishing policies ,in is the number of users sampled at this timestamp for evaluating the release strategy. is the number of remaining users sampled for the release strategy at this timestamp; in order to achieve the goal of robustness evaluation, the evaluator will calculate the release frequency of the data stream local differential privacy algorithm entering the release strategy and the approximate strategy and the proximity distance of the evaluation target frequency, and determine which strategy the algorithm enters at this timestamp to minimize the proximity distance; Step 2: Evaluation of the dissimilarity calculation module of the local differential privacy algorithm for data streams; According to the judgment rule in step 1, the evaluator can determine which strategy the local differential privacy algorithm of the data stream can enter to minimize the proximity distance. If the approximate strategy can minimize the proximity distance, the evaluator will minimize the dissimilarity. , so that the algorithm enters the approximate strategy; otherwise, the evaluation side will maximize , so that the algorithm enters the release strategy; Step 3: Fine-grained algorithm robustness evaluation of the frequency statistics local differential privacy algorithm; According to step 2, maximize or minimize Based on the judgment, the evaluator will perform different operations to satisfy the fine-grained algorithm robustness evaluation, that is, change the release frequency of the data stream local differential privacy algorithm as much as possible to make it close to the evaluation target frequency; when the evaluator chooses to maximize the dissimilarity, the data stream local differential privacy algorithm will enter the release strategy and use the frequency statistics local differential privacy algorithm for frequency statistics and release; the evaluator needs to carry out a fine-grained algorithm robustness evaluation for the frequency statistics local differential privacy algorithm so that its release frequency distribution is close to the evaluation target frequency distribution; otherwise, when the evaluator chooses to minimize the dissimilarity, the data stream local differential privacy algorithm will enter the approximate strategy and directly use the release frequency of the previous timestamp as the release frequency of this timestamp, so the evaluator does not need to perform additional operations.

[0006] A further improvement of the present invention is that in step 1, the evaluator determines which strategy the algorithm enters at the timestamp based on the calculated release frequency of the data stream local differential privacy algorithm entering the release strategy or approximate strategy and the proximity distance of the evaluation target frequency, including: Step 1.1: Calculate the distance of the data stream local differential privacy algorithm into the approximate strategy: ,in, is the publishing frequency of the previous timestamp, is the target frequency for evaluation at that timestamp, The size of the field for user input; Step 1.2: Calculate the average distance of the data stream local differential privacy algorithm entering the release strategy: ,in, It represents the algorithm release frequency after the fine-grained algorithm robustness evaluation of the local differential privacy algorithm based on the frequency statistics of the release strategy; Step 1.3: Compare the proximity distance of the local differential privacy algorithm for controlling data flow into the approximate strategy and proximity to entry release strategy ,if If is smaller, the evaluator hopes that the data stream local differential privacy algorithm enters the approximation strategy; otherwise, the evaluator hopes that the data stream local differential privacy algorithm enters the release strategy.

[0007] A further improvement of the present invention is that the evaluation method for the data stream local differential privacy algorithm dissimilarity calculation module in step 2 is divided into two methods, including input evaluation and output evaluation.

[0008] A further improvement of the present invention is that input evaluation refers to designing input data for evaluating the user's local differential privacy algorithm to affect the calculation mechanism of dissimilarity to evaluate the robustness of the algorithm; assuming that there are Users are used to calculate the dissimilarity of the data stream local differential privacy algorithm, where Named Evaluation User, It is called residual user. The input distribution of the residual user’s local differential privacy algorithm is , among which Item is , then maximizing or minimizing the dissimilarity is to solve the following optimization problem:

[0009]

[0010] in, is the input frequency distribution of the local differential privacy algorithm, and the evaluator maximizes or minimizes it by influencing the frequency distribution The goal of this frequency distribution is item Calculated as ,in Indicates that there are evaluation users The input value of the local differential privacy algorithm is .

[0011] A further improvement of the present invention is that the output evaluation refers to the evaluator bypassing the local perturbation step of the local differential privacy algorithm and directly sending the calculated data to the server to affect the calculation of the dissimilarity, so as to achieve the purpose of verifying the robustness of the algorithm; minimizing the dissimilarity, that is, minimizing When the evaluator minimizes its upper bound Implementation, where is the release frequency distribution of the local differential privacy algorithm:

[0012] in are the parameters of the local differential privacy algorithm, Indicates that there is The value sent to the server by the user is ; When maximizing the dissimilarity, the evaluator first calculates , maximizing the dissimilarity is to make , , that is, use all evaluation users to send to the server This can achieve the goal of maximizing the dissimilarity.

[0013] A further improvement of the present invention is that the fine-grained algorithm robustness evaluation method for the frequency statistical local differential privacy algorithm in step 3 is divided into two types, input evaluation and output evaluation, and both meet the "privacy security consistency"; Input evaluation is designed to evaluate the frequency distribution of user inputs to achieve the goal of fine-grained algorithm robustness evaluation. The publishing strategy of the local differential privacy algorithm for data streams is used by users, where Named Evaluation User, It is called the residual user. The input distribution of the local differential privacy algorithm for the residual user at this timestamp is , among which Item is , then completing the fine-grained algorithm robustness evaluation at this timestamp is to solve the following optimization problem:

[0014]

[0015] in, is the input frequency distribution of the local differential privacy algorithm. The item is calculated as ,in Indicates that there are evaluation users The input value of the local differential privacy algorithm is ; At the same time, calculate the proximity distance of the input evaluation when the number of evaluation users is large enough ,in, Indicates that the privacy budget used is The frequency statistics local differential privacy algorithm is The error of frequency statistics of users; thus, The larger the size or the more users there are, the closer the distance The smaller it is, the more it meets the goal of robustness evaluation, that is, the evaluation shows "privacy and security consistency"; In the output evaluation, the user is evaluated to bypass the local perturbation of the local differential privacy algorithm and send data directly to the server to achieve the effect of fine-grained algorithm robustness evaluation; in order to make the publishing frequency close to the target frequency, that is, to minimize , the evaluator minimizes its upper bound Implementation, where is the release frequency distribution of the local differential privacy algorithm:

[0016] in are the parameters of the local differential privacy algorithm, Indicates that there is The value sent to the server by the user is ; When there are enough evaluation users, the evaluation proximity distance is output ,in is a constant ;akin, The larger the size or the more total users, The smaller it is, the more it meets the goal of robustness evaluation, that is, the evaluation shows "privacy-security consistency".

[0017] A robustness evaluation system for a data stream local differential privacy algorithm, comprising: Selection module, used for evaluation preparation and selection of branching strategies; At each timestamp, the evaluator first sets the evaluation target frequency for that timestamp. And from the total Select from users users as evaluation users, The users are the remaining users; at the same time, the evaluator calculates the privacy budget used by the server to publish the policy at this timestamp from the evaluated users. And the number of users sampled by the server at that timestamp for publishing policies ,in is the number of users sampled at this timestamp for evaluating the release strategy. is the number of remaining users sampled for the release strategy at this timestamp; in order to achieve the goal of robustness evaluation, the evaluator will calculate the release frequency of the data stream local differential privacy algorithm entering the release strategy and the approximate strategy and the proximity distance of the evaluation target frequency, and determine which strategy the algorithm enters at this timestamp to minimize the proximity distance; The first evaluation module is used to evaluate the dissimilarity calculation module of the local differential privacy algorithm for the data stream; According to the judgment rules of the selection module, the evaluator can determine which strategy the local differential privacy algorithm of the data stream enters to minimize the proximity distance. If the approximate strategy can minimize the proximity distance, the evaluator will minimize the dissimilarity. , so that the algorithm enters the approximate strategy; otherwise, the evaluation side will maximize , so that the algorithm enters the release strategy; The second evaluation module is a fine-grained algorithm robustness evaluation of the frequency statistics local differential privacy algorithm; According to the first evaluation module, the maximum or minimum Based on the judgment, the evaluator will perform different operations to satisfy the fine-grained algorithm robustness evaluation, that is, change the release frequency of the data stream local differential privacy algorithm as much as possible to make it close to the evaluation target frequency; when the evaluator chooses to maximize the dissimilarity, the data stream local differential privacy algorithm will enter the release strategy and use the frequency statistics local differential privacy algorithm for frequency statistics and release; the evaluator needs to carry out a fine-grained algorithm robustness evaluation for the frequency statistics local differential privacy algorithm so that its release frequency distribution is close to the evaluation target frequency distribution; otherwise, when the evaluator chooses to minimize the dissimilarity, the data stream local differential privacy algorithm will enter the approximate strategy and directly use the release frequency of the previous timestamp as the release frequency of this timestamp, so the evaluator does not need to perform additional operations.

[0018] Compared with the prior art, the present invention has at least the following beneficial technical effects: The present invention provides a robustness evaluation method and system for a data stream local differential privacy algorithm. The method first provides a branch judgment mechanism, which can adaptively judge at each timestamp which strategy (publishing strategy or approximate strategy) the data stream local differential privacy algorithm enters to minimize the proximity distance of the algorithm robustness evaluation on the overall data stream and thus achieve the goal of algorithm robustness evaluation. The branch judgment mechanism utilizes the "privacy security consistency" of the fine-grained algorithm robustness evaluation method. Furthermore, the evaluator maximizes or minimizes , thereby controlling the data stream local differential privacy algorithm to enter the strategy expected by the evaluator for data statistics and release. There are two evaluation methods for the dissimilarity calculation mechanism of the data stream local differential privacy algorithm: output evaluation and input evaluation.

[0019] Furthermore, when the data stream local differential privacy algorithm enters the release strategy to release data, this method can influence the frequency statistics local differential privacy algorithm in the release strategy so that its release frequency distribution is close to the target frequency distribution expected by the evaluator. According to the closeness of the two distributions, the evaluator can judge the robustness of the algorithm. It can be proved that the fine-grained algorithm robustness evaluation method for the frequency statistics local differential privacy algorithm meets the "privacy security consistency". The fine-grained algorithm robustness evaluation method for the frequency statistics local differential privacy algorithm is also divided into input evaluation and output evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the process of the present invention.

[0021] Figure 2 It is a schematic diagram of the logical architecture of the present invention.

[0022] Figure 3 This is a trend diagram of the approach distance of the present invention changing with the total privacy budget.

[0023] Figure 4 It is a structural block diagram of the system of the present invention. DETAILED DESCRIPTION

[0024] In the following, only some exemplary embodiments are briefly described. As those skilled in the art will appreciate, the described embodiments may be modified in various ways without departing from the spirit or scope of the present invention. Therefore, the drawings and descriptions are considered to be exemplary and non-restrictive in nature.

[0025] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0026] It should also be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0027] It should be further understood that the term "and / or" used in the present description and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0028] Various structural schematic diagrams of the embodiments disclosed in the present invention are shown in the accompanying drawings. These figures are not drawn to scale, and some details are magnified and some details may be omitted for the purpose of clear expression. The shapes of various regions and layers shown in the figures and the relative sizes and positional relationships therebetween are only exemplary, and may deviate in practice due to manufacturing tolerances or technical limitations, and those skilled in the art may additionally design regions / layers with different shapes, sizes, and relative positions according to actual needs.

[0029] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0030] Example 1 Reference Figure 1 , this embodiment provides a method for evaluating the robustness of a data stream local differential privacy algorithm, which specifically includes the following steps: Step 1: Evaluation preparation and branch strategy selection: At each timestamp, the evaluator first sets the evaluation target frequency for that timestamp. And from the total Select from users users as evaluation users, The users are the remaining users; at the same time, the evaluator calculates the privacy budget used by the server to publish the policy at this timestamp from the evaluated users. And the number of users sampled by the server at that timestamp for publishing policies ,in is the number of users sampled at this timestamp for evaluating the release strategy. is the number of remaining users sampled for the release strategy at this timestamp; in order to achieve the goal of robustness evaluation, the evaluator will calculate the release frequency of the data stream local differential privacy algorithm entering the release strategy and the approximate strategy and the proximity distance of the evaluation target frequency, and determine which strategy the algorithm enters at this timestamp to minimize the proximity distance, specifically following the following steps; Step 1.1: Calculate the distance of the data stream local differential privacy algorithm into the approximate strategy: ,in, is the publishing frequency of the previous timestamp, is the target frequency for evaluation at that timestamp, The size of the field for user input; Step 1.2: Calculate the average distance of the data stream local differential privacy algorithm entering the release strategy: ,in, It represents the algorithm release frequency after the fine-grained algorithm robustness evaluation of the local differential privacy algorithm based on the frequency statistics of the release strategy; Step 1.3: Compare the proximity distance of the local differential privacy algorithm for controlling data flow into the approximate strategy and proximity to entry release strategy ,if If is smaller, the evaluator hopes that the data stream local differential privacy algorithm enters the approximation strategy; otherwise, the evaluator hopes that the data stream local differential privacy algorithm enters the release strategy.

[0031] Step 2: Evaluate the dissimilarity calculation module of the local differential privacy algorithm for data streams; according to the above judgment rules, the evaluator can determine which strategy the local differential privacy algorithm of the data stream enters to minimize the proximity distance for the timestamp. If the approximate strategy can minimize the proximity distance, the evaluator will minimize the dissimilarity. , so that the algorithm enters the approximate strategy; otherwise, the evaluation side will maximize , so that the algorithm enters the release strategy. There are two evaluation methods for the dissimilarity calculation module of the local differential privacy algorithm for data streams: input evaluation and output evaluation.

[0032] Input evaluation refers to designing and evaluating the input data of the user's local differential privacy algorithm to affect the calculation mechanism of dissimilarity to evaluate the robustness of the algorithm. Users are used to calculate the dissimilarity of the data stream local differential privacy algorithm, where Named Evaluation User, It is called residual user. The input distribution of the residual user’s local differential privacy algorithm is , among which Item is , then maximizing or minimizing the dissimilarity is to solve the following optimization problem:

[0033]

[0034] in, is the input frequency distribution of the local differential privacy algorithm. The evaluator can maximize or minimize the frequency distribution by influencing it. The goal of this frequency distribution is item It can be calculated as ,in Indicates that there are evaluation users The input value of the local differential privacy algorithm is .

[0035] Output evaluation means that the evaluator bypasses the local perturbation step of the local differential privacy algorithm and directly sends the calculated data to the server to affect the calculation of dissimilarity, so as to achieve the purpose of testing the robustness of the algorithm. When Implementation, where is the release frequency distribution of the local differential privacy algorithm:

[0036] in are the parameters of the local differential privacy algorithm, Indicates that there is The value sent to the server by the user is When maximizing the dissimilarity, the evaluator first calculates , maximizing the dissimilarity is to make , , that is, use all evaluation users to send to the server This can achieve the goal of maximizing the dissimilarity.

[0037] Step 3: Fine-grained algorithm robustness evaluation method for frequency statistical local differential privacy algorithm; according to the previous step, maximize or minimize The evaluator will perform different operations to satisfy the fine-grained algorithm robustness evaluation, that is, change the release frequency of the data stream local differential privacy algorithm as much as possible to make it close to the evaluation target frequency; when the evaluator chooses to maximize the dissimilarity, the data stream local differential privacy algorithm will enter the release strategy and use the frequency statistics local differential privacy algorithm for frequency statistics and release; the evaluator needs to carry out a fine-grained algorithm robustness evaluation for the frequency statistics local differential privacy algorithm so that its release frequency distribution is close to the evaluation target frequency distribution; otherwise, when the evaluator chooses to minimize the dissimilarity, the data stream local differential privacy algorithm will enter the approximate strategy and directly use the release frequency of the previous timestamp as the release frequency of this timestamp, so the evaluator does not need to perform additional operations. The fine-grained algorithm robustness evaluation method for the frequency statistics local differential privacy algorithm is also divided into two types, input evaluation and output evaluation.

[0038] Input evaluation is designed to evaluate the frequency distribution of user inputs to achieve the goal of fine-grained algorithm robustness evaluation. The publishing strategy of the local differential privacy algorithm for data streams is used by users, where Named Evaluation User, It is called the residual user. The input distribution of the local differential privacy algorithm for the residual user at this timestamp is , among which Item is , then completing the fine-grained algorithm robustness evaluation at this timestamp is to solve the following optimization problem:

[0039]

[0040] in, is the input frequency distribution of the local differential privacy algorithm. The term can be calculated as ,in Indicates that there are evaluation users The input value of the local differential privacy algorithm is At the same time, the proximity distance of the input evaluation can be calculated when the number of evaluation users is large enough ,in, Indicates that the privacy budget used is The frequency statistics local differential privacy algorithm is The error of frequency statistics of users. The larger the size or the more users there are, the closer the distance The smaller it is, the more it meets the goal of robustness evaluation, that is, the evaluation shows "privacy-security consistency".

[0041] In the output evaluation, the user is evaluated to bypass the local perturbation of the local differential privacy algorithm and send data directly to the server to achieve the effect of fine-grained algorithm robustness evaluation. In order to make the publishing frequency close to the target frequency, that is, to minimize , the evaluator can minimize its upper bound Implementation, where is the release frequency distribution of the local differential privacy algorithm:

[0042] in are the parameters of the local differential privacy algorithm, Indicates that there is The value sent to the server by the user is It can be proved that when there are enough evaluation users, the output evaluation is close to ,in is a constant. Similarly, The larger the size or the more users there are, the closer the distance The smaller it is, the more it meets the goal of robustness evaluation, that is, the evaluation shows "privacy-security consistency".

[0043] In the description of the present invention, it is necessary to understand that the term "timestamp" refers to a time point mark taken at a certain time interval in the time stream, which is used to represent and record the changes that occur within a period of time after the mark, "data provider" refers to the person or enterprise organization that provides data, and "local differential privacy" refers to the probability that two data sets with only one inconsistent data obtain similar calculation results after performing a specified calculation. "Robustness evaluation method" refers to the evaluation party judging the robustness of the algorithm by the proximity of the algorithm release frequency distribution at each timestamp to the algorithm robustness evaluation target distribution.

[0044] In order to test the impact of different privacy budgets on the performance of this evaluation method, several algorithm robustness evaluation methods for data stream local differential privacy algorithms are compared. A data set of 700,000 data providers, 432 timestamps, and 150 discrete values ​​is used for experiments. The sliding window of the data stream local differential privacy algorithm is set to 20, and the evaluation target distribution is uniformly distributed on each timestamp. The mean square error is used to calculate the proximity between the release frequency and the evaluation target frequency. The smaller the mean square error, the worse the algorithm robustness. The results are as follows: Figure 3As shown in the figure, it can be seen that as the total privacy budget of the data stream local differential privacy algorithm increases, the mean square error of all evaluations decreases, and the mean square error of this method (input evaluation IAE, output evaluation OAE) is lower than that of other methods (input evaluation IUE and ISE, output evaluation OUE and OSE), which can test the robustness of the privacy algorithm more deeply. At the same time, compared with the input evaluation, the output evaluation has a smaller mean square error, showing better performance.

[0045] refer to Figure 2 , the present invention provides a method for robustness evaluation of a data stream local differential privacy algorithm, and its logical method includes two stages: evaluating the dissimilarity calculation module of the data stream local differential privacy algorithm and evaluating the release strategy of the data stream local differential privacy algorithm. In order to improve the evaluation effect on the overall data stream, the algorithm will first determine which strategy the data stream local differential privacy algorithm enters to minimize the approach distance to achieve the purpose of robustness evaluation before conducting an evaluation on the dissimilarity calculation module. It is worth noting that the evaluator cannot directly control the decision selection mechanism of the data stream local differential privacy algorithm to make it enter its desired strategy, but can make the algorithm enter its desired decision by affecting the size of the dissimilarity, thereby performing a more accurate algorithm robustness evaluation.

[0046] Example 2 Reference Figure 4 , this embodiment provides a robustness evaluation system for a data stream local differential privacy algorithm, including: Selection module, used for evaluation preparation and selection of branching strategies; At each timestamp, the evaluator first sets the evaluation target frequency for that timestamp. And from the total Select from users users as evaluation users, The users are the remaining users; at the same time, the evaluator calculates the privacy budget used by the server to publish the policy at this timestamp from the evaluated users. And the number of users sampled by the server at that timestamp for publishing policies ,in is the number of users sampled at this timestamp for evaluating the release strategy. is the number of remaining users sampled for the release strategy at this timestamp; the evaluator calculates the release frequency of the data stream local differential privacy algorithm entering the release strategy and the approximate strategy and the proximity distance of the evaluation target frequency, and determines which strategy the algorithm enters at this timestamp to minimize the proximity distance; The first evaluation module is used to evaluate the dissimilarity calculation module of the local differential privacy algorithm for the data stream; According to the judgment rules of the selection module, the evaluator can determine which strategy the local differential privacy algorithm of the data stream enters to minimize the proximity distance. If the approximate strategy can minimize the proximity distance, the evaluator will minimize the dissimilarity. , so that the algorithm enters the approximate strategy; otherwise, the evaluation side will maximize , so that the algorithm enters the release strategy; The second evaluation module is a fine-grained algorithm robustness evaluation of the frequency statistics local differential privacy algorithm; According to the first evaluation module, the maximum or minimum Based on the judgment, the evaluator will perform different operations to satisfy the fine-grained algorithm robustness evaluation, that is, change the release frequency of the data stream local differential privacy algorithm as much as possible to make it close to the evaluation target frequency; when the evaluator chooses to maximize the dissimilarity, the data stream local differential privacy algorithm will enter the release strategy and use the frequency statistics local differential privacy algorithm for frequency statistics and release; the evaluator needs to carry out a fine-grained algorithm robustness evaluation for the frequency statistics local differential privacy algorithm so that its release frequency distribution is close to the evaluation target frequency distribution; otherwise, when the evaluator chooses to minimize the dissimilarity, the data stream local differential privacy algorithm will enter the approximate strategy and directly use the release frequency of the previous timestamp as the release frequency of this timestamp, so the evaluator does not need to perform additional operations.

[0047] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention. Therefore, no matter from which point of view, the embodiments should be regarded as exemplary and non-restrictive. The scope of the present invention is defined by the attached claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention. Any figure mark in the claims should not be regarded as limiting the claims involved.

[0048] In addition, it should be understood that although this specification is described in accordance with the implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation modes that can be understood by those skilled in the art. The above content is only to illustrate the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution according to the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A robustness evaluation method for a data stream local differential privacy algorithm, characterized in that: include: Step 1: Assessment preparation and selection of branching strategy; At each timestamp, the evaluator first sets the evaluation target frequency for that timestamp. And from the total Select from users users as evaluation users, The users are the remaining users; at the same time, the evaluator calculates the privacy budget used by the server to publish the policy at this timestamp from the evaluated users. And the number of users sampled by the server at that timestamp for publishing policies ,in is the number of users sampled at this timestamp for evaluating the release strategy. is the number of remaining users sampled for the release strategy at this timestamp; in order to achieve the goal of robustness evaluation, the evaluator will calculate the release frequency of the data stream local differential privacy algorithm entering the release strategy and the approximate strategy and the proximity distance of the evaluation target frequency, and determine which strategy the algorithm enters at this timestamp to minimize the proximity distance; Step 2: Evaluation of the dissimilarity calculation module of the local differential privacy algorithm for data streams; According to the judgment rule in step 1, the evaluator can determine which strategy the local differential privacy algorithm of the data stream can enter to minimize the proximity distance. If the approximate strategy can minimize the proximity distance, the evaluator will minimize the dissimilarity. , so that the algorithm enters the approximate strategy; otherwise, the evaluation side will maximize , so that the algorithm enters the release strategy; Step 3: Fine-grained algorithm robustness evaluation of the frequency statistics local differential privacy algorithm; According to step 2, maximize or minimize Based on the judgment, the evaluator will perform different operations to satisfy the fine-grained algorithm robustness evaluation, that is, change the release frequency of the data stream local differential privacy algorithm as much as possible to make it close to the evaluation target frequency; when the evaluator chooses to maximize the dissimilarity, the data stream local differential privacy algorithm will enter the release strategy and use the frequency statistics local differential privacy algorithm for frequency statistics and release; the evaluator needs to carry out a fine-grained algorithm robustness evaluation for the frequency statistics local differential privacy algorithm so that its release frequency distribution is close to the evaluation target frequency distribution; otherwise, when the evaluator chooses to minimize the dissimilarity, the data stream local differential privacy algorithm will enter the approximate strategy and directly use the release frequency of the previous timestamp as the release frequency of this timestamp, so the evaluator does not need to perform additional operations.

2. According to claim 1, a robustness evaluation method for a data stream local differential privacy algorithm is characterized in that: In step 1, the evaluator determines which strategy the algorithm enters at the timestamp based on the calculated release frequency of the data stream local differential privacy algorithm entering the release strategy or approximate strategy and the proximity of the evaluation target frequency, including: Step 1.1: Calculate the distance of the data stream local differential privacy algorithm into the approximate strategy: ,in, is the publishing frequency of the previous timestamp, is the target frequency for evaluation at that timestamp, The size of the field for user input; Step 1.2: Calculate the average distance of the data stream local differential privacy algorithm entering the release strategy: ,in, It represents the algorithm release frequency after the fine-grained algorithm robustness evaluation of the local differential privacy algorithm based on the frequency statistics of the release strategy; Step 1.3: Compare the proximity distance of the local differential privacy algorithm for controlling data flow into the approximate strategy and proximity to entry release strategy ,if If is smaller, the evaluator hopes that the data stream local differential privacy algorithm enters the approximation strategy; otherwise, the evaluator hopes that the data stream local differential privacy algorithm enters the release strategy.

3. The robustness evaluation method of a data stream local differential privacy algorithm according to claim 2 is characterized in that: There are two evaluation methods for the data stream local differential privacy algorithm dissimilarity calculation module in step 2, including input evaluation and output evaluation.

4. The robustness evaluation method of a data stream local differential privacy algorithm according to claim 3 is characterized in that: Input evaluation refers to the design of input data for evaluating the user's local differential privacy algorithm to affect the calculation mechanism of dissimilarity to evaluate the robustness of the algorithm; assuming that there are Users are used to calculate the dissimilarity of the data stream local differential privacy algorithm, where Named Evaluation User, It is called residual user. The input distribution of the residual user’s local differential privacy algorithm is , among which Item is , then maximizing or minimizing the dissimilarity is to solve the following optimization problem: in, is the input frequency distribution of the local differential privacy algorithm, and the evaluator maximizes or minimizes it by influencing the frequency distribution The goal of this frequency distribution is item Calculated as ,in Indicates that there are evaluation users The input value of the local differential privacy algorithm is .

5. The robustness evaluation method of a data stream local differential privacy algorithm according to claim 4 is characterized in that: Output evaluation means that the evaluator bypasses the local perturbation step of the local differential privacy algorithm and directly sends the calculated data to the server to affect the calculation of dissimilarity, so as to achieve the purpose of testing the robustness of the algorithm; Minimize the dissimilarity, that is, minimize When the evaluator minimizes its upper bound Implementation, where is the release frequency distribution of the local differential privacy algorithm: in are the parameters of the local differential privacy algorithm, Indicates that there is The value sent to the server by the user is ; When maximizing the dissimilarity, the evaluator first calculates , maximizing the dissimilarity is to make , , that is, use all evaluation users to send to the server This can achieve the goal of maximizing the dissimilarity.

6. The robustness evaluation method of a data stream local differential privacy algorithm according to claim 5 is characterized in that: There are two types of fine-grained algorithm robustness evaluation methods for the frequency statistics local differential privacy algorithm in step 3: input evaluation and output evaluation, and both meet the "privacy security consistency" requirement. Input evaluation is designed to evaluate the frequency distribution of user inputs to achieve the goal of fine-grained algorithm robustness evaluation. The publishing strategy of the local differential privacy algorithm for data streams is used by users, where Named Evaluation User, It is called the residual user. The input distribution of the local differential privacy algorithm for the residual user at this timestamp is , among which Item is , then completing the fine-grained algorithm robustness evaluation at this timestamp is to solve the following optimization problem: in, is the input frequency distribution of the local differential privacy algorithm. The item is calculated as ,in Indicates that there are evaluation users The input value of the local differential privacy algorithm is ; At the same time, calculate the proximity distance of the input evaluation when the number of evaluation users is large enough ,in, Indicates that the privacy budget used is The frequency statistics local differential privacy algorithm is The error of frequency statistics of users; thus, The larger the size or the more users there are, the closer the distance The smaller it is, the more it meets the goal of robustness evaluation, that is, the evaluation shows "privacy and security consistency"; In the output evaluation, the user is evaluated to bypass the local perturbation of the local differential privacy algorithm and send data directly to the server to achieve the effect of fine-grained algorithm robustness evaluation; in order to make the publishing frequency close to the target frequency, that is, to minimize , the evaluator minimizes its upper bound Implementation, where is the release frequency distribution of the local differential privacy algorithm: in are the parameters of the local differential privacy algorithm, Indicates that there is The value sent to the server by the user is ; When there are enough evaluation users, the evaluation proximity distance is output ,in is a constant ;akin, The larger the size or the more users there are, the closer the distance The smaller it is, the more it meets the goal of robustness evaluation, that is, the evaluation shows "privacy-security consistency".

7. A robustness evaluation system for a data stream local differential privacy algorithm, characterized in that: include: Selection module, used for evaluation preparation and selection of branching strategies; At each timestamp, the evaluator first sets the evaluation target frequency for that timestamp. And from the total Select from users users as evaluation users, The users are the remaining users; at the same time, the evaluator calculates the privacy budget used by the server to publish the policy at this timestamp from the evaluated users. And the number of users sampled by the server at that timestamp for publishing policies ,in is the number of users sampled at this timestamp for evaluating the release strategy. is the number of remaining users sampled for the release strategy at this timestamp; in order to achieve the goal of robustness evaluation, the evaluator will calculate the release frequency of the data stream local differential privacy algorithm entering the release strategy and the approximate strategy and the proximity distance of the evaluation target frequency, and determine which strategy the algorithm enters at this timestamp to minimize the proximity distance; The first evaluation module is used to evaluate the dissimilarity calculation module of the local differential privacy algorithm for the data stream; According to the judgment rules of the selection module, the evaluator can determine which strategy the local differential privacy algorithm of the data stream enters to minimize the proximity distance. If the approximate strategy can minimize the proximity distance, the evaluator will minimize the dissimilarity. , so that the algorithm enters the approximate strategy; otherwise, the evaluation side will maximize , so that the algorithm enters the release strategy; The second evaluation module is a fine-grained algorithm robustness evaluation of the frequency statistics local differential privacy algorithm; According to the first evaluation module, the maximum or minimum Based on the judgment, the evaluator will perform different operations to satisfy the fine-grained algorithm robustness evaluation, that is, change the release frequency of the data stream local differential privacy algorithm as much as possible to make it close to the evaluation target frequency; when the evaluator chooses to maximize the dissimilarity, the data stream local differential privacy algorithm will enter the release strategy and use the frequency statistics local differential privacy algorithm for frequency statistics and release; the evaluator needs to carry out a fine-grained algorithm robustness evaluation for the frequency statistics local differential privacy algorithm so that its release frequency distribution is close to the evaluation target frequency distribution; otherwise, when the evaluator chooses to minimize the dissimilarity, the data stream local differential privacy algorithm will enter the approximate strategy and directly use the release frequency of the previous timestamp as the release frequency of this timestamp, so the evaluator does not need to perform additional operations.

8. The robustness evaluation system of a data stream local differential privacy algorithm according to claim 7, characterized in that: In the selection module, the evaluator determines which strategy the algorithm enters at the timestamp based on the calculated frequency of the release strategy or approximate strategy entered by the local differential privacy algorithm of the data stream and the proximity of the evaluation target frequency, including: The distance between the local differential privacy algorithm of the data stream and the approximate strategy is calculated as ,in, is the publishing frequency of the previous timestamp, is the target frequency for evaluation at that timestamp, The size of the field for user input; The average distance between the local differential privacy algorithm of the data stream and the release strategy is calculated as ,in, It represents the algorithm release frequency after the fine-grained algorithm robustness evaluation of the local differential privacy algorithm based on the frequency statistics of the release strategy; Comparison of the control data flow local differential privacy algorithm into the proximity distance of the approximate strategy and proximity to entry release strategy ,if If is smaller, the evaluator hopes that the data stream local differential privacy algorithm enters the approximation strategy; otherwise, the evaluator hopes that the data stream local differential privacy algorithm enters the release strategy.

9. The robustness evaluation system of a data stream local differential privacy algorithm according to claim 8, characterized in that: There are two evaluation methods for the dissimilarity calculation module of the data stream local differential privacy algorithm in the first evaluation module, including input evaluation and output evaluation; Input evaluation refers to the design of input data for evaluating the user's local differential privacy algorithm to affect the calculation mechanism of dissimilarity to evaluate the robustness of the algorithm; assuming that there are Users are used to calculate the dissimilarity of the data stream local differential privacy algorithm, where Named Evaluation User, It is called residual user. The input distribution of the residual user’s local differential privacy algorithm is , among which Item is , then maximizing or minimizing the dissimilarity is to solve the following optimization problem: in, is the input frequency distribution of the local differential privacy algorithm, and the evaluator maximizes or minimizes it by influencing the frequency distribution The goal of this frequency distribution is item Calculated as ,in Indicates that there are evaluation users The input value of the local differential privacy algorithm is ; Output evaluation means that the evaluator bypasses the local perturbation step of the local differential privacy algorithm and directly sends the calculated data to the server to affect the calculation of dissimilarity, so as to achieve the purpose of testing the robustness of the algorithm; minimizing dissimilarity, that is, minimizing When the evaluator minimizes its upper bound Implementation, where is the release frequency distribution of the local differential privacy algorithm: in are the parameters of the local differential privacy algorithm, Indicates that there is The value sent to the server by the user is ; When maximizing the dissimilarity, the evaluator first calculates , maximizing the dissimilarity is to make , , that is, use all evaluation users to send to the server This can achieve the goal of maximizing the dissimilarity.

10. The robustness evaluation system of a data stream local differential privacy algorithm according to claim 9, characterized in that: The fine-grained algorithm robustness evaluation methods for the frequency statistics local differential privacy algorithm in the second evaluation module are divided into two types: input evaluation and output evaluation, and both meet the "privacy security consistency" requirement; Input evaluation is designed to evaluate the frequency distribution of user inputs to achieve the goal of fine-grained algorithm robustness evaluation. The publishing strategy of the local differential privacy algorithm for data streams is used by users, where Named Evaluation User, It is called the residual user. The input distribution of the local differential privacy algorithm for the residual user at this timestamp is , among which Item is , then completing the fine-grained algorithm robustness evaluation at this timestamp is to solve the following optimization problem: in, is the input frequency distribution of the local differential privacy algorithm. The item is calculated as ,in Indicates that there are evaluation users The input value of the local differential privacy algorithm is ; At the same time, calculate the proximity distance of the input evaluation when the number of evaluation users is large enough ,in, Indicates that the privacy budget used is The frequency statistics local differential privacy algorithm is The error of frequency statistics of users; thus, The larger the size or the more users there are, the closer the distance The smaller it is, the more it meets the goal of robustness evaluation, that is, the evaluation shows "privacy and security consistency"; In the output evaluation, the user is evaluated to bypass the local perturbation of the local differential privacy algorithm and send data directly to the server to achieve the effect of fine-grained algorithm robustness evaluation; in order to make the publishing frequency close to the target frequency, that is, to minimize , the evaluator minimizes its upper bound Implementation, where is the release frequency distribution of the local differential privacy algorithm: in are the parameters of the local differential privacy algorithm, Indicates that there is The value sent to the server by the user is ; When there are enough evaluation users, the evaluation proximity distance is output ,in is a constant ;akin, The larger the size or the more users there are, the closer the distance The smaller it is, the more it meets the goal of robustness evaluation, that is, the evaluation shows "privacy-security consistency".

Images