Private data protection method based on deep learning

Abstract

The invention provides a private data protection method based on deep learning, and the method comprises the following steps: S11, carrying out the classification and labeling of a sensitive data set, recognizing sensitive information, and carrying out the processing of the sensitive information; s12, sensitive information in the sensitive data set is divided into a discrete data set and a continuous data set, and different types of noise are added to the continuous data set and the discrete data set for disturbance; s13, encrypting the disturbed data set, and sharing and publishing the data set to a data user; s14, monitoring the privacy risk of the shared data in real time by using a privacy computing technology; and S15, dynamically adjusting the data sharing strategy according to the privacy risk degree. According to the method, based on the deep learning model and in combination with the differential privacy algorithm, after noise adding processing is carried out on the identified privacy data, dynamic risk monitoring is carried out on the shared data, the privacy of the patient can be protected, and risks existing in the data sharing process can be monitored and managed in real time.

Description

Technical Field

[0001] The present invention relates to the field of data privacy technology, and in particular to a privacy data protection method based on deep learning. Background Art

[0002] Deep learning is widely used in image classification, speech processing, medical image segmentation, natural language processing and other fields. Excellent deep learning models need to rely on a large number of data sets for training. The data sets usually contain various sensitive information. The continuous improvement of model accuracy means that the problem of personal privacy leakage is becoming more and more serious. Especially in today's digital age, a large amount of data is created and collected every day. This data can help data collectors have a better positioning of themselves or users, but at the same time they have to face the risk of data privacy leakage.

[0003] Such privacy and security issues will hinder data sharing and affect the future development of deep learning. On the other hand, due to privacy concerns, personal information is difficult to obtain, and institutions cannot enhance their decision-making or prediction capabilities through learning and training with large amounts of differentiated data, making it difficult to advance the development of artificial intelligence technology. Summary of the invention

[0004] In order to solve the problem of data privacy leakage risk during the training of a large number of data sets, the present invention provides a privacy data protection method based on deep learning, comprising the following steps:

[0005] S11. Classify and label sensitive data sets, identify sensitive information and process it;

[0006] S12, dividing the sensitive information in the sensitive data set into a discrete data set and a continuous data set, allocating the privacy budget to the continuous data set and the discrete data set in proportion through a differential privacy algorithm, and adding different types of noise to the continuous data set and the discrete data set for disturbance respectively;

[0007] S13, encrypt the disturbed data set and share it with data users;

[0008] S14. Use privacy computing technology to monitor the privacy risks of shared data in real time;

[0009] S15. Dynamically adjust data sharing strategies based on the degree of privacy risk.

[0010] Furthermore, in step S11, the method for identifying and processing sensitive information is:

[0011] S21, inputting the training set containing the sample sensitive information into the deep learning model for training to obtain a sensitive information recognition model;

[0012] S22, inputting the sensitive data set into the sensitive information identification model to perform sensitive information identification;

[0013] S23. Anonymize the identified sensitive information, where the anonymization process is used to remove or modify information that can directly identify an individual from the sensitive data set.

[0014] Furthermore, in step S12, the privacy budget is used to measure the degree of privacy protection. The smaller the privacy budget, the higher the degree of privacy protection.

[0015] Furthermore, in step S12, the method of adding different types of noise to the discrete data set for disturbance is:

[0016] S31. Generalize the discrete data set, set the size of the privacy budget, use Bootstrap sampling to extract equal amounts of sample data sets with replacement from the generalized discrete data set to build multiple decision trees, and divide the privacy budget equally into each decision tree;

[0017] S32, the privacy budget equally divided into each decision tree is allocated to each level according to the arithmetic difference privacy budget allocation mechanism, and the privacy budget allocated to each node is divided into two equal parts;

[0018] S33. Calculate the Gini index corresponding to the splitting of the sample data set in different splitting modes, compare it with the Gini index of the local optimal feature, select the global optimal splitting feature through the index mechanism, and the global optimal splitting feature is used for node splitting.

[0019] Furthermore, in step S33, the exponential mechanism evaluates each possible split in the sample data set through a utility function, wherein the utility function is based on information gain, generates noise using an exponential distribution according to the utility score of each split, and selects the global optimal split feature.

[0020] Furthermore, in step S12, the method of adding different types of noise to disturb the continuous data set is:

[0021] S41, constructing a tree with a hierarchical structure according to the semantic attributes of the continuous data set, wherein the leaf nodes of the tree represent all real values, and the upper layer nodes represent the generalization of the lower layer nodes according to the semantic attributes;

[0022] S42, constructing a Bayesian network to calculate the average influence of each leaf node on other leaf nodes, and selecting the leaf node with the greatest influence on other leaf nodes as the first node;

[0023] S43. Calculate the joint probability distribution of each leaf node and its parent node according to the Bayesian network and the continuous data set, and add noise to the joint probability distribution through the Laplace mechanism. The joint probability distribution is used to identify sensitive attributes in the continuous data set data and allocate corresponding privacy budgets for the sensitive attributes.

[0024] Further, in step S13, the data after adding noise is encrypted by an encryption algorithm, and the processed data is stored in an encrypted form. The user needs to decrypt the data according to the authorized password.

[0025] Further, in step S14, the method for real-time monitoring of the privacy risk is as follows:

[0026] S51. In the data sharing stage, use privacy computing technology to detect whether there is a risk of personal information leakage during the use of the shared data;

[0027] S52. Conduct real-time detection of the use of the shared data, and identify and evaluate whether personal information leakage will be caused in the data processing activities;

[0028] S53. According to the attributes of the sensitive data set, define the result of the shared data detection as a and the risk prompt threshold as A. If a≥A, initiate a risk prompt; if a<A, prompt that there is no risk;

[0029] S54. If a risk prompt is initiated, stop data sharing and prompt the administrator to perform security maintenance on the shared data.

[0030] Further, in step S52, for all risk prompts, the publisher needs to manually confirm whether the risk has been lifted. After confirming that the risk has been lifted, continue to publish the sharing task.

[0031] The beneficial effects of the present invention are as follows:

[0032] The present invention provides a privacy data protection method based on deep learning, which combines a deep learning model with a differential privacy algorithm to provide strong privacy protection; by applying different differential privacy mechanisms to discrete data sets and continuous data sets, the amount of noise is more precisely controlled, so as to minimize the impact on data accuracy and usability while protecting privacy; the perturbed data set is encrypted before data sharing, further enhancing the security of data sharing. The present invention not only protects the privacy of data but also ensures the usability of data, promoting the safe and sustainable development of artificial intelligence technology. Description of the Drawings

[0033] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0034] Figure 1 This is a flow chart of a privacy data protection method based on deep learning proposed in the present invention. DETAILED DESCRIPTION

[0035] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only parts related to the present invention, rather than all structures, are shown in the accompanying drawings.

[0036] like Figure 1 As shown, a privacy data protection method based on deep learning, the method comprising:

[0037] S11. Classify and label sensitive data sets according to different data attributes and characteristics, such as identity information, financial status, health status, etc., identify whether there is sensitive information in the classification attributes, and process it to provide a good data environment for subsequent noise processing.

[0038] In this embodiment, the method for identifying and processing sensitive information is:

[0039] S21. Input the training set containing sample sensitive information into the deep learning model for training to obtain a sensitive information recognition model. Input the training set containing sensitive information, including personal information that needs to be protected as stipulated by relevant laws and sensitive information in various industries, into the deep learning model. The deep learning model processes, trains, and learns these data to obtain a sensitive information recognition model.

[0040] S22. Input the sensitive data set into the sensitive information identification model to identify sensitive information. When a group of data is input into the sensitive information identification model, the sensitive information identification model can quickly identify the sensitive information that needs to be protected in this group of data, and provide good data conditions for subsequent noise addition processing. Good data conditions can help the differential privacy algorithm identify the sensitive information that needs to be noised during noise addition, so as to achieve reasonable allocation of the privacy budget.

[0041] S23. Anonymize the identified sensitive information, where the anonymization process is used to remove or modify information that can directly identify an individual from the sensitive data set. Before performing differential privacy processing, the sensitive information is appropriately anonymized to ensure the confidentiality of personal information. Anonymizing sensitive information helps to improve the privacy protection level of data, reduce the risk of data leakage, and ensure the compliance and availability of data.

[0042] S12. The sensitive information in the sensitive data set is divided into discrete data sets and continuous data sets. By distinguishing between continuous data sets and discrete data sets, the optimal algorithm can be selected for each data type, and the privacy budget can be allocated more accurately, thereby improving the overall performance and ensuring privacy protection while maximizing the practicality of the data. The privacy budget is proportionally allocated to the continuous data set and the discrete data set through the differential privacy algorithm, and different types of noise are added to the continuous data set and the discrete data set for disturbance. The privacy budget is used to measure the degree of privacy protection. The smaller the privacy budget, the higher the degree of privacy protection.

[0043] In this embodiment, the method of adding different types of noise to the discrete data set for disturbance is:

[0044] S31. Generalize the discrete data set. Generalization helps to preserve the statistical analysis value of the data to the maximum extent while protecting personal privacy, and achieve a balance between privacy protection and data utility. Set the size of the privacy budget. Use Bootstrap sampling to extract equal amounts of sample data sets from the generalized discrete data set with replacement to build multiple decision trees, and divide the privacy budget equally into each decision tree.

[0045] S32. The privacy budget equally divided into each decision tree is allocated to each level according to the arithmetic difference privacy budget allocation mechanism, and the privacy budget allocated to each node is divided into two equal parts. According to the sequential combination characteristics unique to differential privacy, the privacy protection budget consumed by the algorithm as a whole is equal to the sum of the privacy protection budget consumed by each layer of the decision tree. Since the data sets used for splitting and counting by the nodes in each layer are disjoint, according to the parallel combination characteristics of differential privacy, the privacy budget allocated to each level does not need to be evenly distributed to each node, but directly gives each node the same privacy budget that is equal to itself. That is, according to the number of layers and branches of the decision tree, the privacy budget allocation mechanism becomes the same privacy budget for nodes in the same layer, and the nodes between different layers add noise that obeys the arithmetic difference distribution according to the depth of the decision tree, so as to achieve the addition of a small amount of noise when the amount of data is small. While minimizing the impact of noise addition on the performance of the algorithm, the classification accuracy of the algorithm is improved.

[0046] The upper nodes of the decision tree contain more information and have a higher demand for privacy protection, while the lower nodes contain more specific information and require more privacy budget to protect. Dividing the privacy budget allocated to each node into two equal parts can adapt to privacy needs at different levels and add an appropriate amount of noise to each node to reduce the impact of excessive noise on the classification results at a single node, ensuring balanced privacy protection for the entire tree, which helps to improve the classification performance and accuracy of the model while protecting privacy.

[0047] S33. Calculate the Gini index corresponding to the splitting of the sample data set in different splitting methods, and compare it with the Gini index of the local optimal feature. In the process of building the decision tree, calculate the Gini index for each possible splitting point of each feature, and select the splitting point with the smallest Gini index, that is, the maximum purity, as the optimal splitting point of the feature. This process involves traversing all possible values ​​of each feature in the data set, and calculating the Gini index after the data is split according to this feature value, and then select the feature and splitting point with the smallest Gini index as the basis for node splitting of the decision tree.

[0048] The global optimal split feature is selected through an exponential mechanism, and the global optimal split feature is used for node splitting. The global optimal split feature is the feature that maximizes the purity of the data set after splitting during the entire decision tree construction process, that is, the feature that brings the greatest gain on the entire data set, and this feature will be used for node splitting. Node splitting is a step in the recursive construction process of a decision tree. It divides the data set into two or more subsets based on the selected optimal feature, each of which has consistency or high purity in feature values. At each node of the decision tree, the algorithm evaluates all possible features to determine which feature split can maximize the prediction accuracy of the model. The exponential mechanism evaluates each possible split in the sample data set through a utility function, which is based on information gain. According to the utility score of each split, it uses exponential distribution to generate noise and select the global optimal split feature.

[0049] In this embodiment, the method of adding different types of noise to the continuous data set for disturbance is:

[0050] S41. A tree with a hierarchical structure is constructed according to the semantic attributes of the continuous data set, wherein the leaf nodes of the tree represent all real values, and the upper-layer nodes are represented as generalizations of the lower-layer nodes according to the semantic attributes. By using attribute features with hierarchical semantic relationships, the intrinsic semantic structure of the data can be captured more accurately, which helps to retain the important statistical characteristics of the data during data processing, thereby reducing data errors after privacy protection.

[0051] S42. Construct a Bayesian network to calculate the average influence of each leaf node on other leaf nodes. By evaluating the mutual influence between leaf nodes, a more personalized and flexible privacy protection strategy can be implemented to adapt to different data characteristics and privacy requirements. Select the leaf node with the greatest influence on other leaf nodes as the first node to solve the problem of network instability caused by random selection of leaf nodes, reduce the risk of differential attacks to a certain extent, and thus protect the privacy of individuals in the data set; at the same time, more effectively allocate the privacy budget, more accurately control the distribution of noise, and ensure that the focus of privacy protection is on the part that has the most significant impact on the overall model, which is used to enhance the generalization ability on unseen data and maintain data availability while protecting privacy.

[0052] S43. Calculate the joint probability distribution of each leaf node and its parent node based on the Bayesian network and the continuous data set, add noise to the joint probability distribution through the Laplace mechanism, and the joint probability distribution is used to identify sensitive attributes in the continuous data set data and allocate corresponding privacy budgets for the sensitive attributes. The Laplace mechanism is an effective method to achieve differential privacy. By adding Laplace noise, the definition of differential privacy is met, and the availability of data is retained while protecting privacy.

[0053] In this embodiment, taking medical data sharing as an example, medical data often contains highly sensitive personal information, such as patients' medical records, diagnosis results, treatment records, etc. The sensitive data contained in the medical data is divided into discrete data sets and continuous data sets, where the discrete data sets include the number of patients, gender, blood type, disease diagnosis code, etc., and the continuous data sets include the patient's body temperature, blood pressure, height and weight, etc. After the discrete data and continuous data are noised separately, the data are presented in the form of a whole column, retaining the statistical properties. For example, when obtaining gender data, the overall gender ratio or the total number of women is obtained, which is data that does not disclose personal privacy.

[0054] S13, encrypt the disturbed data set and share it with data users;

[0055] The processed data uses data encryption technology to encrypt the data set and store it in encrypted form, that is, the noisy data is encrypted and stored through an encryption algorithm. In this embodiment, a symmetric encryption algorithm is used to encrypt the disturbed data set. The user needs to decrypt the data according to the corresponding decryption key for subsequent analysis and use. That is, after receiving the encrypted data, the user needs to decrypt the data according to the authorization password to ensure that only authorized encrypted data can be accessed by the user.

[0056] S14. Use privacy computing technology to monitor the privacy risks of shared data in real time;

[0057] In this embodiment, the method for real-time monitoring of the privacy risk is as follows:

[0058] S51. During the data sharing phase, use privacy computing technology to detect whether there is a risk of personal information leakage during the use of the shared data. Privacy computing technologies include multi-party secure computing, federated learning, trusted execution environments, etc., which can perform data analysis and calculation without revealing the original data, effectively preventing data leakage and helping the publisher avoid legal liabilities and economic losses caused by data leakage.

[0059] S52. Conduct real-time detection of the use of shared data. The administrator publishes the data sharing through a specific platform. By setting up a supervision mechanism on the platform, such as deploying a data leakage prevention system, it is possible to detect the usage of data in real time during the data sharing process. Automatically detect and identify abnormal behaviors in data access. By identifying and evaluating potential risks, corresponding preventive measures can be taken and early warnings and prevention can be carried out for potential security threats, identifying and evaluating whether personal information leakage will occur in data processing activities, and avoiding infringement of the personal information subject due to personal information leakage.

[0060] S53. According to the attributes of the sensitive data set, define the result of the shared data detection as a, and the risk warning threshold as A. If a≥A, initiate a risk warning; if a<A, it is indicated that there is no risk. Define the mechanism for the result of the shared data detection and the risk warning threshold based on the attributes of the sensitive data set, analyze each attribute in the sensitive data set to determine their sensitivity levels. Scan the shared data set, and generate a quantified detection result a according to the quantity and type of sensitive attributes in the data set. Set a risk warning threshold A according to the publisher's risk tolerance, legal and regulatory requirements, and industry best practices. When the detection result a exceeds this threshold, it is considered that there is an unacceptable risk in data sharing. If a≥A, it is considered that there is a risk, initiate a risk warning, further review the data sharing request, and take necessary risk mitigation measures. If a<A, it is considered that the risk is within an acceptable range, indicate no risk, and allow data sharing.

[0061] S54. If a risk warning is initiated, stop the data sharing and prompt the administrator to perform security maintenance on the shared data. All risk warnings require the publisher to manually confirm whether the risk has been lifted. After confirming that the risk has been lifted, continue to issue the sharing task. Performing security maintenance on the shared data includes setting the privacy risk level and taking corresponding preventive measures according to the corresponding privacy risk level. Through this detection mechanism, the administrator can more effectively manage and control the risks in the data sharing process while ensuring the reasonable utilization and circulation of data.

[0062] S15. Dynamically adjust the data sharing strategy according to the degree of privacy risk. The dynamic adjustment strategy allows the data publisher to flexibly adjust its data sharing policy according to the current security environment and threat intelligence, ensuring that the data sharing strategy always meets the latest security requirements, which can more effectively protect data and reduce the risk of data leakage. In this embodiment, the privacy risk level is set to three levels: low, medium, and high; when the user triggers the low risk level, the platform sends a security verification, including graphic verification, dynamic password verification, etc., to remind the user that the operation has triggered the privacy risk of the platform; when the user triggers the medium risk level, the platform directly jumps to the state where the key needs to be re-authorized before the platform can be accessed; when the user triggers the high risk level, the user is locked and cannot use the key to log in to the platform to obtain shared data.

[0063] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A privacy data protection method based on deep learning, characterized in that: The method comprises: S11. Classify and label sensitive data sets, identify sensitive information and process it; S12, dividing the sensitive information in the sensitive data set into a discrete data set and a continuous data set, allocating the privacy budget to the continuous data set and the discrete data set in proportion through a differential privacy algorithm, and adding different types of noise to the continuous data set and the discrete data set for disturbance respectively; S13, encrypt the disturbed data set and share it with data users; S14. Use privacy computing technology to monitor the privacy risks of shared data in real time; S15. Dynamically adjust data sharing strategies based on the degree of privacy risk.

2. The privacy data protection method based on deep learning according to claim 1, characterized in that: In step S11, the method for identifying and processing sensitive information is: S21, inputting the training set containing the sample sensitive information into the deep learning model for training to obtain a sensitive information recognition model; S22, inputting the sensitive data set into the sensitive information identification model to perform sensitive information identification; S23. Anonymize the identified sensitive information, where the anonymization process is used to remove or modify information that can directly identify an individual from the sensitive data set.

3. The privacy data protection method based on deep learning according to claim 1, characterized in that: In step S12, the privacy budget is used to measure the degree of privacy protection. The smaller the privacy budget, the higher the degree of privacy protection.

4. The privacy data protection method based on deep learning according to claim 1, characterized in that: In step S12, the method of adding different types of noise to the discrete data set for disturbance is: S31. Generalize the discrete data set, set the size of the privacy budget, use Bootstrap sampling to extract equal amounts of sample data sets with replacement from the generalized discrete data set to build multiple decision trees, and divide the privacy budget equally into each decision tree; S32, the privacy budget equally divided into each decision tree is allocated to each level according to the arithmetic difference privacy budget allocation mechanism, and the privacy budget allocated to each node is divided into two equal parts; S33. Calculate the Gini index corresponding to the splitting of the sample data set in different splitting modes, compare it with the Gini index of the local optimal feature, select the global optimal splitting feature through the index mechanism, and the global optimal splitting feature is used for node splitting.

5. The privacy data protection method based on deep learning according to claim 4 is characterized in that: In step S33, the exponential mechanism evaluates each possible split in the sample data set through a utility function, where the utility function is based on information gain, generates noise using an exponential distribution according to the utility score of each split, and selects the global optimal split feature.

6. The privacy data protection method based on deep learning according to claim 1, characterized in that: In step S12, the method of adding different types of noise to disturb the continuous data set is: S41, constructing a tree with a hierarchical structure according to the semantic attributes of the continuous data set, wherein the leaf nodes of the tree represent all real values, and the upper layer nodes represent the generalization of the lower layer nodes according to the semantic attributes; S42, constructing a Bayesian network to calculate the average influence of each leaf node on other leaf nodes, and selecting the leaf node with the greatest influence on other leaf nodes as the first node; S43. Calculate the joint probability distribution of each leaf node and its parent node according to the Bayesian network and the continuous data set, and add noise to the joint probability distribution through the Laplace mechanism. The joint probability distribution is used to identify sensitive attributes in the continuous data set and allocate corresponding privacy budgets for the sensitive attributes.

7. The privacy data protection method based on deep learning according to claim 1, characterized in that: In step S13, the data after adding noise is encrypted through an encryption algorithm, and the processed data is stored in an encrypted form. The user needs to decrypt the data according to the authorized password.

8. The privacy data protection method based on deep learning according to claim 1, characterized in that: In step S14, the method for real-time monitoring of the privacy risk is as follows: S51. In the data sharing stage, detect whether there is a risk of personal information leakage during the use of the shared data through privacy computing technology; S52. Conduct real-time detection of the use of the shared data, and identify and evaluate whether personal information leakage will occur in the data processing activities; S53. According to the attributes of the sensitive data set, define the result of the shared data detection as a, and the risk prompt threshold as A. If a≥A, initiate a risk prompt; if a<A, prompt no risk; S54. If a risk prompt is initiated, stop data sharing and prompt the management personnel to perform security maintenance on the shared data.

9. The privacy data protection method based on deep learning according to claim 8, characterized in that: In step S52, for all risk prompts, the publisher needs to manually confirm whether the risk has been lifted. If the risk is confirmed to be lifted, continue to publish the sharing task.

Images