Privacy processing method and device for high-dimensional data

Abstract

The embodiment of the invention discloses a privacy processing method and device for high-dimensional data, and relates to the technical field of Internet data analysis. The method comprises the following steps: collecting high-dimensional data generated by each participant; based on the information entropy of the high-dimensional data, distributing a corresponding privacy budget for the features of the high-dimensional data; according to the privacy budget obtained by each feature, noise is added to each piece of high-dimensional data for data disturbance; according to the disturbed high-dimensional data, frequency estimation is carried out on discrete features; and according to the disturbed high-dimensional data, training a deep learning model to obtain a target deep learning model, the target deep learning model being used for completing a preset task. According to the method, the privacy budget can be effectively distributed, and the high-dimensional data is reasonably disturbed, so that the effectiveness of the noise-added high-dimensional data can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of Internet data analysis, and particularly relates to a privacy processing method and device based on high-dimensional data. BACKGROUND

[0002] With the rapid development of information technology and big data analysis technology, more and more user sensitive data are frequently collected and analyzed, such as personal health information, financial records and personal behavior data related to user privacy. Improper collection and use of these data lead to frequent data leakage and privacy infringement, causing serious user privacy security problems. Therefore, how to effectively protect personal privacy while efficiently analyzing user data has become an urgent security problem to be solved.

[0003] Differential privacy is an advanced privacy protection technology at present, which realizes privacy protection by adding noise to data. However, compared with low-dimensional data, high-dimensional data has too many attributes, and more noise needs to be added to achieve the same privacy protection effect as low-dimensional data. Although increasing noise can improve the privacy protection level, it will reduce the utility of data. SUMMARY

[0004] The present application aims to solve the problems existing in the prior art, and provides a privacy processing method and device for high-dimensional data.

[0005] The privacy processing method for high-dimensional data provided by the present application is realized by the following technical scheme: collecting high-dimensional data generated by each participant; allocating corresponding privacy budgets to the features of the high-dimensional data based on the information entropy of the high-dimensional data; adding noise to each piece of high-dimensional data for data perturbation according to the privacy budget obtained by each feature; estimating the frequency of discrete features according to the perturbed high-dimensional data; and training a deep learning model according to the perturbed high-dimensional data to obtain a high-dimensional data privacy processing model.

[0006] Further, the allocation of corresponding privacy budgets to the features of the high-dimensional data based on the information entropy of the high-dimensional data comprises: calculating the information entropy of each feature in the high-dimensional data according to the formula , wherein A[j] represents a feature in the high-dimensional data, and H(A[j]) represents the information entropy of the feature; normalizing the information entropy of each feature in the high-dimensional data to obtain the privacy budget allocation rate corresponding to each feature in the high-dimensional data; and allocating corresponding privacy budgets to each feature of the high-dimensional data based on the privacy budget allocation rate corresponding to each feature in the high-dimensional data.

[0007] Further, the privacy budget obtained according to each feature adds noise to the data perturbation, comprising: judging whether each feature of the current record is a continuous feature or a discrete feature; when the feature is a continuous feature, using Laplace mechanism for perturbation; or, when the feature is a discrete feature, using an optimized k-RR mechanism for perturbation; the using an optimized k-RR mechanism for perturbation, comprising: calculating the probability p of each candidate value in the data set k ∈{p1,p2,…,p m};According to the formula The sum of the probabilities of other candidate values except the data to be perturbed r is calculated p, and the probability of perturbing the data to be perturbed r k to other candidate values r According to the formula The data to be perturbed r is perturbed to r'.

[0008] Further, the frequency estimation for discrete features according to the perturbed high-dimensional data, comprising: counting the number of times each candidate value appears in the noisy data; according to the formula count t [r k ] = counts[r k ] - count f [r k ], the frequency estimation of the original data is performed; wherein, count t [r k ] represents the number of true values r k and perturbed values r k , counts[r k ] represents the number of times r k is counted after noise, and count f [r k ] represents the number of true values that are not r k and perturbed values r k .

[0009] Further, the high-dimensional data privacy processing model is obtained by training the deep learning model according to the perturbed high-dimensional data, comprising: training the deep learning model according to the perturbed high-dimensional data, calculating whether the loss function reaches the preset range; when the loss function does not reach the preset range, performing gradient calculation, gradient noise addition, and gradient clipping on the deep learning model, and updating the model parameters for iterative training; when the loss function reaches the preset range, the high-dimensional data privacy processing model is obtained.

[0010] The high-dimensional data privacy processing device provided by this invention is implemented through the following technical solutions: a data collection module for collecting high-dimensional data generated by each participant; a privacy allocation module for allocating a corresponding privacy budget to the features of the high-dimensional data based on the information entropy of the high-dimensional data; a data perturbation module for adding noise to each piece of high-dimensional data to perturb the data according to the privacy budget obtained for each feature; a frequency estimation module for estimating the frequency of discrete features based on the perturbated high-dimensional data; and a model training module for training a deep learning model based on the perturbated high-dimensional data to obtain a high-dimensional data privacy processing model.

[0011] Furthermore, the privacy allocation module includes: an information entropy calculation submodule, used to calculate according to the formula The information entropy of each feature in the high-dimensional data is calculated, where A[j] represents a feature in the high-dimensional data and H(A[j]) represents the information entropy of that feature; the privacy budget allocation rate calculation submodule is used to normalize the information entropy of each feature in the high-dimensional data to obtain the privacy budget allocation rate corresponding to each feature in the high-dimensional data; the privacy budget allocation submodule is used to allocate the corresponding privacy budget to each feature of the high-dimensional data based on the privacy budget allocation rate corresponding to each feature in the high-dimensional data.

[0012] Furthermore, the data perturbation module includes: a feature judgment submodule, used to determine whether each feature of the current record is a continuous feature or a discrete feature; a continuous feature perturbation submodule, used to perform perturbation using the Laplace mechanism when the feature is continuous; and a discrete feature perturbation submodule, used to perform perturbation using an optimized k-RR mechanism when the feature is discrete. The discrete feature perturbation submodule is also used to calculate the probability p of each candidate value in the dataset. k ∈{p1,p2,…,p m}; According to the formula Calculate the sum of probabilities p of all candidate values ​​other than the data r to be perturbed, and calculate the perturbation to the other candidate value r. k probability According to the formula The data r to be perturbed is perturbed to r'.

[0013] Furthermore, the frequency estimation module includes: a discrete feature candidate value statistics submodule, used to count the number of occurrences of each candidate value in the noisy data; and a discrete feature frequency statistics submodule, used to calculate the frequency of occurrence of each candidate value according to the formula count. t [r k ] = counts[r k ]-count f [r k Perform frequency estimation on the original data; where, countt [r k ] represents the number of real values r k and the disturbance value is r k , counts[r k ] represents the number of r k after adding noise, count f [r k ] represents the number of real values that are not r k and the disturbance value is r k .

[0014] Further, the model training module comprises: a forward propagation submodule for training a deep learning model according to the perturbed high-dimensional data, and calculating whether a loss function reaches a preset range; a backward propagation submodule for gradient calculation, gradient noise addition, and gradient clipping on the deep learning model when the loss function does not reach the preset range, and updating model parameters for iterative training; and a target model training submodule for obtaining the high-dimensional data privacy processing model when the loss function reaches the preset range.

[0015] Compared with the prior art, the beneficial effects of the present application include:

[0016] 1. By allocating corresponding privacy budgets to high-dimensional data according to the actual data situation, less privacy budget is allocated to data with lower information entropy, and more privacy budget is allocated to data with higher information entropy, so that the privacy processing efficiency of the data as a whole is improved.

[0017] 2. For the data perturbation problem of discrete features, by comprehensively considering the probability distribution of each perturbation candidate value, the result of data perturbation is more in line with the actual situation, the rationality of data perturbation for discrete features is improved, and the efficiency of data privacy processing is further improved.

[0018] 3. The deep learning model is trained based on the reasonably perturbed data, so that the model obtained by training can resist inversion attacks, and the security and efficiency of the model are improved. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a flowchart of the high-dimensional data privacy processing method in the present embodiment;

[0020] Figure 2 is another flowchart of the high-dimensional data privacy processing method in the present embodiment;

[0021] Figure 3 is a structural diagram of the high-dimensional data privacy processing device in the present embodiment;

[0022] Figure 4 is a structural schematic diagram of a privacy allocation module in the embodiment;

[0023] Figure 5 is a structural schematic diagram of a data perturbation module in the embodiment;

[0024] Figure 6 is a structural schematic diagram of a frequency estimation module in the embodiment;

[0025] Figure 7 is a structural schematic diagram of a model training module in the embodiment. EMBODIMENT

[0026] The application will be described in further detail below with reference to the drawings:

[0027] Reference Figure 1 The privacy processing method for high-dimensional data in the embodiment includes:

[0028] 101. Collect high-dimensional data generated by each participant.

[0029] In the embodiment, the high-dimensional data has a complex structure and relationship, contains rich semantic information, and usually has tens or hundreds of characteristic data.

[0030] 102. Assign a corresponding privacy budget to each characteristic of the high-dimensional data based on the information entropy of the high-dimensional data.

[0031] For the embodiment, the information entropy is a measure of information uncertainty, used to represent the uncertainty or complexity of a variable. When the value of the information entropy of a characteristic is larger, the corresponding variable is more unstable, indicating that the information amount of the characteristic is larger. Therefore, in the embodiment, the allocation of the privacy budget is controlled by the value of the information entropy. If the information entropy of the high-dimensional data is larger, the privacy budget allocated to it is also larger, and at this time, the noise is relatively smaller, and accordingly, the utility of the high-dimensional data after adding noise is higher.

[0032] In the embodiment, the corresponding privacy budget is assigned to the high-dimensional data according to the actual data situation of the high-dimensional data, less privacy budget is assigned to characteristics with lower information entropy, and more privacy budget is assigned to characteristics with higher information entropy, so that the utility of the privacy processing of the data is improved as a whole.

[0033] 103. Add noise to each piece of high-dimensional data according to the privacy budget obtained for each characteristic to perform data perturbation.

[0034] In the embodiment of the present application, firstly, it is judged whether each feature of the record is a continuous feature or a discrete feature. For continuous data, a Laplace mechanism is used for perturbation, that is, the value after perturbation is equal to the true value plus a noise obeying Laplace distribution. For discrete data, an optimized k-RR mechanism is used for perturbation.

[0035] The k-RR mechanism solves the problem that RR (Random Response mechanism) can only be used to process binary variables, and extends RR to the case of multiple candidate values. In the embodiment of the present application, before performing the optimized k-RR mechanism, the probability p k ∈{p1,p2,…,p m of each candidate value in the data set needs to be calculated first. The probability coefficient of perturbing other candidate values r k is When the data is published, p k is perturbed by the Laplace mechanism, and then normalized to obtain p k ′∈{p1′,p2′,…,p m ′}, which is published together with the perturbed data, so as to facilitate the data user to use the data for distribution estimation. For any input r, the probability formula of the output variable r' of the optimized k-RR is: That is, the probability that the value before and after perturbation is the same is The probability of perturbing other variable values is determined based on the probability distribution of the variable values.

[0036] For the data perturbation problem of discrete features in the embodiment of the present application, by comprehensively considering the probability distribution of each perturbation candidate value, the result of data perturbation is more in line with the actual situation, and the rationality of discrete feature data perturbation is improved, so as to further improve the utility of data privacy processing. In the embodiment of the present application, by using the optimized k-RR mechanism, the problem of low utility of noisy discrete data is solved; and further according to the characteristics of the optimized k-RR method, an estimation method for the optimized k-RR is proposed for the estimation of discrete data.

[0037] In this embodiment of the invention, the purpose of perturbing the data is to protect data privacy. Perturbation strategies typically involve adding a certain amount of random noise to the data. This noise is sufficient to mask the specific values ​​of individual data points without completely distorting the statistical characteristics of the entire dataset. In practical applications, optimizing the k-RR mechanism can yield the most accurate data estimates possible, which not only helps in better understanding the data but also positively impacts subsequent decision-making processes. In this embodiment, continuous data is perturbed using the Laplace mechanism, while discrete data is perturbed using an optimized k-RR mechanism.

[0038] 104. Based on the perturbed high-dimensional data, perform frequency estimation for discrete features.

[0039] For embodiments of the present invention, the frequency estimation method for optimizing k-RR designs can be simply divided into two steps: first, count the number of occurrences of each candidate value (counts[r]). k Then, using the statistical candidate value r k The number of times is equal to the original value of r. k The disturbance is r k The number of times plus the original value not being r k The disturbance is r k The number of times this is performed can be used to estimate the original state of the data. That is, the true value is r. k The disturbance value is r k Quantity count t [r k [equals the statistical value of r after adding noise] k counts[r k Subtracting the true value is not r k The disturbance value is r k Quantity count f [r k ]: count t [r k ] = counts[r k ]-count f [r k ].

[0040] In this embodiment of the invention, based on the characteristics of the k-RR optimization method, an estimation method for optimizing k-RR is proposed for the estimation of discrete data.

[0041] 105. Based on the perturbed high-dimensional data, the deep learning model is trained to obtain the target deep learning model, which is used to complete the preset task.

[0042] For the embodiment of the application, by means of the deep learning method based on differential privacy for high-dimensional data, the deep learning model is trained based on the data after reasonable disturbance, so that the model obtained by training can resist inversion attack, and the security and efficiency of the model are improved.

[0043] Compared with the prior art, the embodiment of the application can assign corresponding privacy budgets to high-dimensional data according to the actual data situation of the high-dimensional data, assign less privacy budget to data with lower information entropy, and assign more privacy budget to data with higher information entropy, so that the privacy processing utility of the data is improved as a whole. At the same time, for the data disturbance problem of discrete features, by comprehensively considering the probability distribution of each disturbance candidate value, the result of data disturbance is more in line with the actual situation, the rationality of data disturbance for discrete features is improved, and the utility of data privacy processing can be further improved. And the deep learning model is trained based on the data after reasonable disturbance, so that the model obtained by training can resist inversion attack, and the security and efficiency of the model are improved.

[0044] Reference Figure 2 Another high-dimensional data privacy processing method in the specific embodiment includes:

[0045] 201, collecting high-dimensional data generated by each participant.

[0046] In the embodiment of the application, the high-dimensional data has a complex structure and relationship, contains rich semantic information, and usually has tens or hundreds of characteristic data.

[0047] 202, calculating the information entropy of each feature in the high-dimensional data.

[0048] Step 202 can be specifically: according to the formula calculating the information entropy of each feature in the high-dimensional data, wherein A[j] represents a feature in the high-dimensional data, and H(A[j]) represents the information entropy of the feature.

[0049] 203, normalizing the information entropy of each feature in the high-dimensional data to obtain the privacy budget allocation rate corresponding to each feature in the high-dimensional data.

[0050] 204, based on the privacy budget allocation rate corresponding to each feature in the high-dimensional data, assigning corresponding privacy budgets to each feature of the high-dimensional data.

[0051] For the embodiment of the present application, the information entropy is a measure of the uncertainty of information, which is used to represent the uncertainty or complexity of a variable. When the value of the information entropy of a feature is larger, it means that the information amount of the feature is larger, i.e., the corresponding variable is more unstable. Therefore, in the embodiment of the present application, the allocation of the privacy budget is controlled by the value of the information entropy. If the information entropy of the high-dimensional data is larger, the privacy budget allocated to the high-dimensional data is also larger, and at this time, the noise is relatively smaller, and correspondingly, the utility of the high-dimensional data after adding the noise is higher.

[0052] In the embodiment of the present application, the corresponding privacy budget is allocated to the high-dimensional data according to the actual data situation of the high-dimensional data, less privacy budget is allocated to the feature with lower information entropy, and more privacy budget is allocated to the feature with higher information entropy, so that the utility of the privacy processing of the data is improved as a whole.

[0053] 205, judge whether each feature in the current record is a continuous feature or a discrete feature.

[0054] 206, when the feature is a continuous feature, use the Laplace mechanism to disturb.

[0055] 207, when the feature is a discrete feature, use the optimized k-RR mechanism to disturb.

[0056] For the embodiment of the present application, step 207 can be specifically: calculating the probability p k ∈{p1,p2,…,p m of each candidate value in the data set, calculating the sum p of the probabilities of the other candidate values except the data to be disturbed r according to the formula , and calculating the probability coefficient of disturbing the data to be disturbed r k to the other candidate values r according to the formula .

[0057] In the embodiment of the present application, before the optimized k-RR mechanism is executed, the probability p k ∈{p1,p2,…,p m of each candidate value in the data set needs to be calculated first, and then the other candidate values are disturbed in direct proportion to the probability of the distribution of the other candidate values themselves. Specifically, the probability of the other candidate values except the data to be disturbed r is summed up The probability coefficient of disturbing the data to be disturbed r k to the other candidate values r is When the data is published, p k needs to be disturbed by the Laplace mechanism, and then p k ′∈{p1′,p2′,…,p mand the perturbed data are published together, facilitating the data users to use the data for distribution estimation. For any input r, the probability formula for optimizing the k-RR output variable r' is as follows: that is, the probability that the value before and after the perturbation is the same is The probability of the perturbation being other variable values is determined based on the probability distribution of the variable values.

[0058] For the embodiments of the present application, for the data perturbation problem of discrete features, by comprehensively considering the probability distribution of each perturbation candidate value, the result of data perturbation is more in line with the actual situation, and the rationality of data perturbation for discrete features is improved, so that the utility of data privacy processing can be further improved. In the embodiments of the present application, by optimizing the k-RR mechanism, the problem of low utility of noisy discrete data is solved.

[0059] In the embodiments of the present application, the purpose of perturbing data is to protect the privacy of reading data, and the perturbation strategy usually involves adding a certain amount of random noise to the data. Such noise is sufficient to mask the specific value of an individual data point, but it does not distort the statistical properties of the entire data set. In actual use, by optimizing the k-RR mechanism, the most accurate data estimation can be obtained, which not only helps to better understand the data, but also has a positive impact on the subsequent decision-making process. In the embodiments of the present application, continuous data is perturbed by the Laplace mechanism, and discrete data is perturbed by the optimized k-RR mechanism.

[0060] 208. According to the perturbed high-dimensional data, frequency estimation is performed for discrete features.

[0061] For the embodiments of the present application, step 208 is specifically: counting the number of times each candidate value appears in the noisy data counts[r k ]; according to the formula count t [r k ]=counts[r k ]-countf[r k ], the frequency of the original data is estimated; wherein count t [r k ] represents the number of real values r k and perturbed values r k , counts[r k ] represents the number of times r k is counted after noise, and count f [r k ] represents the number of real values that are not r k and perturbed values r k .

[0062] In the embodiment of the present application, according to the characteristics of the method for optimizing k-RR, an estimation method for optimizing k-RR is proposed for the estimation of discrete data.

[0063] 209. According to the perturbed high-dimensional data, the deep learning model is trained, and whether the loss function reaches the preset range is calculated.

[0064] 210. When the loss function reaches the preset range, the high-dimensional data privacy processing model is obtained; or, when the loss function does not reach the preset range, gradient calculation, gradient noise addition, and gradient clipping are performed on the deep learning model, and the model parameters are updated for iterative training.

[0065] For the embodiment of the present application, through the deep learning method based on differential privacy for high-dimensional data, the deep learning model is trained based on the data after reasonable perturbation, so that the model obtained by training can resist inversion attack, and the security and efficiency of the model are improved.

[0066] In the embodiment of the present application, differential privacy refers to a random algorithm A, which satisfies the condition that: only when, for any adjacent data sets D and D', and any output O, the condition is satisfied, then the random algorithm A satisfies differential privacy. Wherein, Pr[A(*)=O] represents the probability of output O when the input of algorithm A is data set *. The above condition limits the probability that the query results of adjacent data sets D and D' are the same within a certain range, so that no matter how much background knowledge the attacker has, he cannot infer individual information with high confidence, thereby protecting the data set. When e is closer to 0, i.e. e ε is closer to 1, the probability of output O should also be closer, so that the influence of the input of the random algorithm A on the output O should be smaller. The most ideal case is that e=0, Pr[A(D)=O]=Pr[A(D')=O], that is, no matter how similar the input data sets are, the probability of the output O of the random algorithm A should be the same.

[0067] Compared with the prior art, the embodiment of the present application can allocate corresponding privacy budget to high-dimensional data according to the actual data situation of the high-dimensional data, allocate less privacy budget to data with low information entropy, and allocate more privacy budget to data with high information entropy, so that the privacy processing efficiency of the data as a whole is improved. At the same time, for the data perturbation problem of discrete features, by comprehensively considering the probability distribution of each perturbation candidate value, the result of data perturbation is more in line with the actual situation, the rationality of data perturbation for discrete feature data is improved, so that the utility of data privacy processing can be further improved. And, based on the data after reasonable perturbation, the deep learning model is trained, so that the model obtained by training can resist inversion attack, and the security and efficiency of the model are improved.

[0068] Reference Figure 3 As shown in the specific embodiments, the high-dimensional data privacy processing device includes:

[0069] A data collection module 301 is configured to collect high-dimensional data generated by each participant.

[0070] A privacy allocation module 302 is configured to allocate a corresponding privacy budget to each feature of the high-dimensional data based on information entropy of the high-dimensional data.

[0071] A data perturbation module 303 is configured to add noise to each piece of high-dimensional data according to the privacy budget obtained by each feature to perform data perturbation.

[0072] A frequency estimation module 304 is configured to perform frequency estimation on discrete features according to the perturbed high-dimensional data.

[0073] A model training module 305 is configured to train a deep learning model according to the perturbed high-dimensional data to obtain a target deep learning model, which is used to complete a preset task.

[0074] Reference Figure 4 Further, the privacy allocation module 302 includes an information entropy calculation submodule 3021, a privacy budget allocation rate calculation submodule 3022, and a privacy budget allocation submodule 3023.

[0075] The information entropy calculation submodule 3021 is configured to calculate information entropy of each feature in the high-dimensional data according to a formula wherein A[j] represents a feature in the high-dimensional data, and H(A[j]) represents information entropy of the feature.

[0076] The privacy budget allocation rate calculation submodule 3022 is configured to perform normalization processing on the information entropy of each feature in the high-dimensional data to obtain a corresponding privacy budget allocation rate of each feature in the high-dimensional data.

[0077] The privacy budget allocation submodule 3023 is configured to allocate a corresponding privacy budget to each feature of the high-dimensional data based on the corresponding privacy budget allocation rate of each feature in the high-dimensional data.

[0078] Reference Figure 5 Further, the data perturbation module 303 includes a feature judgment submodule 3031, a continuous feature perturbation submodule 3032, and a discrete feature perturbation submodule 3033.

[0079] The feature judgment submodule 3031 is configured to judge whether each feature of a current record is a continuous feature or a discrete feature.

[0080] The continuous feature perturbation submodule 3032 is used to perform perturbation using the Laplace mechanism when the feature is continuous.

[0081] Discrete feature perturbation submodule 3033 is used to perform perturbation using an optimized k-RR mechanism when the feature is discrete.

[0082] The discrete feature perturbation submodule 3033 is also used to calculate the probability p of each candidate value in the dataset. k ∈{p1,p2,…,p m}; According to the formula Calculate the sum of probabilities p of all candidate values ​​other than the data r to be perturbed, and calculate the perturbation to the other candidate value r. k probability According to the formula The data r to be perturbed is perturbed to r'.

[0083] refer to Figure 6 As shown, the frequency estimation module 304 further includes: a discrete feature candidate value statistics submodule 3041 and a discrete feature frequency statistics submodule 3042.

[0084] Discrete feature candidate value statistics submodule 3041 is used to count the number of occurrences of each candidate value in the noisy data. k ].

[0085] Discrete feature frequency statistics submodule 3042 is used to count according to the formula count t [r k ] = counts[r k ]-count f [r k Perform frequency estimation on the original data; where, count t [r k ] represents the number of values ​​where the true value is also the perturbation value, counts[r k [] indicates the statistical value of r after adding noise. k The number of times, count f [r k ] indicates that the true value is not r. k And the disturbance value is r k The quantity.

[0086] refer to Figure 7 As shown, the model training module 305 further includes: a forward propagation submodule 3051, a backward propagation submodule 3052, and a target model training submodule 3053.

[0087] The forward propagation submodule 3051 is configured to train the deep learning model according to the perturbed high-dimensional data, and to calculate whether the loss function reaches a preset range.

[0088] The back propagation submodule 3052 is configured to, when the loss function does not reach the preset range, perform gradient calculation, gradient noise addition, and gradient clipping on the deep learning model, and update the model parameters, and then perform forward propagation.

[0089] The target model training submodule 3053 is configured to, when the loss function reaches the preset range, obtain the high-dimensional data privacy processing model.

[0090] The adaptive high-dimensional data processing apparatus based on differential privacy provided by the specific embodiment can implement the method embodiments provided above, and the specific function implementation can be referred to the description in the method embodiments, which will not be repeated here.

[0091] The technical solution described above is only one embodiment of the present application. For those skilled in the art, on the basis of the principles disclosed in the present application, various types of improvements or modifications can be easily made, and are not limited to the technical solution described in the above specific embodiments. Therefore, the above description is only preferred, and is not limiting.

Claims

1. A method for privacy processing of high-dimensional data, characterized in that, include: Collect high-dimensional data generated by each participant; Based on the information entropy of the high-dimensional data, a corresponding privacy budget is allocated to the features of the high-dimensional data. Based on the privacy budget obtained for each feature, noise is added to each piece of high-dimensional data to perturb the data. Frequency estimation is performed on discrete features based on the perturbated high-dimensional data. Based on the perturbed high-dimensional data, a deep learning model is trained to obtain a target deep learning model, which is used to complete a preset task.

2. The method for privacy processing of high-dimensional data according to claim 1, characterized in that, The allocation of a corresponding privacy budget to the features of the high-dimensional data based on the information entropy of the high-dimensional data includes: According to the formula Calculate the information entropy of each feature in the high-dimensional data, where A[j] represents a feature in the high-dimensional data and H(A[j]) represents the information entropy of that feature; The information entropy of each feature in the high-dimensional data is normalized to obtain the privacy budget allocation rate corresponding to each feature in the high-dimensional data. Based on the privacy budget allocation rate corresponding to each feature in the high-dimensional data, a corresponding privacy budget is allocated to each feature of the high-dimensional data.

3. The method for privacy processing of high-dimensional data according to claim 1, characterized in that, The step of adding noise to each piece of high-dimensional data to perturb the data based on the privacy budget obtained for each feature includes: Determine whether each feature in the current record is a continuous feature or a discrete feature; When the feature is continuous, the Laplace mechanism is used for perturbation; or, when the feature is discrete, the optimized k-RR mechanism is used for perturbation. The perturbation using the optimized k-RR mechanism includes: Calculate the probability p of each candidate value in the dataset. k ∈{p1,p2,…,p m }; According to the formula Calculate the sum of probabilities p of all candidate values ​​other than the data r to be perturbed, and calculate the perturbation to the other candidate value r. k probability According to the formula The data r to be perturbed is perturbed to r'.

4. The method for privacy processing of high-dimensional data according to claim 1, characterized in that, The frequency estimation based on the perturbed high-dimensional data for discrete features includes: Count the number of times each candidate value appears in the noisy data. k ]; According to the formula count t [r k ] = counts[r k ]-count f [r k ], to perform frequency estimation of the original data; where, count t [r k ] indicates that the actual value is r k And the disturbance value is r k The number, counts[r k [] indicates the statistical value of r after adding noise. k The number of times, count f [r k ] indicates that the true value is not r. k And the disturbance value is r k The quantity.

5. The method for privacy processing of high-dimensional data according to claim 1, characterized in that, The step involves training a deep learning model based on the perturbed high-dimensional data to obtain a target deep learning model, which is used to complete a preset task, including: Based on the perturbed high-dimensional data, the deep learning model is trained, and the loss function is calculated to see if it reaches a preset range. When the loss function reaches the preset range, the high-dimensional data privacy processing model is obtained; or, when the loss function does not reach the preset range, gradient calculation, gradient noise addition, and gradient clipping are performed on the deep learning model, and the model parameters are updated for iterative training.

6. A privacy processing device for high-dimensional data, characterized in that, include: The data collection module is used to collect high-dimensional data generated by each participant; The privacy allocation module is used to allocate a corresponding privacy budget to the features of the high-dimensional data based on the information entropy of the high-dimensional data. The data perturbation module is used to add noise to each piece of high-dimensional data to perturb the data based on the privacy budget obtained for each feature; The frequency estimation module is used to estimate the frequency of discrete features based on the perturbed high-dimensional data. The model training module is used to train a deep learning model based on perturbed high-dimensional data to obtain a target deep learning model, which is used to complete a preset task.

7. The privacy processing apparatus for high-dimensional data according to claim 6, characterized in that, The privacy allocation module includes: The information entropy calculation submodule is used to calculate the entropy based on the formula. Calculate the information entropy of each feature in the high-dimensional data, where A[j] represents a feature in the high-dimensional data and H(A[j]) represents the information entropy of that feature; The privacy budget allocation rate calculation submodule is used to normalize the information entropy of each feature in the high-dimensional data to obtain the privacy budget allocation rate corresponding to each feature in the high-dimensional data. The privacy budget allocation submodule is used to allocate a corresponding privacy budget to each feature of the high-dimensional data based on the privacy budget allocation rate corresponding to each feature in the high-dimensional data.

8. The privacy processing apparatus for high-dimensional data according to claim 6, characterized in that, The data perturbation module includes: The feature determination submodule is used to determine whether each feature of the current record is a continuous feature or a discrete feature. The continuous feature perturbation submodule is used to perform perturbation using the Laplace mechanism when the feature is continuous. The discrete feature perturbation submodule is used to perform perturbation using an optimized k-RR mechanism when the feature is discrete. The discrete feature perturbation submodule is also used to calculate the probability p of each candidate value in the dataset. k ∈{p1,p2,…,p m }; According to the formula Calculate the sum of probabilities p of all candidate values ​​other than the data r to be perturbed, and calculate the perturbation to the other candidate value r. k probability According to the formula The data r to be perturbed is perturbed to r'.

9. The privacy processing apparatus for high-dimensional data according to claim 6, characterized in that, The frequency estimation module includes: The discrete feature candidate value statistics submodule is used to count the number of times each candidate value appears in the noisy data (counts[r]). k ]; The discrete feature frequency statistics submodule is used to count according to the formula count. t [r k ] = counts[r k ]-count f [r k Perform frequency estimation on the original data; where, count t [r k ] indicates that the actual value is r k And the disturbance value is r k The number, counts[r k [] indicates the statistical value of r after adding noise. k The number of times, count f [r k ] indicates that the true value is not r. k And the disturbance value is r k The quantity.

10. The privacy processing apparatus for high-dimensional data according to claim 6, characterized in that, The model training module includes: The forward propagation submodule is used to train the deep learning model based on the perturbed high-dimensional data and calculate whether the loss function reaches the preset range. The backpropagation submodule is used to perform gradient calculation, gradient noise addition, and gradient clipping on the deep learning model when the loss function does not reach the preset range, and then perform forward propagation after updating the model parameters. The target model training submodule is used to obtain the high-dimensional data privacy processing model when the loss function reaches a preset range.

Images