Privacy data hierarchical protection method for big data intelligent detection

Abstract

This invention discloses a privacy-graded data protection method for intelligent detection of big data. In the field of information security technology, the method first segments the original data text into words, extracts high-dimensional semantic features of each word based on a word embedding model and a context-aware encoder, and calculates dynamic sensitivity scores using a pre-trained deep neural network. Based on this, the words are divided into five increasing sensitivity categories. A differentiated privacy budget strategy is adopted, and intensity-adapted Laplacian noise is injected based on a differential privacy mechanism to generate privacy-preserving word features and construct privacy-protected text. This allows for direct input into a big data intelligent detection system to perform intelligent analysis tasks such as named entity recognition, keyword extraction, relation extraction, and sensitive content identification. Accurate detection is achieved while ensuring privacy security, balancing privacy protection strength with data usability. This method is suitable for privacy computing and intelligent analysis needs in large-scale data scenarios.

Description

Technical Field

[0001] This invention belongs to the field of information security technology, specifically, it relates to a method for graded protection of privacy data for intelligent detection of big data. Background Technology

[0002] While intelligent detection systems provide users with accurate and efficient services by collecting and analyzing massive amounts of user data, their in-depth utilization of data also brings the risk of privacy leaks.

[0003] Application CN110020546B discloses a method for graded protection of privacy data. Addressing the issue that most existing models treat all privacy data equally without considering the degree of privacy sensitivity, this method treats the degree of privacy as a metric. By calculating the sensitivity level of privacy data and limiting the frequency of highly sensitive data to a lower value, it provides a higher level of protection for highly sensitive privacy data. Furthermore, this method takes a user-centric approach. After implementing graded privacy protection, it directly calculates the benefits users can gain, linking the results of graded privacy protection to the user's service quality and privacy losses. This ensures that users receive satisfactory service from the system while guaranteeing their data privacy.

[0004] In practice, this method only considers features that satisfy monotonicity. However, in complex big data scenarios, the sensitivity of data is often multidimensional and nonlinear. For example, if we consider three features at the same time, such as disease type, age and region, the sensitivity of the combination may be much higher than the sum of the individual features. Moreover, it is not a simple monotonic relationship. Therefore, it will ignore complex combined features, leading to protection vulnerabilities. Secondly, this method simply limits the frequency of highly sensitive data. While protecting privacy, it severely damages the statistical distribution and availability of the original dataset. For example, in a medical dataset, if heart disease is excessively suppressed as highly sensitive data, then the disease prediction model trained on this dataset will fail for heart disease patients, leading to a decline in service fairness and availability. Finally, this method mainly prevents attacks that infer privacy based on the frequency of data occurrence. However, in the era of big data, there are many more advanced attacks, such as association attacks that link multiple seemingly insensitive data together, and member inference attacks that determine whether a record is in the training set. Simply using frequency limits is extremely ineffective in defending against these attacks.

[0005] To address the aforementioned issues, this invention proposes a privacy data tiered protection method for intelligent big data detection. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a privacy data tiered protection method for intelligent big data detection, which solves the problems of low accuracy in data sensitivity characterization, poor detection task effectiveness, and insufficient robustness of tiered privacy protection mechanisms in complex big data environments.

[0007] The objective of this invention can be achieved through the following technical solutions: A privacy-graded data protection method for intelligent big data detection, the method comprising: Step 1: Extract the original data text, perform word segmentation on the original data text, determine the word segmentation set, extract the high-dimensional features of each word segmentation, and combine the pre-built deep neural network to analyze the dynamic sensitivity score of each word segmentation. Sensitive words are extracted using the scoring distribution and classified into sensitivity categories to form a set of sensitivity category word segments; Step 2: Allocate differentiated privacy budgets to word segments of different sensitivity categories, inject noise of corresponding intensity into word segments of different sensitivity categories based on differential privacy mechanism, generate privacy-preserving word segmentation features associated with each word segmentation, and construct privacy-preserving text; Step 3: Use the privacy-protected text as input to apply to the big data intelligent detection system, execute intelligent analysis tasks, extract detection results based on the output of the intelligent analysis tasks, and complete the hierarchical privacy protection and intelligent detection of the original data.

[0008] As a further aspect of the present invention, in step one, the original data text is extracted, and word segmentation is performed on the original text to determine the word segmentation set in the following specific way: Lock the raw data text to be processed; Extract the text content of the original data text in the original text order, and denote it as Q; Output the text content Q to the word segmentation library, perform word segmentation on the text content Q, determine all the words in the text content Q, and sort them according to the original text order, which is recorded as the initial word segmentation set; The initial word segmentation set is input into the large model, and effective word segmentation filtering is performed in combination with the word segmentation filtering rules predefined by the operator. The filtered effective words are output, arranged in the original order, and labeled as the word segmentation set R={R1,R2,...,Rj}, where j is the total number of effective words.

[0009] As a further aspect of the present invention, the specific method for extracting the high-dimensional features of each word segmentation and then combining them with a pre-constructed deep neural network to analyze the dynamic sensitivity score of each word segmentation is as follows: Extract any one word Ri from the word segmentation set R, where i is the counting index, with a value ranging from 1 to j; The word segmentation Ri output is converted into a word embedding model pre-trained by the operator based on the existing model and transformed into high-dimensional word vectors; Obtain the context information of the word segment Ri in the text content Q, where the coverage of the context information is a preset number of characters; The semantic features of the word segment Ri based on context information are extracted by the context-aware encoder, and then the encoder is fused with the high-dimensional word vector to generate the high-dimensional features associated with the word segment Ri, denoted as ZRi; The high-dimensional feature ZRi is input into a pre-constructed deep neural network, which contains multiple fully connected layers and an output layer. The output layer uses the softmax function in combination with the high-dimensional feature ZRi to generate a probability distribution of the word segment Ri belonging to a preset sensitivity category. The sensitivity categories include Category 1, Category 2, Category 3, Category 4, and Category 5, and the sensitivity relationship of each category is Category 1 < Category 2 < Category 3 < Category 4 < Category 5; Extract the probability distribution of the word segment Ri belonging to each category in the sensitivity category, and label them as θ1, θ2, θ3, θ4, and θ5 in order from the first to the fifth category; Extract the probability distribution preset by the operator to calculate the weights, which are ω1, ω2, ω3, ω4, and ω5 in sequence, and 0 < ω1 < ω2 < ω3 < ω4 < ω5, ω1 + ω2 + ω3 + ω4 + ω5 = 1; The dynamic sensitivity score of word segment Ri is calculated using MG_i=θ1×ω1+θ2×ω2+θ3×ω3+θ4×ω4+θ5×ω5. The higher the MG_i score, the higher the sensitivity of the privacy information contained in the word segment Ri, and vice versa. Similarly, determine the dynamic sensitivity scores of all word segments in the word segmentation set R, denoted as the dynamic sensitivity score sequence MG_1,MG_2,...,MG_j.

[0010] As a further aspect of the present invention, in step one, the specific method for extracting sensitive words using the scoring distribution and classifying them into sensitivity categories to form a sensitivity category word segmentation queue set is as follows: In the word segmentation set R, words with a dynamic sensitivity score ≥ T are marked as sensitive words, and the remaining words are marked as non-sensitive words, where T is a sensitivity threshold preset by combining the sensitivity category and the dynamic sensitivity score; Extract all sensitive words and construct a subset of sensitive words R'={R1',R2',...,Rm'} in the original text order, where m is the total number of sensitive words; Extract the sensitivity categories and the corresponding score ranges for each sensitivity category. The score ranges are non-overlapping and cover the score range of all sensitive words. Take any sensitive word Rn' from the sensitive word subset R', where n is the counting index, and the value ranges from 1 to m; Based on the dynamic sensitivity score MG_n' of the sensitive word segment Rn', determine its score range and assign it the corresponding sensitivity category; Sensitive words of the same sensitivity category are arranged in the original text order in the original text content Q to form a word segmentation queue Ek corresponding to the sensitivity category, where k=1,2,3,4,5, corresponding to the first category, the second category, the third category, the fourth category, and the fifth category. Construct a set of word segmentation queues for all sensitivity categories {E1, E2, ..., E5}, denoted as the sensitivity category word segmentation queue set E.

[0011] As a further aspect of the present invention, the specific method for allocating differentiated privacy budgets to word segmentation of different sensitivity categories in step two is as follows: Extract the pre-defined privacy budget allocation table corresponding to the sensitivity category, as follows: Category 1: Privacy Budget ε1; Category 2: Privacy Budget ε2; Category 3: Privacy Budget ε3; Category 4: Privacy Budget ε4; Category 5: Privacy Budget ε5; And it satisfies ε1<ε2<ε3<ε4<ε5; Extract the sensitivity category word segmentation queue set E={E1,E2,...,E5}; Take the first segmentation queue E1, and based on the first category to which the segmentation queue belongs, assign a privacy budget ε1 to each segmentation in the segmentation queue, and generate privacy budget labels associated with each segmentation; Similarly, privacy budget labels are generated for all word segments in the word segmentation queue to complete the differentiated privacy budget allocation.

[0012] As a further aspect of the present invention, the specific method for generating the privacy-related word segmentation features of each word segmentation in step two is as follows: Extract the high-dimensional features ZRi of any word Ri in the sensitive word segmentation subset R' and the privacy budget εu corresponding to the privacy budget label, where u takes values ​​from 1 to k; Extract the global sensitivity Δ of the operator to the high-dimensional features, and calculate the noise scale parameter b=Δ / ε based on the privacy budget εu; Generate a Laplacian noise vector Lap(0,b) with the same dimension as the high-dimensional feature ZRi of the word segmentation Ri, and add it to the high-dimensional feature ZRi to obtain the noisy high-dimensional feature ZRi', which is used as the privacy segmentation feature of the word segmentation Ri; Similarly, determine the privacy segmentation features of all segments in the sensitive segmentation subset R'.

[0013] As a further aspect of the present invention, the specific method for constructing the privacy-protected text in step two is as follows: All privacy-preserving segmentation features are associated with the original text content Q in the original order of the segmentation and stored after effective segmentation filtering, generating privacy-preserving text QX. In this case, non-sensitive segmentation features are stored in the form of high-dimensional features.

[0014] As a further aspect of the present invention, the specific method for completing the hierarchical privacy protection and intelligent detection of the original data text in step three is as follows: Extract the representation information of all valid word segments in the privacy-preserving text QX, where sensitive word segments are represented as privacy-preserving word segmentation features and non-sensitive word segments are represented as high-dimensional features; Arrange the representation information of all effective word segments according to the original text order in the original text content Q to form a feature sequence; The feature sequence is used as input data and loaded into the detection model of the big data intelligent detection system. The detection model is an intelligent analysis model pre-trained based on a deep neural network architecture, and its network structure includes an embedding layer, an encoding layer and an output layer. The detection model is used to perform intelligent analysis tasks on feature sequences. The feature sequences are mapped into continuous vector representations through the embedding layer, and then deep features with context dependence are extracted based on the encoding layer. Finally, the output layer generates prediction results corresponding to the preset analysis target. The prediction results are extracted as the detection results output of the intelligent analysis task, thus integrating hierarchical privacy protection and intelligent detection of the original data text.

[0015] As a further aspect of the present invention, the analysis objectives in the intelligent analysis task include named entity recognition, keyword extraction, relation extraction, and sensitive content recognition. Currently, a prediction result is generated only when none of the analysis objectives are detected, and this prediction result is output as the detection result of the intelligent analysis task. Conversely, extract the word segments corresponding to the content detected by the analysis target. If the word segments are non-sensitive, they are promoted to the first category and processed with a privacy budget ε1. If it falls into the second, third, or fourth category, it will be upgraded to the third, fourth, or fifth category respectively, and processed with the corresponding privacy budget, and the above steps will be repeated.

[0016] The beneficial effects of this invention are: This invention utilizes deep neural networks to perform refined sensitivity analysis on text data and implements differentiated privacy protection accordingly. It achieves a balance between privacy and data utility, abandoning the traditional one-size-fits-all protection model. By using deep learning to capture the high-dimensional semantic features of word segmentation to quantify dynamic sensitivity, highly sensitive information is masked by strong noise, while less sensitive information retains more of its original semantics. This enhances the ability to combat privacy leaks during big data mining, maximizes data availability, and ensures that the anonymized text can still effectively support subsequent intelligent detection and analysis tasks. While protecting user privacy, it also considers the accuracy and reliability of big data system analysis, providing a solution for the symbiosis of privacy protection and data value mining. This invention improves the accuracy and granularity of sensitive information detection through multi-dimensional feature fusion and dynamic scoring mechanisms. It utilizes word embedding and context-aware encoders to extract high-dimensional semantic features from word segmentation, enabling the model to not only understand the words themselves but also accurately grasp their sensitive meaning in specific scenarios, effectively avoiding the mechanical misjudgments of traditional keyword matching. By generating a fine-grained probability distribution of sensitive categories through deep neural networks and calculating dynamic sensitivity scores with preset weights, it achieves continuous quantitative evaluation of sensitivity levels from low to high. Based on the score distribution, threshold filtering and category classification are performed to construct a multi-level sensitivity segmentation queue, enabling the location of sensitive information and providing a structured basis for differentiated desensitization processing. This invention achieves a precise match between privacy protection strength and data sensitivity by assigning differentiated privacy budgets to word segments of different sensitivity categories. This allows highly sensitive words to receive stronger noise perturbations to resist privacy attacks, while low-sensitivity words retain more original semantic features, thus improving the overall usability of privacy-protected text. By extracting high-dimensional features from a subset of sensitive words and adding Laplacian noise calculated based on global sensitivity and privacy budget, privacy-protected word features of the same dimension can be generated, mathematically guaranteeing strict differential privacy protection. Finally, the privacy features of all words are recombined in the original text order, while retaining the original high-dimensional features of non-sensitive words. This allows the generated privacy-protected text to effectively hide specific sensitive information while maintaining the original contextual coherence. This invention dynamically adjusts the privacy budget based on the sensitivity of word segmentation through a hierarchical processing mechanism, thereby maximizing the utilization value of non-sensitive information while ensuring the security of sensitive information, thus balancing privacy protection and data availability. Simultaneously, the detection model utilizes embedding, encoding, and output layers to extract deep features with contextual dependencies, enabling accurate multi-task analysis including named entity recognition, keyword extraction, relation extraction, and sensitive content identification. Results are only output when all targets are not detected; otherwise, the detected content undergoes privacy upgrades and is processed repeatedly, forming a closed-loop feedback loop. This enhances the accuracy and adaptability of detection, reduces the risk of privacy leakage, and improves the robustness of the intelligent analysis system. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a schematic diagram of the system described in this invention; Figure 2 This is a flowchart illustrating the content described in Embodiment 2 of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1 As shown, this application provides a method for graded protection of privacy data for big data intelligent detection; As an embodiment 1 of this application, it specifically includes: Step 1: Extract the original data text, perform word segmentation on the original data text, determine the word segmentation set, extract the high-dimensional features of each word segmentation, and combine the pre-built deep neural network to analyze the dynamic sensitivity score of each word segmentation. Sensitive words are extracted using the scoring distribution and classified into sensitivity categories to form a set of sensitivity category word segments; Step 2: Allocate differentiated privacy budgets to word segments of different sensitivity categories, inject noise of corresponding intensity into word segments of different sensitivity categories based on differential privacy mechanism, generate privacy-preserving word segmentation features associated with each word segmentation, and construct privacy-preserving text; Step 3: Use the privacy-protected text as input to apply to the big data intelligent detection system, execute intelligent analysis tasks, extract detection results based on the output of the intelligent analysis tasks, and complete the hierarchical privacy protection and intelligent detection of the original data. Example 2

[0021] This embodiment further discloses a method for identifying and classifying text privacy information based on multi-category sensitivity scoring, as described in Embodiment 1. Figure 2 As shown, it specifically includes the following: First, extract any raw data text to be processed as determined by the operator, such as a medical record or a drug purchase list, or other raw data text containing sensitive content. Extract the original text order (reading order or writing order) from the original data text, extract the text content from the original data text, and record it as the text content Q; Next, the text content Q is input into the prepared word segmentation library (the word segmentation library is the part covered by the existing technology, and this solution will not go into too much detail). The word segmentation library performs word segmentation on the text content Q, and finally extracts all the words in the text content Q. All the words are sorted according to the original text order of the original data text and recorded as the preliminary word segmentation set. Next, the initial word segmentation set is input into the large model, and effective word segmentation filtering is performed in combination with the word segmentation filtering rules predefined by the operator. The large model outputs the filtered effective words, and all effective words are sorted according to the original text order of the original data text and marked as the word segmentation set R={R1,R2,...,Rj}, where j is the total number of effective words; It should be noted that the large model mentioned is an existing large model, which can effectively filter the initial word segmentation set by using the given word segmentation filtering rules; For example: First, input a preliminary word segmentation set into the large model, then ask it to filter out all adjectives, and output the effective word segments after filtering out all adjectives according to the order of the preliminary word segmentation set. This completes the effective word segmentation filtering. Here, adjectives are the word segmentation filtering rules. Operators can set the word segmentation filtering rules according to actual requirements.

[0022] Then, the high-dimensional features of all words in the word segmentation set R={R1,R2,...,Rj} are extracted, and the dynamic sensitivity score of each word is analyzed in combination with the deep neural network pre-built by the operator. First, extract any word Ri from the word segmentation set R={R1,R2,...,Rj}, where i is the counting index, with a value range from 1 to j. Perform example processing on word Ri, and process the remaining words in the same way and synchronously. In addition, all the text content processing methods described in this scheme can be backtracked based on their processing algorithms or processing models to obtain the original text content. The extracted word segment Ri is input into a word embedding model pre-trained by the operator based on the existing model. The word embedding model then converts the word segment Ri into high-dimensional word vectors. Among them, word embedding model specifically refers to a word embedding model that has been trained on a large-scale text corpus. It is an existing large model that first needs to be pre-trained on a large-scale general corpus in order to understand the general semantic and grammatical features of words. Secondly, by combining existing word embedding models (such as Word2Vec, GloVe, and FastText) with this large model, a comprehensive model capable of performing input and output operations is obtained. This model is a word embedding model pre-trained based on existing models, and it outputs high-dimensional word vectors.

[0023] Then, based on the position of the word segment Ri in the text content Q, the context information of the word segment Ri in the text content Q is extracted. It should be noted that the coverage of the context information is generally a preset number of characters, such as the 5 characters before the word segment Ri plus the 5 characters after the word segment Ri. In addition to the preset number of characters, existing models can also be used to parse the context information associated with the segmented word Ri. Operators can set the method for extracting the context information of the segmented word Ri according to actual requirements.

[0024] The context-aware encoder is obtained (the context-aware encoder belongs to the technical part covered by natural language processing and deep learning, so this solution will not elaborate on the context-aware encoding). The semantic features of the context information of the word segment Ri are extracted by the context-aware encoder, and the semantic features of the word segment Ri are fused with the high-dimensional word vector of the word segment Ri to finally obtain the high-dimensional feature associated with the word segment Ri, denoted as ZRi. The high-dimensional feature ZRi not only contains the meaning of the word segment Ri itself, but also contains the semantics of the word segment Ri in the specific context. Finally, the high-dimensional feature ZRi of the word segment Ri is input into the deep neural network pre-constructed by the operator, and the probability distribution of the word segment Ri belonging to the preset sensitivity category is output in combination with the high-dimensional feature ZRi. It should be noted that the deep neural network contains multiple fully connected layers and an output layer. The output layer uses the softmax function to process the high-dimensional feature ZRi, and finally generates a probability distribution of the word segment Ri belonging to the preset sensitivity category. In other words, when the operator pre-trains the deep neural network, the deep neural network has already been trained to recognize the sensitivity category associated with the high-dimensional feature. The sensitivity category specifically includes the first category, the second category, the third category, the fourth category, and the fifth category. The sensitivity relationship of each category is as follows: the first category < the second category < the third category < the fourth category < the fifth category. The higher the category, the higher the sensitivity. Obtain the probability distribution of the word segment Ri output by the output layer belonging to each category of the sensitivity category, and arrange them in the order of the first category to the fifth category, denoted as: θ1, θ2, θ3, θ4, θ5; The weights of the probability distribution preset by the operator based on actual needs are calculated and arranged as ω1, ω2, ω3, ω4, ω5, where ω1 to ω5 correspond to the first to fifth categories respectively, and 0 < ω1 < ω2 < ω3 < ω4 < ω5, ω1 + ω2 + ω3 + ω4 + ω5 = 1; Then, the dynamic sensitivity score of the word segment Ri is calculated using MG_i=θ1×ω1+θ2×ω2+θ3×ω3+θ4×ω4+θ5×ω5. Based on the above, it can be seen that the higher the dynamic sensitivity score MG_i of the word segment Ri, the higher the sensitivity of the privacy information contained in the word segment Ri, and vice versa.

[0025] Repeat the above steps to determine the dynamic sensitivity scores of all words in the word segmentation set R, and denote them as the dynamic sensitivity score sequence MG_1,MG_2,...,MG_j.

[0026] Then, based on the word segmentation set R and the dynamic sensitivity score sequence MG_1, MG_2, ..., MG_j, all word segments with a dynamic sensitivity score ≥ T are extracted and marked as sensitive word segments. The remaining word segments with a dynamic sensitivity score less than T are marked as non-sensitive word segments. Here, T is a sensitivity threshold preset by the operator based on the sensitivity category and the dynamic sensitivity score. The operator needs to make adaptive adjustments according to the application scenario. For example, medical records require extremely high privacy protection, so the sensitivity threshold T needs to be set low enough to ensure extremely high privacy protection.

[0027] Extract all the sensitive words finally determined in the above steps, and arrange them according to the original text order of the original data text to form a sensitive word subset R'={R1',R2',...,Rm'}, where m is the total number of sensitive words; Next, all sensitivity categories and their corresponding score ranges (dynamic sensitivity score ranges) are extracted. It should be noted that the score ranges (dynamic sensitivity score ranges) do not overlap and cover the score range of all sensitive words.

[0028] Then, extract any one sensitive word Rn' from the sensitive word subset R', where n is the counting index, with a value range of 1 to m, determine the dynamic sensitivity score MG_n' of the sensitive word Rn', and evaluate the dynamic sensitivity score MG_n' in combination with the score intervals of all sensitivity categories to determine the score interval to which the dynamic sensitivity score MG_n' of the sensitive word Rn' belongs, and assign the corresponding sensitivity category to the sensitive word Rn'; Then, all sensitive words in the sensitive word segmentation subset R' are processed synchronously. Finally, sensitive words of the same sensitivity category are arranged according to the original text order of each sensitive word in the original text content Q, and finally form a word segmentation queue Ek corresponding to the sensitivity category, where k=1,2,3,4,5, 1 corresponds to the first category, 2 corresponds to the second category, 3 corresponds to the third category, 4 corresponds to the fourth category, and 5 corresponds to the fifth category. Finally, a set of word segmentation queues for all sensitivity categories is constructed, {E1,E2,...,E5}, and denoted as the sensitivity category word segmentation queue set E={E1,E2,...,E5}. In this way, the redundant original data text is normalized into a set of sensitive word queues E with clear levels and orderly order. Example 3

[0029] This embodiment, based on embodiment 2, further discloses a method for outputting turn-on and turn-off instructions based on a first-level signal and a second-level signal, specifically including the following: First, obtain the privacy budget allocation table preset by the operator, which corresponds to the sensitivity category. The privacy budget allocation table corresponds to the sensitivity category, as shown below: Category 1: Privacy Budget ε1; Category 2: Privacy Budget ε2; Category 3: Privacy Budget ε3; Category 4: Privacy Budget ε4; Category 5: Privacy Budget ε5; Furthermore, all privacy budgets satisfy ε1<ε2<ε3<ε4<ε5. The privacy budget corresponds to the level of privacy protection. The larger the privacy budget, the greater the noise added and the stronger the privacy protection, and vice versa. This is to avoid using a uniform privacy budget for all word segments, which could lead to excessive loss of information for highly sensitive words or insufficient protection for low-sensitive words. Next, obtain the sensitivity category word segmentation queue set E={E1,E2,...,E5}. Starting from the first word segmentation queue E1, assign the corresponding privacy budget ε1 to all words (sensitive words) in word segmentation queue E1 based on the first category to which word segmentation queue E1 belongs, and generate privacy budget labels for all words, marking the corresponding privacy budget as ε1. Based on the above method, determine the privacy budget for each segmentation in all segmentation queues in the sensitivity category segmentation queue set E={E1,E2,...,E5}, and generate corresponding privacy budget labels to complete the differentiated privacy budget allocation.

[0030] Next, obtain the high-dimensional feature ZRi of any word Ri in the sensitive word segmentation subset R' (here, word Ri and high-dimensional feature ZRi are used as examples for processing), obtain the sensitive category to which word Ri belongs, and determine the privacy budget εu corresponding to the privacy budget label of word Ri based on the corresponding sensitive category, where u takes the value range from 1 to k; Then, the global sensitivity Δ preset by the operator for the high-dimensional features is extracted. The global sensitivity Δ is preset based on the Laplacian mechanism of differential privacy. The noise scale parameter b corresponding to the privacy budget εu is calculated by using b=Δ / ε; Based on the calculated noise scale parameter b, a Laplacian noise vector Lap(0,b) with the same dimension as the high-dimensional feature ZRi of the word segmentation Ri is generated. In this way, by adding noise directly at the feature level instead of modifying the original text, the semantic structure of the feature is preserved, making it easy to use the noise-added feature directly in the future. Next, the Laplacian noise vector Lap(0,b) is added to the high-dimensional feature ZRi of the word segmentation Ri to obtain the high-dimensional feature ZRi' after noise processing, and the high-dimensional feature ZRi' is used as the privacy segmentation feature of the word segmentation Ri.

[0031] Repeat the above steps to determine the privacy features of all words in the sensitive word segmentation subset R'. Then, associate the privacy features of all words with the effective word segmentation filtering and remove words in the original text content Q and store them to generate privacy-protected text QX. In this case, non-sensitive words are stored in the form of high-dimensional features corresponding to the non-sensitive words, rather than directly in the form of the original text, so as to avoid the direct exposure of non-sensitive information. Example 4

[0032] This embodiment further discloses a closed-loop method for achieving intelligent detection based on differential privacy protection, building upon Embodiment 3. Specifically, it includes the following: Based on the content described in Example 3, the privacy-preserving text QX is extracted, and the representation information of all effective word segments in the privacy-preserving text QX is further extracted. It should be noted that sensitive word segments are represented as privacy-preserving word segmentation features, and non-sensitive word segments are represented as high-dimensional features, ensuring that all effective word segments exist in a unified vector form, which is convenient for deep learning models to process. Then, the representation information of all effective word segments is arranged according to the original text order in the text content Q corresponding to the original text data to form a feature sequence. The feature sequence reflects the original structure and contextual relationship of the representation information of effective word segments. Next, the feature sequence is used as input data and loaded into the detection model of the big data intelligent detection system. The detection model of the big data intelligent detection system is an intelligent analysis model pre-trained based on a deep neural network architecture. Its network structure includes an embedding layer, an encoding layer, and an output layer. Essentially, it is a general analysis tool that has been pre-trained on a large-scale corpus. It can extract high-level semantic features from the feature sequence and complete specific tasks. The embedding layer is responsible for mapping the input feature sequence into a continuous vector representation, the encoding layer is responsible for capturing context-dependent deep features, and the output layer generates prediction results corresponding to the preset analysis target. It should be noted that, since the detection model of the big data intelligent detection system is an intelligent analysis model pre-trained based on a deep neural network architecture, which is a general analysis tool pre-trained on a large-scale corpus, it is essentially a part of the existing technology in the big data intelligent detection system. Therefore, this solution will not elaborate further. Finally, the prediction results are extracted and output as the detection results of the intelligent analysis task, thus completing the integration of hierarchical privacy protection and intelligent detection of the original data text.

[0033] It should also be explained that the analytical objectives in intelligent analysis tasks specifically include named entity recognition, keyword extraction, relation extraction, and sensitive content recognition; For example: identifying personal names, place names, and organization names; extracting core keywords; identifying relationships between entities; detecting illegal content such as political and terrorist content; etc. Operators can also add analysis targets based on actual needs; A prediction result is generated only when all analytical targets are not detected, and the prediction result is then used as the output of the intelligent analysis task. If any item in the representation information corresponding to any word segment is detected in the intelligent analysis task, it indicates a detection. In this case, dynamic feedback adjustments are needed to ensure that the appearance of any potentially sensitive content will trigger further intervention by the system, rather than directly outputting results that may contain privacy information. Specifically: If the detected content corresponds to a non-sensitive word segment, the non-sensitive word segment will be promoted to the first category of the sensitivity category and processed with the privacy budget ε1 corresponding to the first category. If the corresponding word segmentation belongs to the second, third, or fourth category, it is upgraded to the third, fourth, or fifth category respectively, and processed with the privacy budget of the corresponding sensitivity category. The above steps are repeated until all analysis targets are no longer detected, thus completing the privacy data classification protection of the original data text.

[0034] All data in the formulas described above have been calculated with dimensions removed. Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0035] The above description is merely an example and illustration of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0036] It should be stated that all user data collected in this application was collected with the user's consent and authorization. Furthermore, the uses of user data are legal and compliant, and the use and processing of user data comply with the relevant laws, regulations, and standards of the relevant regions.

Claims

1. A method for graded protection of privacy data for intelligent detection of big data, characterized in that, The method includes: Step 1: Extract the original data text, perform word segmentation on the original data text, determine the word segmentation set, extract the high-dimensional features of each word segmentation, and combine the pre-built deep neural network to analyze the dynamic sensitivity score of each word segmentation. Sensitive words are extracted using the scoring distribution and classified into sensitivity categories to form a set of sensitivity category word segments; Step 2: Allocate differentiated privacy budgets to word segments of different sensitivity categories, inject noise of corresponding intensity into word segments of different sensitivity categories based on differential privacy mechanism, generate privacy-preserving word segmentation features associated with each word segmentation, and construct privacy-preserving text; Step 3: Use the privacy-protected text as input to apply to the big data intelligent detection system, execute intelligent analysis tasks, extract detection results based on the output of the intelligent analysis tasks, and complete the hierarchical privacy protection and intelligent detection of the original data.

2. The method according to claim 1, characterized in that, In step one, the original data text is extracted, and word segmentation is performed on the original text. The specific method for determining the word segmentation set is as follows: Lock the raw data text to be processed; Extract the text content of the original data text in the original text order, and denote it as Q; Output the text content Q to the word segmentation library, perform word segmentation on the text content Q, determine all the words in the text content Q, and sort them according to the original text order, which is recorded as the initial word segmentation set; The initial word segmentation set is input into the large model, and effective word segmentation filtering is performed in combination with the word segmentation filtering rules predefined by the operator. The filtered effective words are output, arranged in the original order, and labeled as the word segmentation set R={R1,R2,...,Rj}, where j is the total number of effective words.

3. The method according to claim 2, characterized in that, In step one, the specific method for extracting the high-dimensional features of each word segment and then combining them with a pre-constructed deep neural network to analyze the dynamic sensitivity score of each word segment is as follows: Extract any one word Ri from the word segmentation set R, where i is the counting index, with a value ranging from 1 to j; The word segmentation Ri output is converted into a word embedding model pre-trained by the operator based on the existing model and transformed into high-dimensional word vectors; Obtain the context information of the word segment Ri in the text content Q, where the coverage of the context information is a preset number of characters; The semantic features of the word segment Ri based on context information are extracted by the context-aware encoder, and then the encoder is fused with the high-dimensional word vector to generate the high-dimensional features associated with the word segment Ri, denoted as ZRi; The high-dimensional feature ZRi is input into a pre-constructed deep neural network, which contains multiple fully connected layers and an output layer. The output layer uses the softmax function in combination with the high-dimensional feature ZRi to generate a probability distribution of the word segment Ri belonging to a preset sensitivity category. The sensitivity categories include Category 1, Category 2, Category 3, Category 4, and Category 5, and the sensitivity relationship of each category is Category 1 < Category 2 < Category 3 < Category 4 < Category 5; Extract the probability distribution of the word segment Ri belonging to each category in the sensitivity category, and label them as θ1, θ2, θ3, θ4, and θ5 in order from the first to the fifth category; Extract the probability distribution preset by the operator to calculate the weights, which are ω1, ω2, ω3, ω4, and ω5 in sequence, and 0 < ω1 < ω2 < ω3 < ω4 < ω5, ω1 + ω2 + ω3 + ω4 + ω5 = 1; The dynamic sensitivity score of word segment Ri is calculated using MG_i=θ1×ω1+θ2×ω2+θ3×ω3+θ4×ω4+θ5×ω5. The higher the MG_i score, the higher the sensitivity of the privacy information contained in the word segment Ri, and vice versa. Similarly, determine the dynamic sensitivity scores of all word segments in the word segmentation set R, denoted as the dynamic sensitivity score sequence MG_1,MG_2,...,MG_j.

4. The method according to claim 3, characterized in that, In step one, the specific method for extracting sensitive words using the scoring distribution and classifying them into sensitivity categories to form a sensitivity category word segmentation queue set is as follows: In the word segmentation set R, words with a dynamic sensitivity score ≥ T are marked as sensitive words, and the remaining words are marked as non-sensitive words, where T is a sensitivity threshold preset by combining the sensitivity category and the dynamic sensitivity score; Extract all sensitive words and construct a subset of sensitive words R'={R1',R2',...,Rm'} in the original text order, where m is the total number of sensitive words; Extract the sensitivity categories and the corresponding score ranges for each sensitivity category. The score ranges are non-overlapping and cover the score range of all sensitive words. Take any sensitive word Rn' from the sensitive word subset R', where n is the counting index, and the value ranges from 1 to m; Based on the dynamic sensitivity score MG_n' of the sensitive word segment Rn', determine its score range and assign it the corresponding sensitivity category; Sensitive words of the same sensitivity category are arranged in the original text order in the original text content Q to form a word segmentation queue Ek corresponding to the sensitivity category, where k=1,2,3,4,5, corresponding to the first category, the second category, the third category, the fourth category, and the fifth category. Construct a set of word segmentation queues for all sensitivity categories {E1, E2, ..., E5}, denoted as the sensitivity category word segmentation queue set E.

5. The method according to claim 4, characterized in that, In step two, the specific method for allocating differentiated privacy budgets to word segmentation of different sensitivity categories is as follows: Extract the pre-defined privacy budget allocation table corresponding to the sensitivity category, as follows: Category 1: Privacy Budget ε1; Category 2: Privacy Budget ε2; Category 3: Privacy Budget ε3; Category 4: Privacy Budget ε4; Category 5: Privacy Budget ε5; And it satisfies ε1<ε2<ε3<ε4<ε5; Extract the sensitivity category word segmentation queue set E={E1,E2,...,E5}; Take the first segmentation queue E1, and based on the first category to which the segmentation queue belongs, assign a privacy budget ε1 to each segmentation in the segmentation queue, and generate privacy budget labels associated with each segmentation; Similarly, privacy budget labels are generated for all word segments in the word segmentation queue to complete the differentiated privacy budget allocation.

6. The method according to claim 5, characterized in that, In step two, the specific method for generating the privacy-related word segmentation features of each word segmentation is as follows: Extract the high-dimensional features ZRi of any word Ri in the sensitive word segmentation subset R' and the privacy budget εu corresponding to the privacy budget label, where u takes values ​​from 1 to k; Extract the global sensitivity Δ of the operator to the high-dimensional features, and calculate the noise scale parameter b=Δ / ε based on the privacy budget εu; Generate a Laplacian noise vector Lap(0,b) with the same dimension as the high-dimensional feature ZRi of the word segmentation Ri, and add it to the high-dimensional feature ZRi to obtain the noisy high-dimensional feature ZRi', which is used as the privacy segmentation feature of the word segmentation Ri; Similarly, determine the privacy segmentation features of all segments in the sensitive segmentation subset R'.

7. The method according to claim 6, characterized in that, In step two, the specific method for constructing the privacy-protected text is as follows: All privacy-preserving segmentation features are associated with the original text content Q in the original order of the segmentation and stored after effective segmentation filtering, generating privacy-preserving text QX. In this case, non-sensitive segmentation features are stored in the form of high-dimensional features.

8. The method according to claim 7, characterized in that, In step three, the specific method for completing the hierarchical privacy protection and intelligent detection of the original data text is as follows: Extract the representation information of all valid word segments in the privacy-preserving text QX, where sensitive word segments are represented as privacy-preserving word segmentation features and non-sensitive word segments are represented as high-dimensional features; Arrange the representation information of all effective word segments according to the original text order in the original text content Q to form a feature sequence; The feature sequence is used as input data and loaded into the detection model of the big data intelligent detection system. The detection model is an intelligent analysis model pre-trained based on a deep neural network architecture, and its network structure includes an embedding layer, an encoding layer and an output layer. The detection model is used to perform intelligent analysis tasks on feature sequences. The feature sequences are mapped into continuous vector representations through the embedding layer, and then deep features with context dependence are extracted based on the encoding layer. Finally, the output layer generates prediction results corresponding to the preset analysis target. The prediction results are extracted as the detection results output of the intelligent analysis task, thus integrating hierarchical privacy protection and intelligent detection of the original data text.

9. The method according to claim 8, characterized in that, The analysis objectives in the intelligent analysis task include named entity recognition, keyword extraction, relation extraction, and sensitive content recognition. Currently, a prediction result is generated only when none of the analysis objectives are detected, and this prediction result is output as the detection result of the intelligent analysis task. Conversely, extract the word segments corresponding to the content detected by the analysis target. If the word segments are non-sensitive, they are promoted to the first category and processed with a privacy budget ε1. If it falls into the second, third, or fourth category, it will be upgraded to the third, fourth, or fifth category respectively, and processed with the corresponding privacy budget, and the above steps will be repeated.

Images