Data desensitization processing system based on self-supervised learning

Abstract

The invention discloses a data desensitization processing system based on self-supervised learning, and relates to the technical field of data privacy and desensitization, and the system comprises the following steps: S1, based on to-be-desensitized original data, executing preprocessing, and generating a standardized data set; s2, executing data enhancement processing of random shielding, random feature replacement and random noise injection; s3, iteratively optimizing the BYOL pre-training model, and extracting a feature vector of each data sample after dimension reduction; s4, executing an improved DENCLUE algorithm, and generating a sensitive information mark list; s5, executing data shielding, data generalization and data replacement desensitization processing; and S6, carrying out adaptive adjustment iteration on the local density threshold. According to the method, the limitation of a traditional method in processing complex and heterogeneous data is effectively solved, the intelligent level of the desensitization process is improved, and a more efficient and more accurate solution is provided for data privacy protection.

Description

Technical Field

[0001] This invention relates to the fields of data privacy and de-identification technology, and in particular to a data de-identification processing system based on self-supervised learning. Background Technology

[0002] With increasingly stringent data privacy protection requirements, traditional data anonymization methods face unprecedented challenges. In the modern field of data protection, real-time sensitive data anonymization and precise privacy protection are not only crucial means to ensure data security but also key factors in improving data usage efficiency and compliance. However, most current data anonymization methods rely on manual intervention or traditional rule-driven processes, which are cumbersome, inefficient, and fail to meet the ever-growing demands for data privacy protection.

[0003] Traditional data anonymization methods typically rely on manually configured rules and thresholds to identify and process sensitive data. While these methods may work in certain specific scenarios, their lack of automation and intelligent mechanisms prevents them from identifying and processing potential privacy risks in real time and efficiently. Especially when dealing with large-scale, multi-dimensional, and complex data, traditional methods suffer from severely insufficient adaptability and flexibility, making it difficult to address the challenges posed by data diversity and complexity.

[0004] Furthermore, traditional data anonymization methods often overlook the underlying patterns and intrinsic relationships within the data. They rely solely on simple manual settings or rule-driven methods for data replacement, anonymization, or encryption, making it difficult to capture the deep-seated patterns and structural features within the data. For example, the complex relationships between different modalities such as text, images, and audio cannot be effectively extracted, resulting in low accuracy and efficiency in data anonymization, and a high risk of information loss or incorrect anonymization.

[0005] Even with the introduction of traditional machine learning methods for data anonymization, existing approaches still fail to fully leverage the advantages of self-supervised learning techniques, struggling to efficiently handle complex and dynamically changing sensitive data. Traditional methods fail to delve into the underlying structure and patterns of the data, resulting in ineffective anonymization and significantly impacting the accuracy and timeliness of data privacy protection.

[0006] Therefore, how to provide a data anonymization system based on self-supervised learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] This invention proposes a data anonymization system based on self-supervised learning. By combining advanced unsupervised learning techniques, it can efficiently identify and process sensitive data, ensuring data privacy and security. This method not only automatically extracts latent features from data but also effectively learns unsupervised representations of the data through a BYOL pre-trained model, further improving the accuracy and precision of the anonymization process. Simultaneously, combined with an improved DENCLUE algorithm, the system can efficiently identify clustering structures in high-dimensional, highly uncertain data environments and perform dynamic sensitive data anonymization based on the clustering results. This self-supervised learning-based data anonymization system can automatically identify and effectively anonymize sensitive data without relying on manual rules. Compared to traditional rule-based anonymization methods, this invention effectively addresses the limitations of traditional methods when handling complex, heterogeneous data, improving the intelligence level of the anonymization process. This system provides a more efficient and accurate solution for data privacy protection, maximizing the preservation of data value while ensuring user privacy and security.

[0008] A data anonymization processing system based on self-supervised learning according to an embodiment of the present invention includes the following modules: The preprocessing module is used to perform preprocessing on the original data to be de-identified, and generate a standardized dataset. The data augmentation module is used to perform random occlusion, random feature permutation and random noise injection on a standardized dataset to generate occluded data augmentation view, permutation data augmentation view and Gaussian noise data augmentation view. The feature vector extraction module is used to iteratively optimize the BYOL pre-trained model based on occluded data augmentation view, permutation data augmentation view, and Gaussian noise data augmentation view. It inputs the standardized dataset into the trained BYOL pre-trained model and extracts the dimensionality-reduced feature vector of each data sample. The sensitive information labeling module is used to generate a list of sensitive information labels based on the dimensionality-reduced feature vector of each data sample using an improved DENCLUE algorithm. The desensitization module is used to perform data masking, data generalization, and data replacement desensitization processing on the location of sensitive information in the original data based on the sensitive information tag list, and output preliminary desensitized data; The local density threshold iteration module is used to compare the information entropy difference before and after desensitization and to adaptively adjust the local density threshold in the improved DENCLUE algorithm.

[0009] A data anonymization method based on self-supervised learning according to an embodiment of the present invention includes the following steps: S1. Based on the original data to be de-identified, perform preprocessing to generate a standardized dataset; S2. Based on the standardized dataset, perform data augmentation processing with random occlusion, random feature permutation, and random noise injection respectively to generate occluded data augmentation view, permutation data augmentation view, and Gaussian noise data augmentation view; S3. Iteratively optimize the BYOL pre-trained model based on occluded data augmentation view, permutation data augmentation view, and Gaussian noise data augmentation view. Input the standardized dataset into the trained BYOL pre-trained model and extract the dimensionality-reduced feature vector of each data sample. S4. Based on the dimensionality-reduced feature vector of each data sample, execute the improved DENCLUE algorithm to generate a list of sensitive information tags; S5. Based on the list of sensitive information markers, perform data masking, data generalization, and data substitution desensitization processing on the location of sensitive information in the original data, and output preliminary desensitized data; S6. Compare the information entropy difference before and after desensitization, and adaptively adjust the local density threshold in the improved DENCLUE algorithm iteratively.

[0010] Optionally, the preprocessing includes cleaning, normalization, removal of outliers, and data format standardization.

[0011] Optionally, S2 specifically includes: S21. Input a standardized dataset, set a random occlusion ratio, randomly select data items with the corresponding ratio, set a fixed occlusion marker value and replace the original value of the data item to obtain an enhanced view of the occlusion data. S22. Input a standardized dataset, set a random permutation ratio, randomly select data items of the corresponding ratio and swap the positions of the data items to obtain a permuted data augmented view; S23. Input a standardized dataset, set the random noise injection ratio, noise mean and variance, randomly select the corresponding proportion of numerical data items and overlay Gaussian noise to obtain a Gaussian noise data augmentation view.

[0012] Optionally, S3 specifically includes: S31. From the occlusion data augmentation view, the permutation data augmentation view, and the Gaussian noise data augmentation view, randomly select two different data augmentation views and input them into the feature encoders of the online network and the target network of the BYOL pre-trained model. Perform convolution and pooling operations respectively to obtain the feature encoding vectors of the online network and the target network. S32. Input the feature encoding vectors of the online network and the target network into the projection heads of the online network and the target network respectively. Perform linear transformation, batch normalization and ReLU activation function calculation through the first fully connected layer. Perform linear transformation through the second fully connected layer to output the projection feature vectors of the online network and the target network respectively. S33. Input the projection feature vector of the online network into the prediction head, perform linear transformation, batch normalization and ReLU activation function calculation through the first fully connected layer, perform linear transformation through the second fully connected layer, and output the prediction feature vector of the online network. S34. Calculate the magnitude length values ​​of the predicted feature vector of the online network and the projected feature vector of the target network respectively, and multiply them to obtain the magnitude product value; perform an inner product operation on the predicted feature vector of the online network and the projected feature vector of the target network to obtain the inner product value; divide the inner product value by the magnitude product value to obtain the cosine similarity value; define the loss function value as the negative value of the cosine similarity value. S35. Based on the numerical value of the loss function, the Adam optimization algorithm is used to calculate the gradient through backpropagation. The parameters are updated based on the exponential weighted average of the first and second moments. Combined with adaptive learning rate and bias correction, the parameters of each layer of the online network are iteratively optimized until the loss function converges. S36. Continuously update the parameters of each layer of the target network based on the EMA method. Set the smoothing coefficient value of the EMA method, and set the weights of the target network and the online network to the smoothing coefficient value and the complement of the smoothing coefficient value, respectively. After each update of the parameters of each layer of the online network, calculate the weighted average value of the parameter values ​​for each layer of the online network and the target network in the feature encoder and projection head, based on the weights of the target network and the online network. Use the weighted average value to update the parameters of the feature encoder and projection head in the target network, and output the target network after parameter update. S37. Continue to iterate through steps S35 and S36 until the loss function value reaches a convergent state, and obtain the trained BYOL pre-trained model. S38. Input the standardized dataset into the online network of the trained BYOL model. For each data sample in the standardized dataset, pass it through the feature encoder in the online network in turn. After convolution and pooling operations, obtain the feature encoding vector of each data sample. S39. Input the feature encoding vector of each data sample into the projection head of the online network, and after linear transformation through two layers of fully connected networks, obtain the dimensionality-reduced feature vector of each data sample.

[0013] Optionally, the improved DENCLUE algorithm specifically includes: Define the kernel function type and initialize the global bandwidth parameter and momentum coefficient of the kernel function; A local adaptive bandwidth parameter adjustment mechanism is introduced. The dimensionality-reduced feature vectors of each data sample are iterated sequentially. The distance between the current feature vector and other vectors is calculated. A neighborhood size parameter is set, and the nearest feature vectors of the specified neighborhood size parameter are selected to form a neighborhood. The density values ​​of all nearest neighbor feature vectors within the neighborhood are calculated, and the variance of the density within the neighborhood is statistically obtained. Based on the variance of the density within the neighborhood, a density variance threshold is set. If the density variance within the neighborhood exceeds the density variance threshold, the local bandwidth is reduced; otherwise, the local bandwidth is increased. Finally, the local adaptive bandwidth parameter corresponding to each feature vector is obtained. A momentum coefficient mechanism is introduced. Based on the kernel function and the local adaptive bandwidth parameter corresponding to the current feature vector, the current density gradient direction is calculated. A momentum coefficient is set, and the current density gradient direction is combined with the density gradient direction of the previous iteration. The momentum coefficient is used to perform weighted fusion to form the update direction of the momentum. The position of the feature vector is updated along the direction of the momentum update, and the iterative movement continues until the position of the feature vector converges, finally obtaining the density attraction point corresponding to each feature vector. Set a local density threshold, iterate through the density attraction points corresponding to the dimensionality-reduced feature vector of each data sample, calculate the local density value of each density attraction point, and mark the density attraction points whose local density values ​​exceed the corresponding local density threshold as local high density center points, and finally obtain all local high density center points. Using the local high-density center point as a reference, an attraction radius is set, and feature vectors whose density attraction points are located within the attraction radius of the local high-density center are aggregated into corresponding high-density clustering regions, and the clustering results are output. The feature vectors within all local high-density clustering regions are traversed, and the corresponding data sample numbers and data item numbers are extracted to generate a list of sensitive information tags.

[0014] Optionally, S5 specifically includes: S51. Perform numeric replacement desensitization. Iterate through all the marked field positions in the sensitive information mark list. When the marked field is numeric data, determine the upper and lower bounds of the numeric range where the data is located. Based on the upper and lower bounds, construct the range description and replace the original numeric value with the corresponding range. S52. Perform text replacement and desensitization. Iterate through all the marked field positions in the sensitive information mark list. When the marked field is text data, count and construct the non-sensitive text data set of the current field. Randomly select non-sensitive data items from the non-sensitive text data set and replace the original value of the current sensitive text data item. S53. Perform desensitization of highly sensitive fields. Iterate through all the marked field positions in the sensitive information mark list. When the marked field is a highly sensitive data field, set a fixed specific symbol as the masking mark value and completely replace the value of the original data item with the masking mark value. S54. After performing numeric replacement desensitization, text replacement desensitization, and strong sensitive field desensitization, update the positions of the corresponding data sample numbers and data item numbers in the original dataset, and then summarize to obtain the preliminary desensitized dataset.

[0015] Optionally, S6 specifically includes: S61. Input the initial desensitized data and the original data before desensitization, and count the frequency of each data field. Calculate the information entropy value of each field based on its frequency of value, and summarize the information entropy values ​​of all fields to obtain the information entropy of the initial desensitized data and the information entropy of the original data before desensitization. S62. Input the information entropy of the original data before desensitization and the information entropy of the preliminary desensitized data, calculate the numerical difference between the information entropy of the original data and the information entropy of the preliminary desensitized data, and obtain the information entropy difference value. S63. Set an information entropy difference threshold. If the information entropy difference is greater than the information entropy difference threshold, increase the density threshold set in the improved DENCLUE algorithm to obtain an updated local density threshold. Use the updated density threshold to perform the operations in steps S4 and S5; otherwise, output the data that has been desensitized as the final desensitized data.

[0016] The beneficial effects of this invention are: (1) This invention, by combining a BYOL pre-trained model, successfully overcomes the shortcomings of traditional data anonymization methods when dealing with complex data. Traditional anonymization methods typically rely on manual rules or fixed anonymization strategies, making them difficult to adapt to complex and ever-changing data environments. By introducing a BYOL pre-trained model, this invention can automatically learn the latent representations of data through self-supervised learning and effectively identify sensitive data based on these representations. BYOL achieves unsupervised feature learning of data by maximizing similarity without relying on labels, enabling data anonymization to automatically identify and extract implicit sensitive information. This method significantly improves the intelligence level of anonymization processing, effectively handling multimodal and dynamically changing complex data while maintaining efficient data privacy protection.

[0017] (2) This invention significantly improves the accuracy and robustness of data anonymization by combining an improved DENCLUE algorithm. Traditional anonymization methods often fail to effectively uncover potential patterns and privacy risks in high-dimensional and complex data. In contrast, the DENCLUE algorithm utilizes density clustering to efficiently identify outliers and cluster structures in the data, thereby more accurately grasping the distribution characteristics of sensitive data. Through the improved DENCLUE algorithm, this invention can handle multidimensional, uncertain, and heterogeneous data types, and automatically adjust clustering strategies and anonymization methods according to data characteristics to ensure the privacy protection effect after data anonymization. The improved algorithm also effectively optimizes the computational efficiency and accuracy in the clustering process, enabling the anonymization process to maximize the preservation of data validity and compliance while ensuring data security. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0019] Figure 1 This is a schematic diagram of the structure of a data desensitization processing system based on self-supervised learning proposed in this invention; Figure 2 This is an overall flowchart of a data anonymization method based on self-supervised learning proposed in this invention; Figure 3 This is a flowchart of the improved DENCLUE algorithm for a data desensitization processing system based on self-supervised learning proposed in this invention. Detailed Implementation

[0020] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1 A data anonymization system based on self-supervised learning includes the following modules: The preprocessing module is used to perform preprocessing on the original data to be de-identified, and generate a standardized dataset. The data augmentation module is used to perform random occlusion, random feature permutation and random noise injection on a standardized dataset to generate occluded data augmentation view, permutation data augmentation view and Gaussian noise data augmentation view. The feature vector extraction module is used to iteratively optimize the BYOL pre-trained model based on occluded data augmentation view, permutation data augmentation view, and Gaussian noise data augmentation view. It inputs the standardized dataset into the trained BYOL pre-trained model and extracts the dimensionality-reduced feature vector of each data sample. The sensitive information labeling module is used to generate a list of sensitive information labels based on the dimensionality-reduced feature vector of each data sample using an improved DENCLUE algorithm. The desensitization module is used to perform data masking, data generalization, and data replacement desensitization processing on the location of sensitive information in the original data based on the sensitive information tag list, and output preliminary desensitized data; The local density threshold iteration module is used to compare the information entropy difference before and after desensitization and to adaptively adjust the local density threshold in the improved DENCLUE algorithm.

[0022] refer to Figure 2 and Figure 3 A data anonymization method based on self-supervised learning includes the following steps: S1. Based on the original data to be de-identified, perform preprocessing to generate a standardized dataset; S2. Based on the standardized dataset, perform data augmentation processing with random occlusion, random feature permutation, and random noise injection respectively to generate occluded data augmentation view, permutation data augmentation view, and Gaussian noise data augmentation view; S3. Iteratively optimize the BYOL pre-trained model based on occluded data augmentation view, permutation data augmentation view, and Gaussian noise data augmentation view. Input the standardized dataset into the trained BYOL pre-trained model and extract the dimensionality-reduced feature vector of each data sample. S4. Based on the dimensionality-reduced feature vector of each data sample, execute the improved DENCLUE algorithm to generate a list of sensitive information tags; S5. Based on the list of sensitive information markers, perform data masking, data generalization, and data substitution desensitization processing on the location of sensitive information in the original data, and output preliminary desensitized data; S6. Compare the information entropy difference before and after desensitization, and adaptively adjust the local density threshold in the improved DENCLUE algorithm iteratively.

[0023] In this embodiment, the preprocessing includes cleaning, normalization, removal of outliers, and data format unification.

[0024] This preprocessing step effectively improves the quality and processing efficiency of equipment status data. Data cleaning removes noise and errors from the raw data, ensuring accuracy and reliability. Data normalization unifies data from different scales to a standardized range, facilitating subsequent analysis and processing and avoiding problems caused by differences in data scale. Outlier removal effectively eliminates deviations caused by acquisition or transmission issues, further improving data authenticity and consistency. Data format standardization ensures consistency and compatibility, enabling seamless integration and further processing of data from different sources. Through these preprocessing steps, data quality is significantly improved, and the accuracy and efficiency of analysis and modeling are greatly enhanced, ensuring the reliability and stability of equipment status data and ultimately improving the accuracy and operability of subsequent analysis results.

[0025] In this embodiment, S2 specifically includes: S21. Input a standardized dataset, set a random occlusion ratio, randomly select data items with the corresponding ratio, set a fixed occlusion marker value and replace the original value of the data item to obtain an enhanced view of the occlusion data. S22. Input a standardized dataset, set a random permutation ratio, randomly select data items of the corresponding ratio and swap the positions of the data items to obtain a permuted data augmented view; S23. Input a standardized dataset, set the random noise injection ratio, noise mean and variance, randomly select the corresponding proportion of numerical data items and overlay Gaussian noise to obtain a Gaussian noise data augmentation view.

[0026] This step significantly improves the model's robustness and generalization ability through three different data augmentation strategies. Masking data augmentation randomly masks some data items in the input dataset and replaces them with fixed label values, generating a masked augmented view. This method effectively prevents the model from overfitting to certain specific data items and improves the model's tolerance to missing or incomplete data. Permutation data augmentation enhances the model's robustness under different feature arrangements by randomly permuting the positions of data items, helping the model understand data relationships from different perspectives and thus improving its ability to handle unstructured data. Gaussian noise injection adds Gaussian noise to randomly selected data items, simulating noise interference in the data and helping the model better cope with data fluctuations in real-world scenarios, enhancing the model's tolerance to noise and accuracy.

[0027] In this embodiment, S3 specifically includes: S31. From the occlusion data augmentation view, the permutation data augmentation view, and the Gaussian noise data augmentation view, randomly select two different data augmentation views and input them into the feature encoders of the online network and the target network of the BYOL pre-trained model. Perform convolution and pooling operations respectively to obtain the feature encoding vectors of the online network and the target network. S32. Input the feature encoding vectors of the online network and the target network into the projection heads of the online network and the target network respectively. Perform linear transformation, batch normalization and ReLU activation function calculation through the first fully connected layer. Perform linear transformation through the second fully connected layer to output the projection feature vectors of the online network and the target network respectively. S33. Input the projection feature vector of the online network into the prediction head, perform linear transformation, batch normalization and ReLU activation function calculation through the first fully connected layer, perform linear transformation through the second fully connected layer, and output the prediction feature vector of the online network. S34. Calculate the magnitude length values ​​of the predicted feature vector of the online network and the projected feature vector of the target network respectively, and multiply them to obtain the magnitude product value; perform an inner product operation on the predicted feature vector of the online network and the projected feature vector of the target network to obtain the inner product value; divide the inner product value by the magnitude product value to obtain the cosine similarity value; define the loss function value as the negative value of the cosine similarity value. S35. Based on the numerical value of the loss function, the Adam optimization algorithm is used to calculate the gradient through backpropagation. The parameters are updated based on the exponential weighted average of the first and second moments. Combined with adaptive learning rate and bias correction, the parameters of each layer of the online network are iteratively optimized until the loss function converges. S36. Continuously update the parameters of each layer of the target network based on the EMA method. Set the smoothing coefficient value of the EMA method, and set the weights of the target network and the online network to the smoothing coefficient value and the complement of the smoothing coefficient value, respectively. After each update of the parameters of each layer of the online network, calculate the weighted average value of the parameter values ​​for each layer of the online network and the target network in the feature encoder and projection head, based on the weights of the target network and the online network. Use the weighted average value to update the parameters of the feature encoder and projection head in the target network, and output the target network after parameter update. S37. Continue to iterate through steps S35 and S36 until the loss function value reaches a convergent state, and obtain the trained BYOL pre-trained model. S38. Input the standardized dataset into the online network of the trained BYOL model. For each data sample in the standardized dataset, pass it through the feature encoder in the online network in turn. After convolution and pooling operations, obtain the feature encoding vector of each data sample. S39. Input the feature encoding vector of each data sample into the projection head of the online network, and after linear transformation through two layers of fully connected networks, obtain the dimensionality-reduced feature vector of each data sample.

[0028] This implementation significantly improves the feature learning and desensitization capabilities of data by introducing a BYOL pre-trained model and various data augmentation techniques. By inputting occluded data augmentation, permutation data augmentation, and Gaussian noise data augmentation views into the model for training, the system can extract different feature information from diverse data views, enhancing the model's robustness and generalization ability, and ensuring efficiency and stability when processing multimodal and heterogeneous data. Through convolutional operations, pooling operations, and feature encoding and linear transformation of the projection head in both the online and target networks, this invention can accurately capture potential relationships and hidden features in the data. Combined with the cosine similarity loss function, the model can efficiently learn data representations, maximizing consistency between different data views. By progressively updating the parameters of the target network using the EMA method, the target network is ensured to be more stable during training, improving training efficiency. The updating and optimization of model parameters further enhance the accuracy and robustness of predictions, ensuring the accuracy and consistency of data desensitization.

[0029] In this embodiment, the improved DENCLUE algorithm specifically includes: Define the kernel function type and initialize the global bandwidth parameter and momentum coefficient of the kernel function; A local adaptive bandwidth parameter adjustment mechanism is introduced. The dimensionality-reduced feature vectors of each data sample are iterated sequentially. The distance between the current feature vector and other vectors is calculated. A neighborhood size parameter is set, and the nearest feature vectors of the specified neighborhood size parameter are selected to form a neighborhood. The density values ​​of all nearest neighbor feature vectors within the neighborhood are calculated, and the variance of the density within the neighborhood is statistically obtained. Based on the variance of the density within the neighborhood, a density variance threshold is set. If the density variance within the neighborhood exceeds the density variance threshold, the local bandwidth is reduced; otherwise, the local bandwidth is increased. Finally, the local adaptive bandwidth parameter corresponding to each feature vector is obtained. A momentum coefficient mechanism is introduced. Based on the kernel function and the local adaptive bandwidth parameter corresponding to the current feature vector, the current density gradient direction is calculated. A momentum coefficient is set, and the current density gradient direction is combined with the density gradient direction of the previous iteration. The momentum coefficient is used to perform weighted fusion to form the update direction of the momentum. The position of the feature vector is updated along the direction of the momentum update, and the iterative movement continues until the position of the feature vector converges, finally obtaining the density attraction point corresponding to each feature vector. Set a local density threshold, iterate through the density attraction points corresponding to the dimensionality-reduced feature vector of each data sample, calculate the local density value of each density attraction point, and mark the density attraction points whose local density values ​​exceed the corresponding local density threshold as local high density center points, and finally obtain all local high density center points. Using the local high-density center point as a reference, an attraction radius is set, and feature vectors whose density attraction points are located within the attraction radius of the local high-density center are aggregated into corresponding high-density clustering regions, and the clustering results are output. The feature vectors within all local high-density clustering regions are traversed, and the corresponding data sample numbers and data item numbers are extracted to generate a list of sensitive information tags.

[0030] This step significantly improves the accuracy and efficiency of data clustering by introducing a local adaptive bandwidth parameter adjustment mechanism and a momentum coefficient mechanism, ensuring the precision and robustness of data anonymization. By setting the kernel function type and initializing the bandwidth parameters, the system can calculate a suitable local bandwidth for each feature vector, thus more accurately reflecting the local structure of the data. The introduced local adaptive bandwidth adjustment dynamically adjusts the bandwidth based on the density variance of the data samples, making the bandwidth of each data point more consistent with the distribution of its surrounding data, further optimizing the clustering effect. The momentum coefficient mechanism, by combining the current and historical gradient directions, effectively reduces oscillations during feature vector updates, improving the convergence speed and stability of the model. In addition, by calculating density attraction points and local high-density centers, the system can identify key clustering regions in the data and use them as the basis for anonymization. Finally, through the aggregation of high-density clustering regions and the generation of sensitive data labels, the sensitive data items that need to be anonymized are successfully marked. This method significantly improves the automation and accuracy of data anonymization, can efficiently process large-scale high-dimensional data, and ensures effective protection of privacy information while preserving data value. Overall, this invention, through refined clustering and adaptive adjustment mechanisms, not only improves data processing efficiency but also enhances the system's adaptability to complex and dynamic data, providing a more efficient and intelligent solution for data anonymization and privacy protection.

[0031] In this embodiment, S5 specifically includes: S51. Perform numeric replacement desensitization. Iterate through all the marked field positions in the sensitive information mark list. When the marked field is numeric data, determine the upper and lower bounds of the numeric range where the data is located. Based on the upper and lower bounds, construct the range description and replace the original numeric value with the corresponding range. S52. Perform text replacement and desensitization. Iterate through all the marked field positions in the sensitive information mark list. When the marked field is text data, count and construct the non-sensitive text data set of the current field. Randomly select non-sensitive data items from the non-sensitive text data set and replace the original value of the current sensitive text data item. S53. Perform desensitization of highly sensitive fields. Iterate through all the marked field positions in the sensitive information mark list. When the marked field is a highly sensitive data field, set a fixed specific symbol as the masking mark value and completely replace the value of the original data item with the masking mark value. S54. After performing numeric replacement desensitization, text replacement desensitization, and strong sensitive field desensitization, update the positions of the corresponding data sample numbers and data item numbers in the original dataset, and then summarize to obtain the preliminary desensitized dataset.

[0032] This step combines three data masking methods—numerical replacement, text replacement, and highly sensitive field masking—to significantly improve data privacy protection and ensure data usability. Numerical replacement replaces sensitive numerical data with corresponding range descriptions, preventing the leakage of the original data's precise values. Text replacement randomly selects data items from a non-sensitive text dataset to replace sensitive fields, effectively reducing the risk of privacy breaches while maintaining data structure. Highly sensitive field masking uses fixed symbols to mask sensitive data, further protecting extremely sensitive information from leakage. Overall, these masking operations effectively reduce the risk of data leakage while maintaining the dataset's overall structure and characteristics, providing secure input for subsequent data analysis, processing, and modeling. Ultimately, the dataset processed through these masking steps retains important data information while protecting privacy, effectively improving the accuracy, intelligence, and automation of data masking.

[0033] In this embodiment, S6 specifically includes: S61. Input the initial desensitized data and the original data before desensitization, and count the frequency of each data field. Calculate the information entropy value of each field based on its frequency of value, and summarize the information entropy values ​​of all fields to obtain the information entropy of the initial desensitized data and the information entropy of the original data before desensitization. S62. Input the information entropy of the original data before desensitization and the information entropy of the preliminary desensitized data, calculate the numerical difference between the information entropy of the original data and the information entropy of the preliminary desensitized data, and obtain the information entropy difference value. S63. Set an information entropy difference threshold. If the information entropy difference is greater than the information entropy difference threshold, increase the density threshold set in the improved DENCLUE algorithm to obtain an updated local density threshold. Use the updated density threshold to perform the operations in steps S4 and S5; otherwise, output the data that has been desensitized as the final desensitized data.

[0034] This step in this implementation method further enhances data security and privacy protection during the de-identification process by calculating the information entropy difference. By statistically analyzing the frequency of values ​​for each data field and calculating the information entropy, the amount of information contained in the data can be quantified, and the difference between the de-identified data and the original data can be evaluated. Through the calculation of the information entropy difference, the system can determine whether the de-identified data is sufficiently privacy-protecting. If the difference exceeds a preset threshold, the local density threshold is dynamically adjusted using an improved DENCLUE algorithm to further optimize the data de-identification effect and enhance data privacy. If the information entropy difference meets the requirements, the data is considered to have been effectively de-identified, and the system outputs the de-identified data as the final data. This method effectively ensures privacy protection and data quality during the de-identification process, maximizing the preservation of effective information while guaranteeing data privacy and security, and improving the accuracy and intelligence of the de-identification process.

[0035] Example 1: To verify the feasibility of this invention in data anonymization, it was applied to the data anonymization module of a company's sensitive data management system (hereinafter referred to as "System M"). Traditional data anonymization methods typically rely on rule-based fixed methods and manual intervention. These methods not only have limited anonymization effects but also lack flexibility when processing complex data, making it difficult to meet the needs for efficient and accurate anonymization. To address these issues, Platform Y decided to adopt the self-supervised learning-based data anonymization system proposed in this invention.

[0036] During implementation, System M collected a large-scale dataset, including textual data, numerical data, and multi-source heterogeneous data, through a data acquisition platform. The system then performed data cleaning, imputation, noise reduction, and normalization to obtain standardized data in a unified format. By extracting high-quality feature representations from the data using a BYOL pre-trained model, the system can learn deeper feature expressions from various data views and generate effective feature vectors. These vectors provide the foundation for subsequent de-identification processing.

[0037] After feature extraction, system M uses an improved DENCLUE algorithm to perform cluster analysis on the data, automatically identifying sensitive information regions and high-density clustering regions. Through algorithm optimization, the system can flexibly adjust bandwidth and density thresholds, automatically optimizing clustering based on data changes. The system can dynamically identify sensitive data fields and perform precise de-identification operations based on local high-density regions, effectively protecting data privacy.

[0038] To further verify the effectiveness of the method of the present invention, system M compared the performance of the method of the present invention with traditional desensitization methods. The specific comparison data is shown in Table 1: Table 1. Performance Comparison of Data Desensitization Processing System M

[0039] As shown in Table 1, the application of the method of this invention significantly improves the data anonymization accuracy from 80.2% to 96.7% compared to traditional methods, substantially reduces the rate of sensitive information leakage, and increases data processing efficiency by approximately 188%, enabling more efficient data anonymization processing. Data noise reduction is significantly improved, with the average processing time reduced from 15.5 seconds per data entry to 3.2 seconds per data entry, significantly enhancing the system's real-time response capability. Simultaneously, the accuracy of data reconstruction after anonymization increases from 85.4% to 96.3%, indicating that the anonymized data better preserves the effective information of the original data. The average monthly attendance of system maintenance personnel decreases by 63.6%, and the compliance check pass rate after anonymization increases from 87.3% to 98.5%, further demonstrating that the method of this invention ensures data compliance and efficiency while protecting privacy.

[0040] Through the method of this invention, system M can complete more accurate data anonymization processing in a shorter time, reducing the need for manual intervention, improving the automation level of data privacy protection, and further ensuring enterprise data security. Overall, this invention provides a more efficient and intelligent solution for data anonymization processing and offers strong technical support for enterprise data management and compliance.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A data anonymization system based on self-supervised learning, characterized in that, Includes the following modules: The preprocessing module is used to perform preprocessing on the original data to be de-identified, and generate a standardized dataset. The data augmentation module is used to perform random occlusion, random feature permutation and random noise injection on a standardized dataset to generate occluded data augmentation view, permutation data augmentation view and Gaussian noise data augmentation view. The feature vector extraction module is used to iteratively optimize the BYOL pre-trained model based on occluded data augmentation view, permutation data augmentation view, and Gaussian noise data augmentation view. It inputs the standardized dataset into the trained BYOL pre-trained model and extracts the dimensionality-reduced feature vector of each data sample. The sensitive information labeling module is used to generate a list of sensitive information labels based on the dimensionality-reduced feature vector of each data sample using an improved DENCLUE algorithm. The desensitization module is used to perform data masking, data generalization, and data replacement desensitization processing on the location of sensitive information in the original data based on the sensitive information tag list, and output preliminary desensitized data; The local density threshold iteration module is used to compare the information entropy difference before and after desensitization and to adaptively adjust the local density threshold in the improved DENCLUE algorithm.

2. A data anonymization system based on self-supervised learning, characterized in that, The modules are connected through the following steps: S1. Based on the original data to be de-identified, perform preprocessing to generate a standardized dataset; S2. Based on the standardized dataset, perform data augmentation processing with random occlusion, random feature permutation, and random noise injection respectively to generate occluded data augmentation view, permutation data augmentation view, and Gaussian noise data augmentation view; S3. Iteratively optimize the BYOL pre-trained model based on occluded data augmentation view, permutation data augmentation view, and Gaussian noise data augmentation view. Input the standardized dataset into the trained BYOL pre-trained model and extract the dimensionality-reduced feature vector of each data sample. S4. Based on the dimensionality-reduced feature vector of each data sample, execute the improved DENCLUE algorithm to generate a list of sensitive information tags; S5. Based on the list of sensitive information markers, perform data masking, data generalization, and data substitution desensitization processing on the location of sensitive information in the original data, and output preliminary desensitized data; S6. Compare the information entropy difference before and after desensitization, and adaptively adjust the local density threshold in the improved DENCLUE algorithm iteratively.

3. The data anonymization system based on self-supervised learning according to claim 2, characterized in that, The preprocessing includes cleaning, normalization, removal of outliers, and data format standardization.

4. The data anonymization system based on self-supervised learning according to claim 2, characterized in that, S2 specifically includes: S21. Input a standardized dataset, set a random occlusion ratio, randomly select data items with the corresponding ratio, set a fixed occlusion marker value and replace the original value of the data item to obtain an enhanced view of the occlusion data. S22. Input a standardized dataset, set a random permutation ratio, randomly select data items of the corresponding ratio and swap the positions of the data items to obtain a permuted data augmented view; S23. Input a standardized dataset, set the random noise injection ratio, noise mean and variance, randomly select the corresponding proportion of numerical data items and overlay Gaussian noise to obtain a Gaussian noise data augmentation view.

5. A data anonymization system based on self-supervised learning according to claim 2, characterized in that, S3 specifically includes: S31. From the occlusion data augmentation view, the permutation data augmentation view, and the Gaussian noise data augmentation view, randomly select two different data augmentation views and input them into the feature encoders of the online network and the target network of the BYOL pre-trained model. Perform convolution and pooling operations respectively to obtain the feature encoding vectors of the online network and the target network. S32. Input the feature encoding vectors of the online network and the target network into the projection heads of the online network and the target network respectively. Perform linear transformation, batch normalization and ReLU activation function calculation through the first fully connected layer. Perform linear transformation through the second fully connected layer to output the projection feature vectors of the online network and the target network respectively. S33. Input the projection feature vector of the online network into the prediction head, perform linear transformation, batch normalization and ReLU activation function calculation through the first fully connected layer, perform linear transformation through the second fully connected layer, and output the prediction feature vector of the online network. S34. Calculate the magnitude length values ​​of the predicted feature vector of the online network and the projected feature vector of the target network respectively, and multiply them to obtain the magnitude product value; perform an inner product operation on the predicted feature vector of the online network and the projected feature vector of the target network to obtain the inner product value; divide the inner product value by the magnitude product value to obtain the cosine similarity value; define the loss function value as the negative value of the cosine similarity value. S35. Based on the numerical value of the loss function, the Adam optimization algorithm is used to calculate the gradient through backpropagation. The parameters are updated based on the exponential weighted average of the first and second moments. Combined with adaptive learning rate and bias correction, the parameters of each layer of the online network are iteratively optimized until the loss function converges. S36. Continuously update the parameters of each layer of the target network based on the EMA method. Set the smoothing coefficient value of the EMA method, and set the weights of the target network and the online network to the smoothing coefficient value and the complement of the smoothing coefficient value, respectively. After each update of the parameters of each layer of the online network, calculate the weighted average value of the parameter values ​​for each layer of the online network and the target network in the feature encoder and projection head, based on the weights of the target network and the online network. Use the weighted average value to update the parameters of the feature encoder and projection head in the target network, and output the target network after parameter update. S37. Continue to iterate through steps S35 and S36 until the loss function value reaches a convergent state, and obtain the trained BYOL pre-trained model. S38. Input the standardized dataset into the online network of the trained BYOL model. For each data sample in the standardized dataset, pass it through the feature encoder in the online network in turn. After convolution and pooling operations, obtain the feature encoding vector of each data sample. S39. Input the feature encoding vector of each data sample into the projection head of the online network, and after linear transformation through two layers of fully connected networks, obtain the dimensionality-reduced feature vector of each data sample.

6. A data anonymization system based on self-supervised learning according to claim 2, characterized in that, The improved DENCLUE algorithm specifically includes: Define the kernel function type and initialize the global bandwidth parameter and momentum coefficient of the kernel function; A local adaptive bandwidth parameter adjustment mechanism is introduced. The dimensionality-reduced feature vectors of each data sample are iterated sequentially. The distance between the current feature vector and other vectors is calculated. A neighborhood size parameter is set, and the nearest feature vectors of the specified neighborhood size parameter are selected to form a neighborhood. The density values ​​of all nearest neighbor feature vectors within the neighborhood are calculated, and the variance of the density within the neighborhood is statistically obtained. Based on the variance of the density within the neighborhood, a density variance threshold is set. If the density variance within the neighborhood exceeds the density variance threshold, the local bandwidth is reduced; otherwise, the local bandwidth is increased. Finally, the local adaptive bandwidth parameter corresponding to each feature vector is obtained. A momentum coefficient mechanism is introduced. Based on the kernel function and the local adaptive bandwidth parameter corresponding to the current feature vector, the current density gradient direction is calculated. A momentum coefficient is set, and the current density gradient direction is combined with the density gradient direction of the previous iteration. The momentum coefficient is used to perform weighted fusion to form the update direction of the momentum. The position of the feature vector is updated along the direction of the momentum update, and the iterative movement continues until the position of the feature vector converges, finally obtaining the density attraction point corresponding to each feature vector. Set a local density threshold, iterate through the density attraction points corresponding to the dimensionality-reduced feature vector of each data sample, calculate the local density value of each density attraction point, and mark the density attraction points whose local density values ​​exceed the corresponding local density threshold as local high density center points, and finally obtain all local high density center points. Using the local high-density center point as a reference, an attraction radius is set, and feature vectors whose density attraction points are located within the attraction radius of the local high-density center are aggregated into corresponding high-density clustering regions, and the clustering results are output. The feature vectors within all local high-density clustering regions are traversed, and the corresponding data sample numbers and data item numbers are extracted to generate a list of sensitive information tags.

7. A data anonymization system based on self-supervised learning according to claim 2, characterized in that, S5 specifically includes: S51. Perform numeric replacement desensitization. Iterate through all the marked field positions in the sensitive information mark list. When the marked field is numeric data, determine the upper and lower bounds of the numeric range where the data is located. Based on the upper and lower bounds, construct the range description and replace the original numeric value with the corresponding range. S52. Perform text replacement and desensitization. Iterate through all the marked field positions in the sensitive information mark list. When the marked field is text data, count and construct the non-sensitive text data set of the current field. Randomly select non-sensitive data items from the non-sensitive text data set and replace the original value of the current sensitive text data item. S53. Perform desensitization of highly sensitive fields. Iterate through all the marked field positions in the sensitive information mark list. When the marked field is a highly sensitive data field, set a fixed specific symbol as the masking mark value and completely replace the value of the original data item with the masking mark value. S54. After performing numeric replacement desensitization, text replacement desensitization, and strong sensitive field desensitization, update the positions of the corresponding data sample numbers and data item numbers in the original dataset, and then summarize to obtain the preliminary desensitized dataset.

8. A data anonymization system based on self-supervised learning according to claim 2, characterized in that, S6 specifically includes: S61. Input the initial desensitized data and the original data before desensitization, and count the frequency of each data field. Calculate the information entropy value of each field based on its frequency of value, and summarize the information entropy values ​​of all fields to obtain the information entropy of the initial desensitized data and the information entropy of the original data before desensitization. S62. Input the information entropy of the original data before desensitization and the information entropy of the preliminary desensitized data, calculate the numerical difference between the information entropy of the original data and the information entropy of the preliminary desensitized data, and obtain the information entropy difference value. S63. Set an information entropy difference threshold. If the information entropy difference is greater than the information entropy difference threshold, increase the density threshold set in the improved DENCLUE algorithm to obtain an updated local density threshold. Use the updated density threshold to perform the operations in steps S4 and S5; otherwise, output the data that has been desensitized as the final desensitized data.

Images